Construct validation of a portable virtual reality simulator for laparoscopic camera navigation via machine learning and latent behavioral modeling Authors: José Ignacio Guzman Montoto, Mauricio Herrera-Marín, Carolina Andrea Donoso Carrasco, Boris Marinkovic Gómez, Santiago Espinosa Peña, Camilo Ignacio Rodríguez Beltrán, Rodrigo Trigo Vilches Published: 2026 ABSTRACT AbstractDedicated training for laparoscopic camera navigation remains limited by the cost of high-fidelity simulators and the absence of continuous objective assessment. We present SECMA, a portable virtual reality platform combining a mechanically constrained interface that reproduces the trocar pivot, decoupled horizon stabilisation, and optical-redirection channels specific to camera handling with high-frequency 6-DoF telemetry. To establish construct-related validity evidence, a six-target navigation task was administered to 20 surgeons and 18 medical students across four practice sessions. Three analytical layers were applied. First, under repeated nested cross-validation (5 × 50 outer folds), Logistic Regression was the best classifier (, ); execution time, path length, and depth-axis velocity were primary discriminating features. Second, Hidden Markov Models (K = 6) identified an expert-enriched coordination regime occupied 12.9 percentage points more by experts (Cohen’s d = 0.707), characterised by purposeful pitch-axis exploration versus diffuse yaw–roll noise in novices. Third, linear mixed-effects models confirmed robust expert–novice differences across four kinematic outcomes (all p ≤ 0.018) and practice-related improvement in temporal and spatial efficiency, with no significant between-group difference in improvement rate. These convergent findings support SECMA as a scalable and analytically rigorous platform for camera navigation training and proficiency-based assessment.Keywords: Virtual reality simulation, Laparoscopic camera navigation, Construct validity, laparoscope, Expert–novice classification, Hidden Markov models, Repeated nested cross-validation, Linear mixed-effects models, Surgical training, Kinematic analysisSubject terms: Biomedical engineering, Health care, Medical research, Mathematics and computing Introduction Laparoscopic camera navigation constitutes a foundational yet persistently under-assessed component of minimally invasive surgery (MIS). During laparoscopic procedures the camera operator effectively controls the visual reference frame of the entire surgical team: a stable, correctly oriented, and precisely directed image is a prerequisite for instrument coordination, tissue identification, and safe dissection1,2. Unlike open surgery, the laparoscopic environment eliminates direct tactile feedback and replaces the three-dimensional operative field with a two-dimensional monitor representation, imposing a demanding visuomotor-mapping problem on the camera operator3,4. The challenge is compounded when a angled laparoscope is used. Unlike a scope, the oblique optic decouples instrument orientation from visual field direction: shaft rotation redirects the viewing axis while simultaneously disturbing the horizon, axial translation changes apparent target size, and the operator must continuously reason about the mapping between proximal hand motion and distal visual consequence4,5. Kengen et al.5 demonstrated that performance and learning curves differ systematically across optical angles, confirming that camera handling constitutes a distinct and more demanding psychomotor skill than straight-lens manipulation. Poor control of the angled scope degrades visualisation, disrupts instrument coordination, and increases the cognitive burden of the operative team2,4, yet dedicated training for this specific skill remains comparatively limited and inconsistently implemented. Traditional training for laparoscopic skills relies on box trainers such as the Fundamentals of Laparoscopic Surgery (FLS) programme and on supervised intraoperative exposure. Although these approaches have demonstrated educational value6–8, they share structural limitations that restrict their scalability and their suitability as platforms for camera-navigation assessment: they depend on manual evaluation by instructors, provide sparse or binary outcome metrics, consume clinical resources, and offer limited opportunities for deliberate repetition in isolation from the broader procedural context9–11. Virtual reality (VR) simulators have emerged as a response to these constraints. A now substantial body of evidence, including systematic reviews and meta-analyses, confirms that VR-based surgical training improves measurable technical performance and supports proficiency-based progression in laparoscopic skills7,12–14. The educational value of simulation in this domain derives not from surface resemblance to the clinical environment but from its capacity to reproduce task-defining motor constraints, support deliberate repetition, and enable objective assessment through quantitative metrics6,15,16. For laparoscopic camera navigation specifically, dedicated simulation curricula have shown measurable benefit. The Tulane Trainer17,18 and the EndoTower system19established early proof of concept for simulator-based camera-handling training. Nilsson et al20. subsequently provided randomised trial evidence that simulation-based camera-navigation training significantly improves technical performance relative to no training. Shetty et al.21 demonstrated construct and face validity for a dedicated camera-navigation curriculum on the LapSim platform. Arikatla et al.9 reported face validation of a virtual camera-navigation task trainer targeting the specific demands of horizon stabilisation and directional control. Most recently, Jiongbiao et al.22 confirmed that a structured VR camera-navigation curriculum can reliably prepare new surgical assistants, providing further evidence that the skill is trainable and transferable. Taken together, this body of work establishes the educational rationale for camera-navigation simulation while also highlighting a persistent gap: the absence of portable, low-cost systems that simultaneously preserve the mechanical constraints of angled laparoscopic handling and provide rich objective assessment beyond binary task-completion metrics. High-fidelity commercial simulators such as LapMentor, LapVR, and LapSim incorporate anatomically detailed virtual environments and provide automated performance feedback, but they consistently exceed USD 100,000 in acquisition cost13,23,24. This creates a structural access problem that is especially acute in resource-limited settings and in training programmes at institutions without dedicated skills laboratories25,26. In response, several low-cost simulators have been developed for core laparoscopic skills. VBLaST10 uses a modified physical handle and a PC-based virtual environment. SinaSim27 employs inertial measurement units and custom tracking hardware. SIMISGEST-VR24 incorporates gesture-based interaction for a broader range of psychomotor tasks. These platforms have demonstrated educational utility while reducing cost barriers. However, important practical and functional limitations remain. Systems such as VBLaST10 and SinaSim27 depend on laboratory-based configurations requiring external computing hardware and, in some cases, dedicated haptic or custom tracking devices, which limits portability, deployability, and scalability outside specialised simulation facilities. Moreover, although later developments of VBLaST introduced a camera-navigation exercise, these systems were not designed around the specific mechanical and control demands of laparoscopic camera navigation in a portable standalone format. In particular, they do not combine, in a single low-cost platform, the fulcrum-mediated pivot constraint, axial translation, and the decoupled control of horizon stabilisation and optical redirection that define angled-scope handling as a psychomotor problem distinct from generic instrument manipulation. By contrast, while SIMISGEST-VR24 offers a more accessible interaction paradigm, its design is oriented toward general psychomotor skill acquisition and does not explicitly address camera navigation under angled optics. More recent hybrid approaches, including augmented-reality systems such as LapAR28 and mixed-reality hysteroscopy trainers29, have further shown that portability and educational value are not mutually exclusive. However, these systems typically focus on different anatomical or procedural contexts and do not target the camera-control mechanics addressed in the present work. Table 1 summarises the key differences between SECMA and the two most-cited low-cost systems in this space. The comparison highlights three dimensions that distinguis the present contribution: a mechanically constrained interface that enforces the specific kinematic coupling of camera handling; fully standalone operation with continuous high-frequency telemetry; and a multi-layered analytic framework combining feature-based classification, latent-state modelling, and longitudinal mixed-effects analysis—none of which is present in either prior system. Table 1Comparison of SECMA with representative low-cost laparoscopic VR simulators. VBLaST10 and SinaSim27 are the most widely cited low-cost systems addressing generic laparoscopic skills; neither was designed specifically for camera navigation with mechanically constrained controllers. Entries marked (†) indicate architectural support not yet formally validated in published transfer studies.FeatureVBLaST10SinaSim27SECMA (present work)Interaction deviceHaptic device (PHANToM or equivalent)Custom handheld or IMU-basedCommercial controllers (Oculus Quest 2) mechanically adaptedPhysical constraints (fulcrum effect)Partially enforced (device-dependent)Not physically enforcedMechanically constrained; telescopic shaft with articulated pivotKinematic couplingModerate (hardware-dependent)MinimalExplicit pitch–yaw–axial translation coupling; decoupled horizon stabilisation and optical redirectionCostMedium–high (dedicated haptic hardware)LowLow (commercial VR headset + 3D-printed interface)PortabilityLimited (laboratory setup)ModerateHigh (compact case, kg, no external PC required)Standalone operationNo (PC-dependent)Typically PC-dependentYes (Qualcomm Snapdragon XR2; 100 Hz telemetry)Objective performance metricsBasic (time, error counts, path length)BasicContinuous 6-DoF kinematic data; 29 engineered featuresLatent behavioural modellingNoNoYes (Hidden Markov Models; states)Longitudinal mixed-effects analysisNot reportedNot reportedYes (LMM across practice sessions P2–P4)Primary training focusFLS instrument tasks (peg transfer, suturing)General laparoscopic skills laparoscopic camera navigationTransfer validity establishedPartial (instrument skills)7Not reportedNot yet established (†); construct-related evidence presented here Comparison of SECMA with representative low-cost laparoscopic VR simulators. VBLaST10 and SinaSim27 are the most widely cited low-cost systems addressing generic laparoscopic skills; neither was designed specifically for camera navigation with mechanically constrained controllers. Entries marked (†) indicate architectural support not yet formally validated in published transfer studies. Beyond cost and portability, a second persistent limitation of existing camera-navigation simulators is the coarseness of their performance metrics. The predominant approach has been to use scalar outcome measures — task completion time, number of targets acquired, or binary success/failure — that capture aggregate efficiency but discard the rich spatiotemporal structure of the underlying trajectories30,31. This creates an assessment gap with important practical consequences. Completion time and counts cannot identify how a movement was performed, which behavioral sub-components contribute to expertise differences, or which specific dimensions are most resistant to improvement under brief practice. They also provide limited basis for adaptive feedback because they cannot localise the source of performance deficits within a trial. Recent advances in motion-capture instrumentation and machine-learning analysis have opened a path beyond this limitation. Tracking-based extensions of box trainers and add-on sensing solutions have demonstrated that continuous instrument kinematics can be recorded and analysed objectively in real training environments31–33. In the domain of VR surgical simulation, Winkler-Schwartz et al.34 and Bogar et al.35 showed that machine-learning classifiers trained on kinematic features can reliably discriminate between expertise levels, supporting the use of AI-driven metrics for proficiency assessment. These contributions suggest that the combination of continuous 6-DoF telemetry and interpretable computational modeling can transform a VR simulator from a controlled training environment into an instrumented assessment platform. The technical infrastructure to do so is now accessible at low cost, but it has not yet been systematically applied to the specific case of laparoscopic camera navigation. A further theoretical consideration motivates the present work. The literature on VR effectiveness in training has consistently shown that educational benefit depends not on visual or geometric resemblance to the clinical environment but on whether the simulator preserves the action–effect contingencies that structure the target skill16,36–38. In the specific case of angled laparoscopic camera handling, these contingencies include the non-linear mapping between shaft rotation and visual-field direction, the horizon disturbance induced by roll, and the depth-dependent change in apparent target position. A system that does not reproduce these relationships cannot elicit the specific visuomotor adaptation that constitutes expert camera handling, regardless of how realistic its visual rendering may be. This principle defines the design space for a functionally valid camera-navigation simulator and motivates the mechanical architecture described in the present study. The present study Building on these considerations, we present the design, implementation, and construct validation of SECMA (Sistema de Entrenamiento Portátil para Cirugías de Mínimo Acceso), a portable VR platform for laparoscopic camera navigation training39. The system addresses simultaneously the three limitations identified above. At the hardware level, a custom mechanical interface attaches Oculus Quest 2 controllers to a telescopic shaft with an articulated fulcrum support, reproducing the pivot constraint, axial translation, and decoupled horizon-stabilisation and optical-redirection channels specific to camera handling. At the software level, independent rotational kinematics implemented in Unity accurately model the oblique optical geometry of the angled scope. At the analytical level, continuous 6-DoF telemetry is processed through a three-layer computational framework: (i) supervised binary classification of expertise level using repeated nested cross-validation; (ii) unsupervised latent-state inference via Hidden Markov Models applied to raw trajectory sequences; and (iii) longitudinal mixed-effects modeling of kinematic outcomes across practice sessions. To assess construct validity, a structured six-target navigation task was administered to 20 board-certified surgeons (Expert group) and 18 medical students (Novice group) across four practice sessions. Engineered kinematic features — including temporal efficiency, path economy, per-axis velocity statistics, angular-speed descriptors, and Euler-angle statistics derived from quaternion time series — were extracted from interpolated trajectories and used as inputs to all three analytical layers. The following sections describe the mechanical design and computational kinematics of the simulator, the experimental protocol, the feature engineering methodology, and the statistical frameworks applied. The results are then presented and discussed in relation to the literature on camera-navigation simulation, low-cost VR training, and AI-based surgical performance assessment. The overarching aim is to demonstrate that a portable, mechanically informed, and telemetry-rich VR platform can generate expertise-sensitive kinematic organisation and short-horizon practice effects in a task that directly reflects the motor demands of laparoscopic camera control. Methods Laparoscopic camera system: design and kinematics The laparoscopic camera system, depicted in Fig. 1, is designed for precise spatial control and optimal visualisation within confined surgical environments. At its proximal end, the system integrates a compact camera head unit coupled to the main laparoscopic cylinder: a rigid shaft housing a rod-lens optical system that is inserted through a trocar to access the abdominal cavity. Crucially, the internal optical configuration incorporates an oblique lens—commonly —which generates an angled field of view relative to the instrument’s longitudinal axis, enabling visualisation of structures lateral to the instrument tip without repositioning the shaft40. Components and kinematic degrees of freedom of the laparoscopic camera system. Roll (R), Yaw (Y), Pitch (coupled to Y), Axial Translation (L/APA), and independent fibre-optic connector rotation (F) are indicated. See main text for functional description. The shaft also features a fibre optic cable connector, positioned perpendicularly, which transmits illumination from an external cold-light source into the surgical field. Independent rotation of this connector around the main shaft axis constitutes a separate degree of freedom (DoF) that redirects the oblique viewing angle laterally without disturbing the primary horizon orientation of the laparoscope. As illustrated in Fig. 1, the motion of the laparoscope is mechanically constrained by the trocar insertion point, which acts as a fixed pivot. This constraint defines four kinematic degrees of freedom: pitch (vertical angulation) and yaw (horizontal angulation) together aim the instrument tip; roll (rotation about the longitudinal axis) redirects the oblique optical field; and axial translation adjusts the working distance. Additionally, the fibre-optic connector can be rotated independently around the shaft, redirecting the illumination and viewing axis laterally without disturbing the primary shaft orientation. The interaction of roll with the oblique lens geometry is the principal source of visuospatial complexity in camera handling: any shaft roll simultaneously displaces the target in the image and tilts the horizon, requiring active compensation to maintain a stable visual reference. Figure 2 illustrates how the three principal rotational manipulations alter the visual field of an angled laparoscope. Panel (A) shows how increasing lens inclination (, , ) progressively displaces a reference target toward the lower edge of the frame, a direct consequence of the oblique optical axis. Panel (B) adds a shaft roll: the visual horizon rotates accordingly across all three lens types, illustrating the intraoperative disorientation that arises from uncompensated roll and the active stabilisation effort it demands. Panel (C) combines that roll with an additional fibre-optic connector rotation, redirecting the field laterally while preserving the tilted horizon — a manoeuvre that enables peripheral exploration without repositioning the entire instrument. Effect of lens inclination, shaft roll, and optical-connector rotation on the laparoscopic visual field (see main text for description of panels A–C). Simulation of a 30-degree angled laparoscopic camera system To emulate the functional behaviour of a laparoscopic camera, a mechanical interface was developed that transforms Oculus Quest 2 controllers into a physically constrained laparoscopic instrument. The core assembly comprises a telescopic metallic shaft integrated with two ball-and-socket articulated joints mounted on a compact, transportable base platform. This configuration reproduces the mechanical constraints and degrees of freedom experienced during real laparoscopic procedures without requiring a full operating-tower setup. The complete simulator is a self-contained, portable unit housed in a rigid carrying case with external dimensions of cm and an assembly weight of approximately 4.4 kg, enabling deployment in training environments without a fixed installation. The manipulable interface—comprising the telescopic shaft, ball-and-socket joints, and attached Oculus controllers—weighs approximately 0.5 kg, which falls within the typical range of clinical laparoscopic camera assemblies (≈ 0.3–0.8 kg for the telescope and camera head combined), providing a comparable physical load during manipulation. The telescopic shaft offers an adjustable working length of approximately 26–60 cm, spanning and exceeding the nominal insertion length of standard laparoscopes (≈ 30–40 cm) to accommodate different training configurations. The design does not aim to replicate the exact geometry or mass distribution of a clinical camera head; rather, it prioritises the reproduction of task-defining functional constraints—the fulcrum effect, coupled pitch–yaw–axial translation, and the decoupled control of optical redirection and horizon stabilisation—that are necessary and sufficient to elicit the visuomotor demands of laparoscopic camera navigation. As illustrated in Fig. 3, panel (A) provides a side-by-side comparison of the simulated and real laparoscopic systems, emphasising the functional mapping between them. Control assignments are partitioned between the two Oculus controllers according to the clinical instrument’s decoupled control logic. The right controller, attached via a snap-fit mechanism at the proximal end of the shaft, governs horizon stabilisation: it simulates the camera operator’s ability to maintain a level visual reference by adjusting roll rotation of the camera about its longitudinal axis. The left controller, positioned along an intermediate segment of the shaft, governs optical redirection: its rotation changes the viewing axis by turning the optic, reproducing the visual-field changes induced by shaft rotation in the clinical instrument. Simulation of laparoscope function using the VR interface: comparative design and kinematics. (A) Comparative view of the simulated (right) and real (left) laparoscopic systems, highlighting the functional mapping of components and control points. (B) Emulation of camera control including pitch, yaw, roll, and horizon stabilisation through the right Oculus controller. (C) Fulcrum-effect simulation through coordinated axial translation and angular articulation at the instrument’s simulated insertion point. The telescopic shaft and ball-socket joints together reproduce the fulcrum effect: the combination of axial translation and angular articulation at the simulated trocar point creates the inverted motion-to-visual mapping characteristic of laparoscopic instrument handling. This mechanical constraint is fundamental to the educational validity of the platform, as it forces the user to reason about proximal-to-distal motion inversion during target acquisition and horizon stabilisation4,5. Figures 3(B) and (C) detail, respectively, the emulation of pitch, yaw, and roll adjustments for the angled camera and the mechanical reproduction of the fulcrum effect at the simulated trocar insertion point. To complement the mechanical and kinematic description of the simulator, we assessed expert-perceived face validity using a post-task questionnaire administered to n = 15 board-certified surgeons and sub-specialists, all with more than 100 laparoscopic procedures performed. The questionnaire comprised 11 items covering physical realism, navigation utility, feedback quality, motivation/adoption, and graphical interface/calibration. Overall ratings were favourable, with a mean battery score of 4.33 out of 5 and acceptable internal consistency (Cronbach α = 0.77). The highest-rated domain was Navigation Utility (mean = 4.64, 95% bootstrap CI: 4.44–4.82), indicating unanimous expert agreement that the task was relevant for localisation, centering, horizon maintenance, and approach skills with a camera. Physical Realism was also rated favourably (mean = 4.30, 95% CI: 4.10–4.50), including instrument manipulation, camera control, and horizon stabilisation. The lowest-rated domain was Graphical Interface and Calibration (mean = 3.77, 95% CI: 3.53–4.00), mainly due to the calibration process, which several respondents indicated benefited from brief operator guidance. Full item wording, ordinal mappings, individual expert scores, and domain-level summaries are provided in Supplementary Table S1 and Supplementary Figure S1. Universal connector and mechanical integration To enable rapid reconfiguration between training modules, a modular universal connector system was developed. Figure 4 illustrates the four assembly stages: panel (A) shows the universal connector and its pressure-fit integration with the Oculus controller, reinforced by screws engaging the female snap-fit support; panel (B) shows the coupled assembly ensuring secure alignment during simulated procedures; panel (C) presents the telescopic tube with snap-fit mounts that allow flexible controller positioning for both primary camera control and optical-fibre rotation; and panel (D) demonstrates the complete mechanical chain that transmits hand motions to the simulated laparoscope, enabling realistic emulation of pivoting and rotational dynamics. Modular universal-connector interface of the SECMA platform (panels A–D; see main text for assembly description). Figure 5 shows the fully assembled platform and a faculty surgeon performing the navigation task, demonstrating both the practical viability of the system and the embodied character of the simulated interaction. SECMA simulation platform in use. (A) Fully assembled simulator with integrated Oculus controllers, universal connectors, and mechanical interface. (B) A faculty surgeon performs the camera navigation task, demonstrating use in construct validation and training scenarios. Computational simulation of endoscopic camera dynamics To accurately reproduce laparoscopic camera behaviour within the Unity virtual environment, a precise kinematic model was developed. Figure 6 illustrates the control architecture. Kinematic architecture of the virtual camera system in the SECMA simulator. (A) Reference frames of the left and right Oculus controllers relative to the shaft axis. (B) Computation of camera optical rotation via projection of the left controller forward vector onto the plane perpendicular to the shaft. (C) Determination of horizon stabilisation angle from the signed angular displacement of the right controller right-axis vector relative to the fixed shaft reference vector. The simulator defines three principal reference vectors (Fig. 6A): the longitudinal axis of the laparoscopic shaft (gray), the global up-vector (orange) orthogonal to the simulator base, and a cross-product reference vector . Each Oculus controller has a fixed body frame composed of unit vectors (right), (up), and (forward), where superscript denotes left and right controllers respectively. Lens orientation control (left controller). The forward vector of the left controller is projected onto the plane perpendicular to to produce the normalised projected vector which encodes the rotation direction of the camera optics. The virtual camera lens is oriented using , accurately emulating changes in visual direction induced by the oblique lens configuration40. Horizon stabilisation (right controller). The angular displacement ϕ of the right controller’s right-axis vector with respect to the fixed reference vector , measured about via the signed-angle function, encodes the current horizon orientation: The final camera orientation is obtained by composing two rotations: alignment of the camera forward axis with , followed by an axial rotation by -ϕ about the camera’s local forward axis. This decoupled structure separates horizon stabilisation from optical redirection, matching the two independent control streams of the physical instrument. All rotations are captured through unit-quaternion representations sampled at uniformly spaced timestamps. The complete control update executed at each Unity frame is summarised in Algorithm 1. Kinematic Control of the Laparoscope and Horizon Stabilisation Task description: virtual navigation with a laparoscopic camera Participants performed a camera navigation task using the simulated laparoscope within a three-dimensional virtual environment. The scene (Fig. 7A) comprised a bounded workspace containing six lidless containers positioned at varying locations and orientations, each displaying a red circular target with a horizontal alignment stripe. Six-target camera navigation task in the SECMA virtual simulator. (A) Workspace layout. (B) Real-time verification interface (centring ring and horizon alignment band). (C) Confirmation marker after successful target completion. See main text for task criteria and data-acquisition details. Task completion required participants to approach each target sequentially and simultaneously satisfy two conditions for at least three seconds: (i) Centring condition: the circular target must remain within the camera’s verification ring (Fig. 7B), ensuring correct translational positioning. (ii) Horizon condition: the target stripe must remain within the alignment band of the same interface, ensuring that horizon stabilisation is maintained concurrently with centring. Both criteria were enforced in real time by the software interface, which continuously monitored camera pose using timestamped position and orientation records. Upon successful completion, a visual confirmation marker was displayed (Fig. 7C) and the simulator archived the corresponding kinematic data. The dual-condition design was deliberate: the simultaneous centring-and-horizon requirement taxes the decoupled control structure of the mechanical interface, obliging participants to manage optical redirection (left controller) and horizon stabilisation (right controller) concurrently. This directly replicates the fundamental visuomotor challenge of laparoscopic camera handling in clinical practice2,4,5. Construct validity: experimental design and analytical framework Participants and experimental design To evaluate the construct validity of the camera navigation task, a controlled experiment was conducted with two independent groups: Expert group (n = 20): board-certified surgeons with demonstrable experience in laparoscopic navigation. Expert group (n = 20): board-certified surgeons with demonstrable experience in laparoscopic navigation. Novice group (n = 18): medical students without prior laparoscopic experience. Novice group (n = 18): medical students without prior laparoscopic experience. Participants performed the six-target navigation task in a standardised three-dimensional virtual environment using the SECMA simulator. Each participant completed four practice sessions after a brief orientation period; Practice 1 was discarded to exclude familiarisation effects, leaving three analysed sessions (Practices 2, 3, and 4). Motion was constrained to the simulated trocar pivot to mimic real laparoscopic kinematics. Telemetry acquisition and interpolation Spatial and rotational data were captured continuously through a rigidly attached pose marker on the virtual camera, recording three spatial coordinates (x, y, z) and four unit-quaternion components at each timestamp. The nominal sampling rate of the Oculus Quest 2 controller tracking system was 80–100 Hz. Because trial durations differed across participants and sessions, all time series were linearly interpolated to a fixed length equal to the mean number of frames observed across expert trials. Quaternion components were re-normalised after interpolation to ensure unit-norm constraints were preserved. From the interpolated sequences, a set of engineered kinematic and rotational features was derived; complete definitions, units, and operational descriptions for all features are provided in Table 2. The feature set spans five conceptual groups: (i) temporal and path-economy descriptors (ExecutionTime, PathLength, PathEfficiency, ExcessPath, DirectDisplacement); (ii) per-axis velocity and displacement statistics (AvgVelocityX/Y/Z, MaxDispX/Y/Z, StartX/Y/Z); (iii) quaternion rotation component maxima (MaxRotation X/Y/Z/W); (iv) Euler-angle statistics in the XYZ extrinsic convention (MaxAngleX/Y/Z, MinAngleX/Y/Z, MeanAngle X/Y/Z, StdAngleX/Y/Z); and (v) instantaneous angular-speed statistics derived from successive quaternion differences (AngularP95, AngularMean, AngularStd). Table 2Definitions of engineered features extracted from interpolated laparoscopic camera trajectories. All temporal features are derived from the interpolated time axis. Euler angles follow the XYZ extrinsic convention and are reported in degrees; angular speeds are computed from successive unit-quaternion differences and reported in rad s. Each group expands into per-axis variants where indicated (X, Y, Z or X, Y, Z, W).Feature (group)UnitsOperational definition(a) Temporal and path-economy descriptorsExecutionTimesTotal trial duration from the first to the last interpolated timestamp.PathLengthsim. unitsCumulative sum of Euclidean step distances along the 3-D trajectory: .DirectDisplacementsim. unitsStraight-line distance between trial start and end positions: .PathEfficiency–Ratio ; a value of 1 denotes perfectly straight motion; lower values indicate redundant trajectory.ExcessPathsim. units; excess distance relative to the straight-line path.(b) Per-axis velocity and displacement statisticsAvgVelocity{X|Y|Z}sim. units sMean absolute velocity along each spatial axis: .MaxDisp{X|Y|Z}sim. unitsRange (max − min) of camera position along each axis.Start{X|Y|Z}sim. unitsCamera position at trial onset along each axis.(c) Quaternion rotation component extremaMaxRotation{X|Y|Z|W}–Maximum value of each unit-quaternion component observed during the trial; captures the extreme rotational excursion of the camera.(d) Euler-angle statistics (XYZ extrinsic)MaxAngle{X|Y|Z}degMaximum Euler angle about each axis observed during the trial.MinAngle{X|Y|Z}degMinimum Euler angle about each axis observed during the trial.MeanAngle{X|Y|Z}degTrial-mean Euler angle about each axis.StdAngle{X|Y|Z}degStandard deviation of the Euler angle about each axis; proxy for rotational variability and control steadiness.(e) Instantaneous angular-speed statisticsAngularP95rad s95th percentile of instantaneous angular speed computed from successive quaternion differences: . Captures peak angular agitation.AngularMeanrad sTrial-mean instantaneous angular speed.AngularStdrad sStandard deviation of instantaneous angular speed; reflects within-trial speed fluctuation. Definitions of engineered features extracted from interpolated laparoscopic camera trajectories. All temporal features are derived from the interpolated time axis. Euler angles follow the XYZ extrinsic convention and are reported in degrees; angular speeds are computed from successive unit-quaternion differences and reported in rad s. Each group expands into per-axis variants where indicated (X, Y, Z or X, Y, Z, W). Supervised classification under repeated nested cross-validation Binary classification of Group (Expert vs. Novice) was performed using seven supervised classifiers: Logistic Regression, Support Vector Machine (SVM, RBF kernel), Random Forest, Gradient Boosting, Decision Tree, Naive Bayes, and k-Nearest Neighbours (KNN). To obtain unbiased and stable performance estimates a repeated nested cross-validation (RNCV) design was adopted. The outer loop consisted of a 5-fold stratified cross-validation repeated 50 times with different random partitions, yielding 250 independent test folds per model. Within each outer training fold, hyperparameters were selected via a 4-fold inner grid search optimising the score. The outer test fold was never exposed during hyperparameter optimisation. Continuous features were standardised within each outer training fold using z-score normalisation; the categorical variable Practice was one-hot encoded; missing values were imputed with fold-wise medians (numeric) or modes (categorical). Performance was quantified by accuracy, balanced accuracy, precision, recall, score, and AUC across the 250 outer folds. To formally compare models, paired Wilcoxon signed-rank tests were applied to per-fold vectors (n = 250 pairs), with Benjamini–Hochberg (BH) correction for six simultaneous comparisons. Standardised Logistic Regression coefficients — computed on features already z-scored by the pipeline’s StandardScaler and therefore directly comparable across features — were used to characterise feature importance from the top-tier classifier, providing signed importance (positive coefficient → Novice class; negative coefficient → Expert class) as reported in Fig. 11. Performance results are reported in Table 3. Top 15 features for expert vs. novice classification ranked by the absolute value of the standardised Logistic Regression coefficient. Bars are coloured by conceptual group: temporal/path, angular/rotational, spatial/positional. Positive values (right) indicate that higher feature values are associated with the Novice class; negative values (left) are associated with the Expert class. The model was refit on the full dataset after inner-loop (5-fold) grid search. Coefficient values are annotated on each bar. Table 3Expert vs. novice classification performance under repeated nested cross-validation. Outer loop: 5-fold × 50 repetitions = 250 test folds per model. Inner loop: 4-fold grid search on training data only. All values are mean ± SD across the 250 outer folds.†Logistic Regression was selected as the reference model (highest mean); see Table 4 for pairwise statistical comparisons.ModelAccuracyBal. AccuracyPrecisionRecallF1 ScoreAUCLogistic Regression†SVMRandom ForestKNNGradient BoostingNaive BayesDecision Tree Expert vs. novice classification performance under repeated nested cross-validation. Outer loop: 5-fold × 50 repetitions = 250 test folds per model. Inner loop: 4-fold grid search on training data only. All values are mean ± SD across the 250 outer folds.†Logistic Regression was selected as the reference model (highest mean); see Table 4 for pairwise statistical comparisons. Linear mixed-effects models for longitudinal performance To quantify learning progression across practice sessions and to test whether practice-related improvement differed between groups, a linear mixed-effects model (LMM) was fitted for each of four kinematic outcomes: , , PathEfficiency, and AngularP95. Log transformations were applied to the first two outcomes to reduce right-skew and stabilise residual variance; the transformed scale enables percentage-change interpretation of coefficients. The model specification for each outcome of participant i at practice session was: where β are population-level fixed effects, is a participant-level random intercept capturing stable baseline heterogeneity, and is the within-participant residual. Practice 2 and the Expert group were used as reference levels. Models were fitted using maximum-likelihood estimation via the statsmodels MixedLM implementation in Python, with sequential optimiser fallback (lbfgs → bfgs → cg → powell → nm) to ensure convergence. All four models converged successfully. The intraclass correlation coefficient (ICC), defined as , was computed from the estimated variance components to quantify the proportion of total outcome variance attributable to stable between-participant differences. Fixed-effect estimates, 95% confidence intervals, and variance components are reported in Table 5; estimated marginal means and individual learning trajectories are shown in Figs. 12 and 13. Table 5Fixed-effect estimates from linear mixed models for four kinematic outcomes across practice sessions 2–4. Reference levels: Practice 2 and Expert group. Log-scale coefficients correspond to multiplicative effects on the original scale. Bold values indicate p<0.05.Termlog(Exec. Time)log(Path Length)Path EfficiencyAngular P95(SE)/p(SE)/p(SE)/p(SE)/pIntercept (P2, Expert)Practice 3-0.143\ (0.091)+0.020\ (0.023)+0.139\ (0.160)Practice 4+0.045\ (0.023)-0.029\ (0.160)Group: NoviceP3 × Novice+0.019\ (0.124)+0.008\ (0.128)+0.013\ (0.033)+0.074\ (0.226)P4 × Novice+0.073\ (0.130)+0.110\ (0.134)-0.004\ (0.034)+0.175\ (0.236)0.08180.06570.00220.22460.06620.07020.00470.2181ICC0.5530.4840.3220.507; ; . Fixed-effect estimates from linear mixed models for four kinematic outcomes across practice sessions 2–4. Reference levels: Practice 2 and Expert group. Log-scale coefficients correspond to multiplicative effects on the original scale. Bold values indicate p<0.05. ; ; . Estimated marginal means for four kinematic outcomes across practice sessions P2–P4 by group. Lines represent model-predicted means and points correspond to individual observations. Individual trajectories for across sessions P2–P4. Thin lines represent participants and thick lines indicate group means. Hidden Markov modeling of laparoscopic camera trajectories To uncover latent temporal structure in task execution that is not captured by trial-level aggregate features, the full interpolated trajectories were modelled as multivariate time series using Hidden Markov Models (HMMs) with multivariate Gaussian emissions41,42. Each observation comprised the three spatial coordinates and four unit-quaternion components , standardised across all sequences prior to fitting. Formally, let denote an observed sequence and the corresponding latent state sequence with . The HMM is parameterised by: An initial state distribution , where . An initial state distribution , where . A transition matrix , where . A transition matrix , where . Emission distributions: for each , , where and are the emission mean and full covariance matrix. Emission distributions: for each , , where and are the emission mean and full covariance matrix. The model parameters were learned via the Expectation–Maximisation (EM) algorithm43, implemented in the hmmlearn library44 with numerically stable log-likelihood estimation and Viterbi decoding. Each model was trained for 300 EM iterations with a fixed random seed. The number of free parameters for a K-state Gaussian HMM with full d-dimensional covariance is yielding for with d = 7. The optimal number of states was selected by evaluating BIC and AIC over ; the model selection results are reported in Table 6. The selected model (K = 6) was decoded using the Viterbi algorithm to assign a latent state label to each frame. Table 6HMM model selection criteria for hidden states. Models used Gaussian emissions with full covariance structure (d = 7 dimensions, 300 EM iterations). The number of free parameters p is computed via Equation (5). The selected model (K = 6, bold) achieves the lowest BIC and AIC within the evaluated range.KpAICBIC2731, 414, 288.951, 414, 985.1231131, 173, 989.311, 175, 066.9541551, 029, 877.711, 031, 355.8951991, 001, 580.336245 HMM model selection criteria for hidden states. Models used Gaussian emissions with full covariance structure (d = 7 dimensions, 300 EM iterations). The number of free parameters p is computed via Equation (5). The selected model (K = 6, bold) achieves the lowest BIC and AIC within the evaluated range. Per-participant state-occupancy proportions (fraction of frames in each state) were computed and compared between groups using Welch two-sample t-tests with BH correction across the K = 6 simultaneous comparisons. State-level feature profiles were characterised by computing the z-scored mean of each engineered feature within each state, visualised as a heatmap (Fig. 15). All features used in this profiling are defined in Table 2. Segments with fewer than five frames were excluded from state-level feature computation to avoid unstable estimates. Pairwise group comparisons within the expert-enriched state were performed using Welch t-tests with BH correction across the 13 compared features, as reported in Table 8 and Fig. 16. Empirical transition matrices were additionally computed separately for each group to characterise inter-state dynamics and self-transition persistence (Fig. 17). Standardised (z-score) feature profiles across the six HMM latent states (K = 6). Each cell encodes the z-score of the state-level mean of a kinematic or rotational feature relative to the cross-state grand mean; warm colours indicate values above, cool colours below the cross-state mean. Complete definitions and units for all features are provided in Table 2. State 5 (bottom row) is distinguished by the highest ExcessPath (z = +1.89), elevated StdAngleZ (z = +1.18) and StdAngleY (z = +1.78), strongly negative MaxRotationX (z = -2.02), and negative StartX (z = -1.61), suggesting a phase of spatially extended, rotationally variable camera motion originating from a leftward initial position. Table 8Group comparisons of kinematic features within HMM State 5 (expert-enriched state), using Welch t-tests with Benjamini–Hochberg correction across 13 comparisons. Δ = Expert mean − Novice mean; positive values indicate experts exceed novices, negative values the reverse. Cohen’s d quantifies the standardised effect size (sign matches Δ). Complete feature definitions are given in Table 2. No comparison reaches after correction; however, several features show medium-to-large effect sizes (|d| ≥ 0.5), warranting attention in the context of limited statistical power.FeatureExpertNoviceΔWelch tpdPathLength1.972 ± 1.4361.595 ± 0.340+0.3770.5460.725+0.361PathEfficiency0.072 ± 0.0480.039 ± 0.007+0.0330.1180.539+0.983ExcessPath1.869 ± 1.4311.532 ± 0.316+0.3370.5830.725+0.325AvgVelocityZ0.013 ± 0.0030.011 ± 0.005+0.0020.6690.725+0.493AngularP951.111 ± 0.2330.912 ± 0.368+0.2000.5800.725+0.648MaxDispX0.151 ± 0.0150.161 ± 0.018-0.009-0.6560.6040.725-0.557MaxDispY0.113 ± 0.0550.079 ± 0.010+0.0340.1660.539+0.859MaxDispZ0.127 ± 0.0180.120 ± 0.017+0.0070.6640.725+0.411StdAngleX37.530 ± 19.37714.877 ± 3.865+22.6540.0230.301+1.621StdAngleY31.003 ± 4.45737.336 ± 3.496-6.333-2.1170.1640.539-1.581StdAngleZ81.263 ± 38.732125.228 ± 39.734-43.966-1.3880.3270.725-1.121StartY0.021 ± 0.1300.008 ± 0.001+0.0130.8060.806+0.137StartZ0.378 ± 0.0620.338 ± 0.087+0.0400.6340.725+0.525 Group comparisons of kinematic features within HMM State 5 (expert-enriched state), using Welch t-tests with Benjamini–Hochberg correction across 13 comparisons. Δ = Expert mean − Novice mean; positive values indicate experts exceed novices, negative values the reverse. Cohen’s d quantifies the standardised effect size (sign matches Δ). Complete feature definitions are given in Table 2. No comparison reaches after correction; however, several features show medium-to-large effect sizes (|d| ≥ 0.5), warranting attention in the context of limited statistical power. Distribution of kinematic and rotational features within HMM State 5 (expert-enriched state), stratified by group. Box plots show median, IQR, and 1.5×IQR whiskers; dots denote outliers. No feature reaches after BH correction, but three features exhibit large effect sizes (|d| ≥ 1): StdAngleX (d = +1.621, experts more variable), StdAngleY (d = -1.581, novices more variable), and StdAngleZ (d = -1.121, novices more variable). Complete feature definitions are provided in Table 2. Empirical transition probability matrices for Expert (A) and Novice (B) groups (K = 6). Cell values are row-normalised transition probabilities; the red box highlights the self-transition of the expert-enriched State 5. Both groups exhibit highly self-absorbing dynamics (diagonal ≥ 0.97), consistent with sustained behavioral episodes. The self-transition probability of State 5 is vs. , indicating virtually identical dwell-time persistence in both groups once this state is entered; group differences in State 5 therefore reflect occupancy frequency (entry rate) rather than dwell time. Results Classification performance and feature importance Primary group comparisons Before proceeding to the machine-learning analyses, we confirmed that the two groups differed significantly on the primary kinematic indicators. Novices exhibited a mean execution time of 134.87 s versus 70.46 s for experts. After a Box–Cox transformation (λ = -0.3727), a two-sample t-test confirmed a significant difference (p < 0.001). Path length was likewise longer in novices (3.67 vs. 2.28 simulator units); significance was established after a Yeo–Johnson transformation (λ = -0.9972, p < 0.001). Engineered feature set Spatial and rotational time series (3-D positions plus unit quaternions) captured at each interpolated time-step were transformed into a fixed-length feature vector per trial. The feature set spans five conceptual groups: (i) temporal and path-economy descriptors; (ii) per-axis velocity and displacement statistics; (iii) quaternion rotation component extrema; (iv) Euler-angle statistics in the XYZ extrinsic convention; and (v) instantaneous angular-speed statistics derived from successive quaternion differences. Table 2 provides a complete definition of each feature group, including units and a precise operational description. Repeated nested cross-validation design To obtain unbiased and stable estimates of classifier performance we adopted a repeated nested cross-validation (RNCV) design. The outer loop consisted of a 5-fold stratified cross-validation repeated 50 times with different random partitions, yielding 250 independent test folds per model. Within each outer training set, hyperparameters were selected via a 4-fold inner grid search optimising the score. Crucially, the outer test fold was never seen during hyperparameter optimisation, guaranteeing that reported metrics are true out-of-sample estimates. The large number of outer folds (250 per model) also ensures that performance estimates are stable with respect to the random partition of the small dataset, as confirmed empirically in Fig. 9. estimate stability across the 250 outer folds of the repeated nested cross-validation for the best classifier (Logistic Regression). The solid line is the cumulative running mean; the shaded band covers ±1 SD. Vertical grid lines mark repetition boundaries (every 5 folds). The estimate stabilises after approximately 50 folds and converges to (dashed grey line), validating the adequacy of the 50-repetition design for this sample size. Seven classifiers were evaluated: Logistic Regression, Support Vector Machine (SVM, RBF kernel), Random Forest, Gradient Boosting, k-Nearest Neighbours (KNN), Naive Bayes, and Decision Tree. Continuous features were standardised within each outer training fold; Practice was one-hot encoded. Missing values were imputed with fold-wise medians (numeric) or modes (categorical). Classifier performance Table 3 reports mean ± SD performance across all 250 outer folds. All models substantially exceeded chance (), confirming that the engineered feature set carries strong discriminative information. Logistic Regression achieved the highest mean score (0.818 ± 0.093) and the highest AUC (0.904 ± 0.061), with balanced accuracy (0.825 ± 0.083) closely matching raw accuracy, indicating that performance was not inflated by class-imbalance artefacts. SVM reached comparable mean (0.817 ± 0.132) and a median AUC of 0.920, while Random Forest attained and AUC = 0.900 ± 0.064. Figure 8(A) illustrates the full distribution across 250 folds for each model, revealing that the three top-performing classifiers—Logistic Regression, SVM, and Random Forest—occupy largely overlapping interquartile ranges, while Decision Tree and Naive Bayes show substantially wider variance. Panel (B) confirms consistent performance across all five metrics for the top-tier models. Expert vs. novice classification performance under repeated nested cross-validation (250 outer folds per model).(A) Box-plot distribution of scores across all 250 outer test folds. Whiskers extend to 1.5 IQR; dots denote outlier folds. The dashed line marks chance level (). (B) Multi-metric bar chart (mean ± SD) showing AUC, balanced accuracy, precision, recall, and for all seven classifiers. Logistic Regression attains the highest mean (0.818 ± 0.093) and AUC (0.904 ± 0.061). Statistical model comparison To rigorously test whether Logistic Regression outperforms the remaining classifiers—as requested by Reviewer 1—we applied paired Wilcoxon signed-rank tests on the per-fold vectors (n = 250 pairs per comparison), with Benjamini–Hochberg (BH) correction for multiple testing across the six comparisons (Table 4). Table 4Paired Wilcoxon signed-rank tests comparing Logistic Regression (reference) against each alternative classifier on per-fold scores (n = 250 fold pairs). p-values are BH-corrected for six simultaneous comparisons. Positive indicates that the reference model outperforms the comparator; bold rows indicate statistically significant results ().ComparatorMean Median Wp-value (BH)Sig.Decision Tree0.0990.0973175.5<0.001<0.001YesNaive Bayes0.0740.0834994.5<0.001<0.001YesGradient Boosting0.0490.0425210.0<0.001<0.001YesKNN0.0410.0337981.0<0.001<0.001YesRandom Forest0.0090.0009573.00.0880.105NoSVM0.0010.0007594.00.1070.107No Paired Wilcoxon signed-rank tests comparing Logistic Regression (reference) against each alternative classifier on per-fold scores (n = 250 fold pairs). p-values are BH-corrected for six simultaneous comparisons. Positive indicates that the reference model outperforms the comparator; bold rows indicate statistically significant results (). Logistic Regression significantly outperformed Decision Tree (, ), Naive Bayes (, ), Gradient Boosting (, ), and KNN (, ). The differences with respect to Random Forest () and SVM () did not reach statistical significance after correction, indicating that these three classifiers form a statistically equivalent top tier. This finding supports the conclusion that linear separability is sufficient to discriminate groups in the engineered feature space—a result of practical interest, as it implies that the expert/novice distinction is robustly encoded in the kinematic features without requiring complex non-linear decision boundaries. Classifier reliability and error characterisation Figure 9 shows the cumulative mean of Logistic Regression as a function of the number of completed outer folds. The estimate stabilises rapidly—converging within approximately 50 folds—and remains virtually constant over the remaining 200 folds (, grey dashed line), confirming that 250 outer folds provide a reliable and reproducible performance estimate despite the modest sample size. Figure 10 shows the confusion matrix for Logistic Regression pooled across all 250 outer folds (4,875 test instances in aggregate). The classifier correctly identified 83% of expert trials (sensitivity) and 82% of novice trials (specificity), with symmetric error rates indicating no systematic bias toward either class. The near-equal diagonal in the row-normalised panel confirms that performance is not driven by a class-imbalance artefact. Aggregated confusion matrix for Logistic Regression, pooled across all 250 outer folds ( test-set assignments). (A) Raw prediction counts. (B) Row-normalised matrix showing per-class rates. The classifier correctly identifies 83% of expert trials and 82% of novice trials, with symmetric error rates confirming the absence of class-imbalance artefacts. Feature importance Because Logistic Regression achieved the highest mean and AUC among all evaluated models (Table 3), feature importance was characterised directly from the coefficients of the fitted LR model rather than from a separate surrogate. After refitting with inner-loop hyperparameter selection on the full dataset, the coefficients of the liblinear solver operate on features that have already been z-scored by the pipeline’s StandardScaler; the resulting standardised coefficients are therefore directly comparable across features and carry a meaningful direction: a positive coefficient increases the log-odds of the Novice class, while a negative coefficient increases the log-odds of the Expert class. Figure 11 displays the top 15 features ranked by . The features with the largest positive coefficients—ExecutionTime, PathLength, and ExcessPath—confirm that novice performance is characterised by longer task duration, more tortuous trajectories, and a higher proportion of redundant motion relative to the straight-line displacement. AvgVelocityZ, by contrast, appears with a positive sign but moderate magnitude, indicating that novices show higher mean depth-axis velocity—not slower movement, but less controlled and more erratic insertion and withdrawal along the laparoscope’s longitudinal axis. This finding is clinically relevant: in 30-degree laparoscopy, uncontrolled axial motion frequently induces unintended horizon shifts, destabilising the visual field. On the negative (Expert) side, PathEfficiency and StartY carry the largest absolute coefficients. The first reflects that expert trajectories are geometrically more economical—closer to the direct path between targets. The second suggests that experts position the camera at a consistently higher y-coordinate at trial onset, a spatial strategy that may afford a wider working angle and reduce the need for subsequent corrective manoeuvres. Several angular descriptors (StdAngleX, MeanAngleZ) also appear with substantial negative coefficients, indicating that reduced rotational variability and a characteristic mean camera orientation are markers of expert-level motor control. The sign and magnitude of these coefficients converge with the primary group comparisons and provide a mechanistically coherent picture: expert laparoscopic camera navigation is distinguished by a combination of temporal efficiency, path economy, deliberate camera positioning, and fine rotational control. This interpretable structure—recoverable from a linear model—supports the clinical face validity of the feature set and offers concrete, quantifiable targets for adaptive training feedback. Summary of construct-validity evidence Taken together, the three layers of evidence—significant primary group comparisons, robust classifier performance (, ) sustained over 250 independent test folds, and statistically validated superiority of the best model over four out of six alternatives—provide strong, internally consistent support for the construct validity of the SECMA camera-navigation task. The finding that a linear classifier suffices, and that the most discriminating features relate to task efficiency (time, path) and depth control (velocity along z), offers concrete and actionable targets for training feedback design. Longitudinal analysis of performance with linear mixed models To quantify learning progression across repeated practice sessions and to assess whether group differences in performance were sustained or moderated by practice, we fitted linear mixed-effects models (LMMs) for four kinematic outcomes: , , PathEfficiency, and AngularP95. Log-transformations were applied to ExecutionTime and PathLength to reduce right-skewness and stabilise residual variance. Importantly, this transformation preserves the relative ordering of effects while allowing coefficients to be interpreted in terms of percentage changes. All four features are defined in Table 2, and model specification followed Eq. 3. In this framework, outcomes were modelled as a function of practice session, group, and their interaction, with participant-specific random intercepts accounting for baseline heterogeneity. Residual variability captured within-participant fluctuations across sessions. Practice 2 and the Expert group were used as reference levels. All models converged successfully under maximum-likelihood estimation using statsmodels MixedLM. Fixed effects Table 5 reports the fixed-effect estimates for all four outcomes. Overall, performance improved with practice, particularly for temporal and spatial efficiency measures, while group differences between novices and experts remained substantial across sessions. For , there was a significant reduction at Practice 3 (, p<0.05) and a larger decrease at Practice 4 (, p<0.001), indicating faster task completion with repeated practice. A similar pattern was observed for , with a significant reduction at Practice 4 (, p<0.01), suggesting more efficient trajectories. Despite these improvements, novices exhibited consistently poorer performance than experts. Specifically, the Novice group showed longer execution times (, p<0.001), longer path lengths (, p<0.01), and lower path efficiency (, p<0.05). Additionally, novices displayed reduced angular speed variability (, p<0.05), reflecting less dynamic control. Importantly, interaction terms between practice and group were not statistically significant across outcomes, indicating that learning rates were broadly comparable between novices and experts. Thus, while practice led to measurable improvements, it did not fully close the performance gap between groups within the observed sessions. Variance components revealed moderate to high between-subject heterogeneity (ICC ≈ 0.32–0.55), supporting the inclusion of participant-level random effects. Group differences The Group (Novice) coefficient is statistically significant for all four outcomes. On the log scale, novices showed a mean 0.589 units higher than experts at the reference session (, SE = 0.131, p < 0.001), corresponding to a multiplicative factor of —novices took approximately 80% longer to complete the task. Similarly, was 0.407 units higher in novices (, SE = 0.126, p = 0.001; ), indicating 50% longer trajectories. On the untransformed scale, PathEfficiency was 0.067 units lower in novices (, SE = 0.028, p = 0.018), confirming geometrically less economical motion. For AngularP95, novices exhibited a 0.549 rad s lower peak angular speed (, SE = 0.227, p = 0.016), indicating reduced peak rotational agility rather than excessive angular noise—consistent with the HMM finding that novices in State 5 showed less purposeful rotational exploration (Section "Latent behavioral profiling through hidden Markov models"). Practice effects and learning curves Figure 12 summarises the estimated marginal means (EMMs) for all four outcomes across sessions P2–P4. Overall, both groups exhibited clear improvements in temporal and spatial performance, as reflected by reductions in and . For execution time, experts showed a substantial decrease by session 4 (, p<0.001 relative to session 2), corresponding to an approximate 32% reduction on the original scale. Novices exhibited a comparable reduction (-0.306 log-units), obtained from the sum of the main practice effect and the interaction term. The difference in learning rates between groups was not statistically significant (p=0.573 for the P4 Novice interaction), indicating that both groups improved at similar rates despite starting from different baselines. A similar pattern was observed for path length, where both groups showed reductions across sessions, with experts maintaining consistently shorter trajectories throughout. These results suggest that practice leads to measurable gains in execution efficiency but does not eliminate baseline performance differences between groups. In contrast, no significant practice effects were observed for PathEfficiency or AngularP95 (all p>0.05). This indicates that higher-level aspects of motor control, such as trajectory economy and rotational agility, are less responsive to short-term repetition and may require longer or more targeted training interventions to improve. Taken together, these results show that while repeated practice improves speed and movement efficiency, it does not substantially reduce the expert–novice performance gap within the three sessions studied. Individual learning trajectories and variance decomposition Figure 13 displays individual trajectories for across practice sessions, revealing substantial inter-individual heterogeneity in both groups. While many participants exhibit decreasing trends consistent with overall learning effects, trajectories are not uniformly monotonic, with several individuals showing plateaus or transient increases between sessions. Among experts, most trajectories follow a downward trend, although moderate cross-over between individuals is observed, indicating variability in baseline performance and learning progression. In contrast, novices exhibit both higher initial values and a wider dispersion of trajectories, reflecting greater heterogeneity in performance levels. This variability is consistent with the variance decomposition obtained from the LMM. In particular, the relatively high intraclass correlation coefficient (ICC ≈ 0.55) indicates that a substantial proportion of total variability is attributable to between-subject differences rather than within-subject fluctuations across sessions. The estimated random-intercept variance () further supports the presence of persistent individual-specific effects that are not explained by practice alone. Taken together, these results indicate that learning trajectories are highly individualised, and that repeated practice does not fully homogenise performance across participants within the observed sessions. The intraclass correlation coefficients quantify the proportion of total outcome variance attributable to stable between-participant differences. ICC values ranged from 0.322 (PathEfficiency) to 0.553 (), indicating that 32–55% of total variance reflects systematic individual differences rather than within-participant session-to-session fluctuation. This magnitude of ICC underscores the appropriateness of the random-intercept structure and highlights that baseline performance heterogeneity is a substantial source of variability, which should be accounted for in training-protocol design through individualised progression thresholds. Implications for training design Taken together, the LMM results indicate three distinct patterns. First, a robust and statistically significant group main effect is present across all four kinematic dimensions, confirming that expert and novice performance differs systematically and cannot be attributed to sampling variation. Second, both groups improve on temporal and spatial efficiency outcomes (execution time, path length) with practice, but the improvement rate does not differ significantly, and the expert–novice gap persists at all observed sessions. Third, rotational agility and path economy show no significant practice-related change within three sessions. These findings suggest that while brief repeated exposure to the task is sufficient to elicit temporal learning, deeper motor dimensions—horizon-stabilisation efficiency, rotational agility, and navigation economy—require either more repetitions, targeted adaptive feedback, or structured task progressions to improve. The state-level HMM results (Section "Latent behavioral profiling through hidden Markov models") provide mechanistic detail on which behavioral sub-phases are most amenable to such interventions. Latent behavioral profiling through hidden Markov models Model selection Hidden Markov Models (HMMs) with multivariate Gaussian emissions and full covariance structure were fitted to the concatenated interpolated trajectories from practices 2–4. Each observation comprised the three spatial coordinates and the four unit-quaternion components , standardised across all sequences prior to fitting. Models were trained for 300 Expectation–Maximisation iterations with a fixed random seed, and the number of hidden states was evaluated over . The number of free parameters for a K-state Gaussian HMM with full d-dimensional covariance is yielding for . Table 6 report the log-likelihood, AIC, and BIC for each candidate model. Both criteria decrease monotonically across the evaluated range without exhibiting a clear elbow, with K = 6 achieving the lowest values (; ; ). The absence of an elbow is common in densely sampled multivariate time series where each additional state captures finer-grained behavioral sub-phases. In such cases, the boundary of the search range becomes the data-driven optimum; extending the range further would risk overfitting given the available sample size (, ). We therefore retained K = 6 as the information-theoretically best model within the evaluated range, while acknowledging this limitation explicitly in Section 3.3. State decoding and occupancy Latent state sequences were decoded for every trajectory via the Viterbi algorithm. For each participant, the proportion of frames assigned to each of the six states was computed as a measure of state occupancy. Table 7 reports the group means, standard deviations, and the results of Welch two-sample t-tests on the per-participant proportions, with Benjamini–Hochberg (BH) correction applied across the six simultaneous comparisons. Table 7Group differences in HMM state occupancy (proportion of frames assigned to each state per participant). Values are mean ± SD. Δ = Expert − Novice. Statistical testing: Welch two-sample t-test with Benjamini–Hochberg correction across K = 6 comparisons. Cohen’s d provides the standardised effect size. Despite wide confidence intervals due to the modest sample size, State 5 shows the largest effect (d = 0.707, medium-to-large), with experts occupying this state ≈ 12.9 percentage points more than novices.StateExpert mean ± SDNovice mean ± SDΔWelch tpd00.232 ± 0.0820.240 ± 0.128-0.007-0.1940.8470.883-0.06810.235 ± 0.1030.204 ± 0.101+0.0310.4070.88320.146 ± 0.1170.141 ± 0.085+0.0060.8830.88330.214 ± 0.1050.226 ± 0.099-0.011-0.2930.7720.883-0.11040.181 ± 0.1380.227 ± 0.280-0.047-0.6280.5360.883-0.2115 Group differences in HMM state occupancy (proportion of frames assigned to each state per participant). Values are mean ± SD. Δ = Expert − Novice. Statistical testing: Welch two-sample t-test with Benjamini–Hochberg correction across K = 6 comparisons. Cohen’s d provides the standardised effect size. Despite wide confidence intervals due to the modest sample size, State 5 shows the largest effect (d = 0.707, medium-to-large), with experts occupying this state ≈ 12.9 percentage points more than novices. Figure 14 illustrates the occupancy profiles. After BH correction, none of the six state comparisons reaches the α = 0.05 significance threshold, a result that should be interpreted in the context of the study’s sample size rather than as evidence of no difference. With and , the achieved statistical power to detect a medium effect (d = 0.5) in a two-sample t-test is approximately at α = 0.05, insufficient to reliably reject the null hypothesis when true moderate group differences exist. Mean proportion of frames in each HMM latent state (K = 6), stratified by group (Expert vs. Novice); bars show ± 1 SD. State 5 shows the largest group difference (Δ = +0.129, Cohen’s d = 0.707), with experts occupying this state markedly more than novices. No comparison reaches after BH correction, attributable to limited statistical power given the modest sample size (, ); see main text for power analysis. Notably, State 5 exhibits both the largest absolute difference (Δ = +0.129; experts spend of frames in this state vs. for novices) and the largest standardised effect size (d = 0.707), which by Cohen’s conventions falls in the medium-to-large range. This effect size is meaningful and provides substantive evidence that State 5 encodes a behavioral mode more characteristic of expert performance, even if the small sample prevents formal statistical confirmation at conventional thresholds. State feature profiles To characterise the kinematic content of each latent state, all trajectory segments decoded into that state were collected (retaining only segments with ≥ 5 frames to ensure stable estimates), and the engineered feature vector (fully defined in Table 2) was computed per segment. State-level feature means were then z-scored across states to produce the standardised profile matrix shown in Fig. 15. The heatmap reveals distinct kinematic signatures for each state, summarised as follows. State 0 is characterised by above-average execution time (z = +0.88), elevated StartX (z = +0.88) and StdAngleZ (z = +1.11), suggesting a spatially rightward, angularly variable phase with moderate duration. State 1 shares the above-average execution time (z = +0.96) of State 0 but presents strongly negative MeanAngleX (z = -1.73) and MaxAngleX (z = -1.34), consistent with a camera orientation phase in which the pitch axis is held at a persistently low angle. State 2 is defined by high AngularMean (z = +1.78) and AngularP95 (z = +1.33) combined with low execution time (z = -0.79), indicating a rapid, angularly agitated phase consistent with short corrective reorientation manoeuvres. State 3 is the “slow-approach” state: it presents the most negative execution time (z = -1.62), strongly negative velocities along all three axes, and elevated MeanAngleY (z = +1.50) and MaxAngleY (z = +1.75), suggesting a deliberate, low-speed phase with the camera pitched upward. State 4 constitutes the highest-energy state, with the largest positive z-scores for all velocity components (AvgVelocityX: z = +1.69; AvgVelocityY: z = +1.67; AvgVelocityZ: z = +1.92), all displacement ranges, and AngularStd (z = +1.99), consistent with large, rapid, and spatially unconstrained camera repositioning. State 5 (the expert-enriched state) presents the highest ExcessPath (z = +1.89) and positive PathLength (z = +1.72) alongside the most negative MaxRotationX (z = -2.02) and negative execution time (z = -1.61), suggesting a behaviorally compact yet rotationally distinctive phase characterised by active horizon management with limited axial translation. Within state 5 group comparisons To characterise the expert advantage within State 5, all trajectory segments decoded into that state were isolated and feature distributions were compared between groups using Welch two-sample t-tests with BH correction across 13 simultaneous comparisons (Table 8; Fig. 16). Although no comparison survives BH correction at α = 0.05, three features exhibit standardised effect sizes in the large range (|d| ≥ 1) and therefore warrant substantive interpretation. StdAngleX shows the strongest separation (d = +1.621, p = 0.023 uncorrected): within State 5, experts exhibit substantially greater variability in the pitch-axis Euler angle ( vs. ), suggesting that experts actively explore a wider pitch range during this state, consistent with deliberate horizon-scanning and target-acquisition strategies for a 30-degree laparoscope. In contrast, StdAngleY (d = -1.581) and StdAngleZ (d = -1.121) are higher in novices, indicating greater uncontrolled yaw and roll variability—noise rather than purposeful exploration. PathEfficiency shows a medium-to-large effect (d = +0.983): experts achieve a ratio of direct displacement to total path length nearly twice that of novices (0.072 vs. 0.039), indicating considerably more economical navigation even within the same latent behavioral state. Similarly, MaxDispY (d = +0.859) and AngularP95 (d = +0.648) are consistently higher in experts, suggesting greater vertical workspace utilisation and higher peak angular agility. Taken together, the within-state evidence converges on a coherent interpretation of State 5: it is a behaviorally rich phase during which experts exhibit purposeful rotational exploration (high StdAngleX, high AngularP95), efficient spatial navigation (PathEfficiency novice), and greater vertical reach (MaxDispY), while novices in the same state show greater uncontrolled yaw and roll variation (StdAngleY, StdAngleZ). Transition structure and temporal dynamics The empirical transition matrices computed separately for each group (Fig. 17) reveal that the overall dynamics are highly self-absorbing in both groups: diagonal entries for all six states exceed 0.97 in experts and 0.97 in novices, indicating that the trajectory data consist of sustained behavioral episodes separated by relatively infrequent transitions. This is consistent with the high sampling density of the interpolated trajectories and the block structure of the navigation task. A noteworthy structural feature of both matrices is the near-zero probability of any transition from States 0–4 into State 5, and conversely, the near-zero probability of exiting State 5 to any state other than State 4 (the high-energy repositioning state). This suggests that State 5 is an absorbing or near-absorbing regime that is entered predominantly through the high-energy repositioning phase (State 4) and exited only rarely. Experts enter State 5 more frequently—consistent with their higher occupancy proportion (48.5% vs. 35.6%)—while once in State 5 both groups sustain it with nearly identical persistence (, ). This dissociation between entry frequency (group-dependent) and dwell time (group-independent) provides mechanistic insight: the expert advantage in State 5 resides in the capacity to enter and re-enter this behavioral regime more consistently across trials, rather than in prolonged single episodes once it is reached. Relationship to the supervised classification analysis The HMM analysis is methodologically independent of the supervised classifiers reported in Section "Classification performance and feature importance", yet the two analyses converge on consistent behavioural dimensions. The features that distinguish experts from novices within State 5—particularly PathEfficiency, StdAngleX, and the angular-speed metrics—are the same dimensions identified as most discriminating by the Logistic Regression coefficients (Figure 11). This convergence between a generative unsupervised model (HMM) and a discriminative supervised model (LR) on the same feature dimensions strengthens the interpretation that these kinematic signatures reflect genuine and stable behavioral differences between the two populations, rather than statistical artefacts of any single modelling approach. Limitations of the HMM analysis Several limitations must be acknowledged. (i) Monotonically decreasing selection criteria. BIC and AIC decreased without a clear elbow across , implying that the optimal K within the evaluated range coincides with the search boundary. The six-state solution should therefore be regarded as the best available rather than a global optimum; higher values of K were excluded to limit overfitting given the sample size but may capture additional behavioral granularity. (ii) Sample-size-limited statistical power. With and , the study is substantially underpowered for the six simultaneous Welch tests on occupancy proportions; post-hoc power for a medium effect at α = 0.05 is estimated at . The large Cohen’s d values observed for State 5 (d = 0.707 for occupancy; |d| ≥ 1 for three within-state features) are therefore practically meaningful despite non-significant adjusted p-values. (iii) Hard state assignment. Viterbi decoding assigns each frame to a single state, suppressing the posterior probability uncertainty; soft decoding would provide more nuanced occupancy estimates but complicates group comparisons. (iv) Post-hoc label semantics. The behavioral labels assigned to each state (e.g., “high-energy repositioning” for State 4, “active horizon management” for State 5) are interpretive and descriptive; they cannot be causally validated without kinematic ground-truth measurements from real surgical procedures. (v) Pool-level model. The HMM was fitted on the pooled trajectory set across participants; individual-level heterogeneity in emission parameters is therefore not modelled. Future work should consider mixed-effects HMMs or hierarchical Bayesian formulations to account for participant-specific behavioral styles. Replication with a larger, independent cohort is needed before these latent state findings can be generalised to other simulators or clinical populations. Discussion and conclusion The construct-validity question The central question addressed here is whether a portable, mechanically constrained VR platform can generate behaviorally organised, expertise-sensitive kinematic structure in a laparoscopic camera navigation task. Following Carter et al.15, construct validity in surgical simulation requires that performance metrics index the specific motor demands of the target skill in a manner that covaries with documented experience differences—not merely that groups can be statistically distinguished. The relevant demands for camera navigation include coordinated management of the fulcrum pivot, horizon stabilisation under roll, optical redirection through shaft rotation, and depth regulation through axial translation4,5,20. The convergent evidence from three independent analytical layers supports this more restrictive interpretation: SECMA reproduces sufficient task-defining control structure to generate measurable expertise-sensitive kinematic organisation and short-horizon practice effects. Functional fidelity over display fidelity High-fidelity commercial platforms reproduce anatomical environments with considerable realism at costs exceeding USD 100,000, limiting accessibility in resource-constrained settings13,23,25. The literature on VR effectiveness has consistently shown that training benefit depends less on visual or geometric duplication than on whether the system preserves the action–effect contingencies that organise the target skill16,36–38. For camera handling, these contingencies are the non-linear mapping between shaft rotation and visual-field direction, the horizon disturbance induced by roll, and the depth-dependent change in apparent target position. SECMA was designed around this principle: the telescopic shaft preserves axial translation and the fulcrum pivot constraint; the decoupled control logic separates horizon stabilisation from optical redirection; and the kinematic model implements oblique optical geometry in the virtual camera. This design philosophy distinguishes SECMA from VBLaST10 and SinaSim27, which emphasise generic instrument handling without task-specific optical constraints, and from SIMISGEST-VR24, where camera navigation is embedded in a broader curriculum. The most closely related specific-purpose systems—the CameraTrainer and the LapSim curriculum validated by Shetty et al.21—have established that dedicated camera-navigation simulation improves technical performance; Kengen et al.5 confirmed that performance and learning curves differ across optical angles, directly supporting angle-specific training; and Jiongbiao et al.22 confirmed construct sensitivity in structured VR camera-navigation tasks. SECMA extends this line of work in three specific ways not present in any single existing platform: a portable, fully self-contained device enforcing the mechanical constraints of handling; continuous 6-DoF telemetry at 80–100 Hz; and a multi-layered analytic framework combining feature-based classification, latent-state modeling, and longitudinal mixed-effects analysis. Classification results and linear separability Under RNCV (5 × 50 = 250 outer folds, inner 4-fold hyperparameter search), Logistic Regression achieved the highest mean (0.818 ± 0.093) and AUC (0.904 ± 0.061), with SVM and Random Forest forming a statistically equivalent top tier (Wilcoxon signed-rank, and 0.107 respectively). The sufficiency of a linearclassifier implies linear separability of the expert and novice feature distributions, consistent with findings from Winkler-Schwartz et al.34 and Bogar et al.35. The standardised LR coefficients identify temporal efficiency (ExecutionTime, PathLength) and depth-axis control (AvgVelocityZ) as the most discriminating dimensions. The prominence of depth-axis velocity is mechanically meaningful: in laparoscopy, uncontrolled axial motion displaces the focal plane and introduces horizon shifts; experts who manage axial insertion more deliberately exhibit lower and more stable Z-axis velocity. Angular descriptors (StdAngleX, MeanAngleZ) contribute additionally, indicating that rotational regulation is a secondary but significant discriminating dimension. These features are directly interpretable in terms of the task demands, supporting the claim that the telemetry samples clinically relevant aspects of camera handling30,31. Latent behavioral states and motor variability State 5 is the most theoretically interesting latent state: experts spent approximately 48.5% of frames in this state versus 35.6% for novices (Δ = +0.129; Cohen’s d = 0.707, medium-to-large). Although the difference does not reach after BH correction—attributable to limited statistical power (estimated for d = 0.5 at the available sample size)— the effect size is practically meaningful. Within State 5, the most striking group difference is in StdAngleX (d = +1.621): experts show greater pitch-axis variability ( vs. ), while novices show greater yaw and roll variability (StdAngleY: d = -1.581; StdAngleZ: d = -1.121). This asymmetry—expert variability structured and directional, novice variability diffuse—is precisely the distinction between skilled and unskilled motor variability described by Bernstein45 and formalised as the difference between “good” (exploratory) and “bad” (error-generating) variability46,47. To our knowledge, this distinction has not been previously reported in the context of laparoscopic camera navigation simulators. The transition analysis reveals that both groups sustain State 5 with nearly identical persistence once entered (, ). The expert advantage therefore lies entirely in the frequency of entry into State 5. Since State 4 (high-energy repositioning) is the primary transition source into State 5, training interventions aimed at reducing the amplitude and frequency of corrective repositioning manoeuvres represent the most direct avenue toward closing this gap. Longitudinal performance and practice effects Experts completed the task approximately 80% faster than novices (; p < 0.001), traversed 50% shorter trajectories (; p = 0.001), showed higher path efficiency (p = 0.018), and maintained higher peak angular speed (p = 0.016), consistent with purposeful rotational exploration. The critical finding is that neither PracticeGroup interaction was statistically significant (all p> 0.41): both groups improved at statistically similar rates across sessions. The expert–novice gap persists not because novices fail to learn, but because they start from a higher baseline and converge toward it proportionally. Path efficiency and AngularP95 showed no significant practice effects in either group, suggesting that trajectory economy and rotational agility require targeted feedback, longer training horizons, or explicit task progressions to develop. ICC values of 0.32–0.55 confirm that 32–55% of total variance reflects stable between-participant differences, reinforcing the need for individualised progression thresholds6,7. Both the LMM and HMM frameworks independently identify path economy, angular variability, and depth-axis control as the primary axes of expert–novice differentiation. This cross-framework convergence strengthens the interpretation that the expertise signal in SECMA’s telemetry is genuine and stable. Positioning SECMA within the surgical training literature SECMA is positioned as a preparatory training tool for laparoscopic camera navigation, a skill typically acquired through intraoperative exposure or physical box trainers where assessment is largely subjective and limited to task completion time. SECMA enables structured, repeatable practice with continuous objective metrics, potentially supporting earlier acquisition of fundamental visuomotor skills. The system is currently in active use in training contexts at the Universidad de Chile and Clínica Alemana de Santiago, providing initial evidence of practical integration into surgical education. The present work adds to the contributions of Nilsson et al.20, Shetty et al.21, Arikatla et al.9, and Jiongbiao et al.22 three elements not present in combination in existing platforms: (i) a fully portable, self-contained device; (ii) continuous high-frequency 6-DoF telemetry; and (iii) a multi-layered analytic framework simultaneously characterising aggregate efficiency, within-trial behavioral state organisation, and longitudinal practice trajectories. An important distinction with prior AI-enhanced simulation work34,35 is that the present contribution combines classifier analysis with unsupervised latent-state modeling, characterising not only whether groups differ but how they differ at the sub-trial level. The HMM framework opens a path toward state-aware adaptive feedback: by identifying in real time which latent behavioral state a trainee currently occupies, the system could provide targeted guidance specific to the behavioral mode rather than relying on global performance summaries. Limitations The sample is modest (, ), limiting precision in interaction estimates and HMM occupancy contrasts; post-hoc power for the state-occupancy comparison is , so non-significant adjusted p-values should be interpreted as a power issue rather than evidence of no effect. The three analysed sessions characterise early adaptation but cannot resolve asymptotic performance, retention, or transfer to the operating room; concurrent, predictive, and transfer validity remain to be established. The HMM was fitted on a pooled trajectory set without individual-level emission parameters; Viterbi hard decoding suppresses posterior probability uncertainty; and the BIC/AIC curves decreased monotonically without a clear elbow, so K = 6 is the best available rather than a globally optimal solution—hierarchical HMM formulations are recommended for future work. Finally, SECMA does not replicate the exact geometry, mass distribution, or haptic profile of a clinical laparoscope; the appropriate claim is preservation of relevant control constraints, not full physical equivalence. Implications for training design and future directions The converging evidence suggests three actionable training targets: (i) reducing redundant translational travel (ExcessPath, practice-sensitive); (ii) developing purposeful pitch exploration rather than diffuse yaw/roll variability (StdAngleX vs. StdAngleY/Z, not practice-sensitive under brief repetition and requiring targeted feedback); and (iii) increasing entry frequency into State 5 by reducing the amplitude of corrective repositioning manoeuvres. Future work should pursue extended training protocols (≥ 10 sessions), experimental manipulation of feedback conditions including state-aware real-time HMM-based guidance, transfer-validity studies, and hierarchical behavioral modeling accounting for participant-level random effects. Integrated analytical perspective Figure 18 summarises the analytical structure of the study and clarifies how the different modelling layers contribute complementary information. The pipeline begins with the VR-based execution of the camera-navigation task under mechanically constrained interaction, followed by continuous acquisition of positional and rotational trajectories. These trajectory data are then analysed along two coordinated branches. In the first, raw sequential data are modelled directly through Hidden Markov Models (HMMs) to characterise recurrent latent coordination regimes and their group-specific occupancy patterns. In the second, interpolated trajectories are transformed into engineered kinematic descriptors that support supervised expert–novice classification and longitudinal mixed-effects modelling across practice sessions. This dual analytical structure is important because it allows the simulator to be evaluated not only in terms of global performance differences, but also in terms of the temporal organisation of behaviour. Accordingly, the construct-related evidence provided by SECMA does not rely on a single endpoint; it emerges from the convergence of discriminative, longitudinal, and latent-state analyses applied to the same task environment. Integrated analytical pipeline of the SECMA camera-navigation study. Trajectory data acquired during the virtual laparoscopic task were processed through two complementary branches: direct sequence modelling via Hidden Markov Models (HMMs) for latent-state inference, and feature-based modelling for supervised classification and longitudinal mixed-effects analysis. The combination of both branches provides convergent construct-related evidence on expertise sensitivity and performance organisation. Conclusion SECMA is a portable, mechanically constrained, and telemetry-rich VR platform for laparoscopic camera navigation training. Its scientific contribution lies in three integrated elements not combined in any existing low-cost portable system: a mechanical design that enforces the key motor constraints of camera handling; continuous 6-DoF telemetry supporting quantitative analysis at the trial and sub-trial level; and a multi-layered analytic framework characterising performance through convergent evidence from supervised classification, unsupervised latent-state inference, and longitudinal mixed modeling. The classification analysis establishes recoverable expertise-related structure (, AUC = 0.904) under rigorous RNCV. The HMM analysis reveals that this structure extends to the temporal organisation of movement, with experts showing d = 0.707 higher occupancy of an efficient behavioral state and purposeful rather than diffuse rotational variability. The LMM analysis confirms practice-related improvement in temporal and spatial efficiency, a persistent expert–novice gap across all observed sessions, and that rotational agility and path economy require more than brief repetition to develop. Together, these characteristics make SECMA a promising and analytically rigorous platform for scalable camera-navigation training, and establish the methodological foundation for future work on adaptive feedback, proficiency-based progression modeling, and transfer validation in minimally invasive surgery education. Supplementary Information Supplementary Information. Author contributions J.I.G.M. conceptualized and implemented the virtual reality simulator for laparoscopic camera control, designed the validation task, and contributed substantially to manuscript writing. M.H.M. conceived and designed the experiments, collected and analyzed data, performed machine learning and hidden Markov model analyzes, prepared visualizations, developed the methodological framework, and contributed substantially to manuscript writing. C.A.D.C. provided surgical expertise for the design and implementation of the 30-degree laparoscopic camera navigation exercise, participated in experimental execution, and reviewed the manuscript. B.M.G. advised on the conceptualization of the laparoscopic camera navigation exercise, and contributed to the organization and execution of experiments. S.E.P. designed the mechanical interface for the laparoscopic simulation system, and constructed the mechanical prototypes for the SECMA simulator. C.I.R.B. contributed to the conception, design, and evaluation of the mechanical interface for the laparoscopic camera simulator, and participated in organizing the experimental session. R.T.V. contributed to the design and implementation of the simulator’s management software and developed the services required for integration with the selected and trained machine learning models. All authors reviewed and approved the final manuscript. Funding This research was supported by the Chilean National Agency for Research and Development (Agencia Nacional de Investigación y Desarrollo, ANID) through the Scientific and Technological Research Promotion Fund (ANID IT 2023), under Grant No. IT23I0018: “Portable training system for minimal access surgeries (SECMA) based on a virtual reality simulator, which incorporates surgical suturing exercises and AI applications in the analysis of data from practical activities.” Data availability The data sets and code generated and analyzed during this study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Ethical approval All procedures involving human participants in the construct validation study were reviewed and approved by the Ethics Committee of the Universidad del Desarrollo (UDD), Chile, and carried out in full accordance with the Declaration of Helsinki, as well as institutional and national ethical guidelines. Written informed consent was obtained from all participants prior to enrollment, including consent for voluntary participation and anonymous processing of their data. In addition, specific written informed consent was obtained from the individual depicted in Figure 5(B) for the publication of the identifiable image. All experimental methods were carried out in strict compliance with applicable ethical regulations and supervision requirements. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary Information The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-55403-1. References Associated Data Data Availability Statement The data sets and code generated and analyzed during this study are available from the corresponding author on reasonable request.