IMU-based identification of rowing conditions through supervised machine learning Authors: Anna Leber, Tobias Siebert, Walter Rapp, Steffen Held Published: 2026 ABSTRACT AbstractBackgroundRowing combines on-water and ergometer-based training, while inertial measurement units (IMUs) offer a practical approach for field-based biomechanical monitoring. This study examined whether trunk-mounted IMU data can distinguish static Concept2 (C2), dynamic RP3, and on-water rowing conditions.MethodsTen youth rowers (16.3 ± 0.8 years) performed high-intensity rowing in all three conditions. Triaxial acceleration and gyroscope data were recorded at 208 Hz. Six classical machine-learning and four deep-learning models classified overlapping three-stroke windows using leave-one-subject-out cross-validation. Repeated-measures analyses assessed condition-specific IMU magnitude features.ResultsThe ensemble achieved the highest mean accuracy (85.5 ± 10.3%). Individual accuracy varied from 24.0% to 97.1% across models and athletes. The 7,545 windows were imbalanced toward Boat, with a no-information rate of 71.8%. Boat was identified most reliably, whereas RP3 was frequently misclassified as Boat. Nine of ten magnitude features remained significant after false-discovery-rate correction.ConclusionsTrunk-mounted IMU signals distinguished the three rowing conditions at group level. However, class imbalance, RP3 familiarity, measurement order, boat-standardisation asymmetry, and athlete-level variability require cautious interpretation. Applied use requires athlete-specific calibration and prospective validation.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13102-026-02082-0.Keywords: Rowing, Wearable sensors, Machine learning, Leave-one-subject-out, Trunk kinematics, Static ergometer, Dynamic ergometer, RP3, Concept2, Biomechanics Introduction Rowing has been practised since antiquity and was scheduled for the first modern Olympic Games in 1896, although the events were cancelled due to weather; it has been contested since 1900 [1]. In recent decades, rowing has evolved, with ergometers complementing traditional training on the water [2]. While rowing in a boat is not always feasible, ergometers provide an alternative training option regardless of weather conditions, offer controlled training environments, and allow standardised performance diagnostics and monitoring [3]. Rowing technique is characterised by a cyclical movement with a complex and structured pattern of repeated strokes [4–6]. Each requires a coordinated interaction between the lower limbs, trunk, and upper body [7]. Biomechanical analyses in rowing are typically used to better understand these movement patterns, focusing on kinematics and kinetics using measurement tools such as force sensors, video analysis, and inertial measurement units (IMUs). The accelerometer and gyroscope combination used in the present sensor system has previously been validated for dynamic movement capture in skiing and daily-life walking contexts, although not under rowing-specific conditions [8]. In rowing, inertial sensors have been applied to monitor stroke-related movement characteristics and performance metrics in on-water and ergometer settings, supporting their use as practical tools for field-based movement analysis in this sport [9–11]. Training on rowing ergometers should approximate the kinematic and kinetic demands of rowing in a boat. Two types of ergometers are commonly used in training and research: the static and the dynamic ergometer. The established static Concept2 (C2) ergometer (Concept2 Inc., Morrisville, VT, USA) is frequently used for training and performance testing in rowing [3, 12]. It features a movable sliding seat while the flywheel-stretcher unit remains fixed, causing the rower’s centre of gravity and a large part of the body mass to shift from anterior to posterior along the sliding rail during each stroke [13]. In contrast, the RP3 Model T dynamic rowing ergometer (RP3 Rowing B.V., The Netherlands) has a movable sliding seat and a freely moving flywheel-foot-stretcher unit. This design produces only minor shifts in the rower’s centre of gravity compared to the static ergometer, more closely resembling on-water dynamics [13, 14]. The flywheel-stretcher unit weighs approximately 17 kg and has been described as approximating the mass of a single scull [15]. For context, World Rowing specifies a minimum boat mass of 14 kg for single sculls [16]. Previous comparisons indicate both similarities and differences across rowing conditions. Trunk, thigh, and lower-leg angles at the catch and finish, as well as stroke length, can be comparable between dynamic ergometers and on-water rowing [17]. Handle-force curves from dynamic ergometers also correlate strongly with on-water curves (r > 0.90), suggesting similar force-application timing despite differences in magnitude [12, 17]. In contrast, static ergometry has been associated with greater lumbar flexion [18], whereas dynamic ergometry can produce higher knee and hip angular accelerations and higher stroke rates [19–22]. Static systems often produce higher peak handle forces, while dynamic systems can reduce handle force and increase foot-stretcher force or alter joint power [19, 20, 23, 24]. However, other studies reported no differences in selected power or force variables [25, 26]. Dynamic ergometry may therefore approximate important aspects of on-water rowing without reproducing all kinematic and kinetic demands. The literature is therefore informative but partly inconsistent and is based on heterogeneous samples, ergometers, and measurement systems. Direct three-condition comparisons remain scarce. It is unclear whether trunk IMU patterns can reliably distinguish C2, RP3, and Boat, whether classification performance is consistent across athletes, and which acceleration or gyroscope features contribute to separation. Inter-individual differences in stroke style create an additional challenge [20, 27]. This uncertainty limits evidence-based interpretation of condition-specific movement signatures. Machine-learning methods can identify nonlinear patterns in high-dimensional biomedical and wearable-sensor data [28, 29]. They may therefore distinguish closely related rowing conditions while feature-based analyses can indicate which acceleration and gyroscope descriptors differ between equipment settings. Combining both approaches provides classification performance and an interpretable biomechanical context while permitting direct comparison of feature-based and sequence-based representations. Against this background, we examine whether classical machine-learning and deep-learning models can distinguish rowing movements using IMU-derived data. We address four research questions: (i)Can classical machine-learning and deep-learning models distinguish between the three rowing conditions (C2, RP3, Boat)? Can classical machine-learning and deep-learning models distinguish between the three rowing conditions (C2, RP3, Boat)? (ii)What is the accuracy of the distinction at the individual level? What is the accuracy of the distinction at the individual level? (iii)Does classification accuracy differ across the three conditions? Does classification accuracy differ across the three conditions? (iv)Which acceleration and gyroscope parameters differ between conditions? Which acceleration and gyroscope parameters differ between conditions? For this study, adequate classification performance was defined as a mean accuracy of at least 70% across athletes. We hypothesised that athlete-level accuracy would vary because of inter-individual stroke patterns [27]. Based on reported mechanical similarities between dynamic ergometry and on-water rowing [14, 17], we expected the strongest overlap between RP3 and Boat and clearer separation of static C2. We also expected higher acceleration magnitudes on C2 because more of the rower’s body mass moves relative to the fixed flywheel–foot-stretcher unit [13]. These hypotheses concern trunk movement signatures and do not imply full biomechanical equivalence between training modes. Methods This cross-sectional study examined condition-specific rowing movement patterns without a training intervention. We conducted data collection in cooperation with the Olympic Training Center Freiburg-Schwarzwald (Germany). Participants We recruited ten youth rowers (four female, six male; 16.3 ± 0.8 years; 1.77 ± 0.13 m; 70 ± 11.8 kg) from the regional training centres Breisgau-Hochrhein and Karlsruhe. Participants belonged to the U17 and U19 training squads and reported a mean rowing experience of 3.3 ± 1.6 years. Inclusion criteria required at least one year of competitive rowing experience and an age of 18 years or younger. At the time of data collection, all athletes were in good physical condition and free of acute injuries; therefore, no exclusions were necessary. All athletes were familiar with the static C2 ergometer, whereas only 40% reported prior experience with the dynamic RP3 ergometer. Participation was voluntary. As all athletes were minors, we obtained written informed consent from their parents or legal guardians prior to participation. Athletes and guardians received written and verbal information about the study aims, experimental design, and the potential benefits and risks of the testing procedures. Following data collection, the Internal Ethics Committee of IST University of Applied Sciences granted positive ethical approval for the exact study procedures, measurements, Archinisis sensor system, instrumentation, data analyses, and consent arrangements reported in this manuscript (approval no. IR100726IST233, issued on 10 July 2026). Measuring instruments Movement data were recorded using wearable inertial sensors (Archinisis GmbH, Düdingen, Switzerland). Each sensor unit (78 g) included a global navigation satellite system (GNSS) receiver (u-blox M10 chip, 10 Hz) and an inertial measurement unit (IMU) (ST ISM330DLC, 208 Hz). The IMU combined an accelerometer, which measured linear accelerations, and a gyroscope, which recorded angular velocities, each along three spatial axes. The Archinisis sensor system has previously been validated for dynamic movement capture in skiing and daily-life walking contexts rather than under rowing-specific conditions [8]. Prior to each measurement, sensors were recalibrated for fifteen minutes to ensure a sufficient GPS signal. The sensors were placed in a pocket on the athlete’s back in a tight-fitting garment to reduce interference with the rowing motion. Global Navigation Satellite System (GNSS) data were not used for classification because translational displacement would trivially separate Boat from the two ergometer conditions. The analysis was restricted to IMU signals to test the more demanding trunk-movement classification problem, particularly the distinction between C2 and RP3. Rowing conditions We collected data under three conditions: static ergometer (C2), dynamic ergometer (RP3), and on-water rowing (Boat) in a single scull (Fig. 1). Rowing conditions: (A) Dynamic rowing ergometer with a movable foot-stretcher-flywheel unit (RP3 Model T [RP3]; RP3 Rowing B.V., The Netherlands). B Static rowing ergometer with a fixed foot-stretcher and flywheel (Concept2 [C2], Model D; Concept2 Inc., Morrisville, VT, USA) provided by the Breisacher Rowing Club. C On-water rowing in individually adjusted boats We checked all ergometers for proper functioning prior to testing, including smooth chain movement and unobstructed seat rail movement. Boat dimensions averaged 7.95 ± 0.13 m in length and 15.79 ± 0.98 kg in mass, distributed across six different boat types (Supplementary Table S1A). We did not standardise the boat across athletes, as individual rigging parameters such as oarlock height and span, oarlock pitch and angle, and foot-stretcher position, inclination, and height are essential for technically sound rowing [30, 31]. Air-damper settings were not standardised across ergometers. Athletes selected these settings themselves according to their usual training practice. The recorded values were air-damper settings, not machine-calculated drag-factor values. Because the settings were self-selected rather than experimentally controlled, they may have contributed to C2 and RP3 inertial signatures and should be considered when interpreting ergometer-specific differences. The corresponding boat, ergometer, sensor, and damper settings are reported in Additional file 1, Supplementary Table S1A. Stroke rate (strokes/min [spm]) was monitored in all three conditions. Both ergometers featured an integrated display, while on-water stroke rate was continuously monitored using a Speed Coach GPS monitor (Nielsen-Kellerman, 92 × 67 × 31 mm, 150 g). Testing procedure Data collection took place on 10, 12, 13, and 28 October 2024 at the Landesstützpunkt Breisach rowing facility on the Rhine River. Athletes 1–8 were assessed during the first data-collection period, whereas athletes 9 and 10 were assessed on 28 October 2024. Environmental conditions were comparable across sessions, with calm water and no wind influence. The mean Rhine discharge was 77.33 ± 19.74 m³/s. Each session began with a preparation phase, during which we equipped athletes with the IMU sensors. Due to additional boat handling procedures (e.g., transporting boat and sculls from the boathouse to the dock, launching and docking the boat, cleaning the equipment), on-water measurements required approximately 60 min per athlete, whereas ergometer trials lasted approximately 35–40 min. To accommodate the logistical constraints of boat handling and to ensure that all athletes completed the on-water condition under comparable environmental conditions (water level, time of day, daylight), we conducted on-water measurements first in all athletes. The order of the two ergometer conditions was subsequently randomised between athletes by coin tossing to minimise order effects between C2 and RP3. Since Boat was always measured first, order effects cannot be fully separated from condition effects. This design choice reduced environmental variation during on-water testing but may have introduced freshness, learning, or time-of-day confounding. Prior to each measurement, athletes completed a standardised 15-minute warm-up in the corresponding condition to familiarise them with the equipment, the environmental conditions, and the target stroke rate of 32 to 36 spm, and to minimise learning effects. During the measurement phase, we instructed athletes to row at a high dynamic effort while maintaining the target stroke rate. The prescribed range was achieved in all cases except for athlete 9 during the on-water condition. Across all three conditions, the mean stroke rate was 33.9 ± 1.8 strokes per minute. On-water measurements were conducted on the same section of the Rhine (river kilometres 223–224) in both directions. Before switching between ergometer conditions, a rest period of at least 15 min was provided, during which the inertial sensors were recalibrated. Data processing We transferred IMU data from the Archinisis sensors to a computer and organised them into athlete-specific folders in the proprietary .nao format. We imported the data into the Archinisis software [32] for initial processing. As no built-in algorithms were available for analysing trunk movements in rowing, we exported all recordings as raw data in HDF5 format for subsequent custom processing. All subsequent processing was implemented in Python 3.12 using NumPy, SciPy, and pandas [33–35]. For each HDF5 recording, we loaded timestamps and triaxial accelerometer and gyroscope signals and ensured a consistent channel-by-time layout. We used the gyroscope y-axis, which represented rotation about the sensor’s medio-lateral axis, to identify strokes. Angular velocity was converted from radians per second to degrees per second. Rowing IMU studies commonly use low-pass Butterworth filtering, although reported specifications vary [9]. We applied a zero-phase, second-order Butterworth low-pass filter with a study-specific 3 Hz cutoff using SciPy’s filtfilt. This cutoff retained the prescribed stroke-rate fundamental of approximately 0.53–0.60 Hz and its first harmonics while suppressing higher-frequency noise. Candidate boundaries were local minima of the filtered signal, detected as peaks in the inverted signal with a minimum separation of 0.5 s. We retained minima below the 0.30 quantile of all candidate-minimum values. Consecutive retained minima delimited candidate segments. A segment was accepted when its duration was 1.3–3.75 s and the absolute filtered gyroscope y-signal exceeded 40 deg/s at least once. The duration bounds correspond to approximately 16–46 spm. The 40 deg/s threshold was a heuristic exclusion criterion rather than a literature-derived cutoff and was applied uniformly to remove low-amplitude segments without a clear drive phase. For each valid stroke, we stored the raw accelerometer and gyroscope segment with its stroke index, sampling frequency, athlete identifier, and condition. We then formed overlapping windows of three chronologically consecutive strokes with a step size of one, only within each recording. The analysed dataset contained 7,605 valid strokes and yielded 7,545 three-stroke windows. Detected stroke boundaries and retained stroke segments were plausibility-checked against synchronised video and outputs from the RowerUp video-analysis platform [36] before the final analysis. For classical machine learning models, we computed a fixed-length feature vector for each detected stroke segment based on triaxial accelerometer and gyroscope signals. In addition to the three axes, we computed the Euclidean magnitude for accelerometry and gyroscopy, resulting in eight signals per stroke [37]: accx, accy, accz, accmag, gyrx, gyry, gyrz, gyrmag, where the index (mag) stands for magnitude. For each of these eight signals, we extracted time-domain descriptors commonly used in wearable sensor feature-based classification, including skewness as a standardised third central moment and kurtosis as excess kurtosis, obtained by subtracting 3 from the standardised fourth central moment so that the normal distribution has kurtosis 0 [38–41]. We further computed within-modality axis correlations using Pearson correlation coefficients for accelerometry and gyroscopy, specifically x with y, x with z, and y with z [42]. Frequency-domain features were computed for the magnitude signals only. We removed the mean from the magnitude signal, computed the power spectrum via the fast Fourier transform, and retained the non-negative frequencies up to the Nyquist frequency [43]. From this spectrum we derived dominant frequency as the frequency with maximum power, spectral centroid as the power-weighted mean frequency, spectral entropy as minus the sum of normalised power times its logarithm, and relative band power from 0 to 5 Hz as the proportion of total spectral power within this band [37, 42, 44]. For each stroke, we retained segment duration, number of samples, and estimated sampling frequency. Estimated sampling frequency was used as a diagnostic variable and excluded from the classical classifier inputs, leaving 96 classification features per stroke. For window-level feature classification, the ordered feature vectors of three consecutive strokes were concatenated into one fixed-length vector. We evaluated six classical models: (1) multinomial logistic regression [45], (2) a support vector machine with a radial basis kernel [46], (3) k-nearest neighbours with k = 5 [47], (4) a decision tree [48], (5) a random forest with 100 decision trees [49], and (6) extreme gradient boosting [50] implemented with XGBoost [51]. For time-series modelling, each stroke was linearly resampled to 200 time points for the three accelerometer and three gyroscope channels. Three consecutive strokes were concatenated into a 600-time-step, six-channel sequence; no sequence was truncated after this transformation. We evaluated (7) a one-dimensional convolutional neural network (CNN) [52], (8) a long short-term memory (LSTM) network [53], (9) a CNN–LSTM hybrid [54, 55], and (10) an ensemble that averaged the class probabilities of the CNN, LSTM, and hybrid models [56]. Data analysis We evaluated all models using leave-one-subject-out validation implemented with a Leave-One-Group-Out scheme [57]. In each fold, one athlete formed the test set and the remaining nine athletes formed the training set. Feature inputs were z-standardised within each fold using parameters fitted only on the training set [58]. For each model, we calculated accuracy, balanced accuracy, macro-F1, and Cohen’s kappa for every held-out athlete. We summarised athlete-level performance as the mean, sample standard deviation, and 95% confidence interval using the t distribution. Confusion matrices pooled out-of-sample predictions across folds so that each window contributed once and only as a held-out observation [59, 60]. The window-level distribution was Boat: 5,421, RP3: 1,160, and C2: 964, yielding a 71.8% no-information rate for the majority Boat class. Counts by athlete and condition and all imbalance-robust metrics are reported in Supplementary Tables S1B, S1C, and S3. To quantify condition-specific differences in acceleration and gyroscope magnitude, we analysed the minimum, maximum, mean, median, and range. Residual normality was assessed with Shapiro–Wilk tests and Q–Q plots, and residual plots were visually inspected [61]. Sphericity was assessed with Mauchly’s test, with Greenhouse–Geisser correction when required. Ten one-way repeated-measures analyses of variance (rmANOVAs) tested the main effect of condition. Partial eta-squared (ηp²) was interpreted as small (≥ 0.01), moderate (≥ 0.06), or large (≥ 0.14) [62]. Benjamini–Hochberg false-discovery-rate (FDR) correction was applied across the ten omnibus tests. We calculated three paired t-tests per feature and applied Benjamini–Hochberg FDR correction across all 30 pairwise comparisons. Pairwise results include 95% confidence intervals for the mean differences and Cohen’s dz [62]. Analyses were conducted using Python 3.12 with NumPy, SciPy, pandas, scikit-learn, TensorFlow/Keras 2.19.0, statsmodels, and XGBoost [33–64]. Deep-learning models used an NVIDIA T4 GPU. Two-sided p < 0.05 was considered statistically significant. Given the athlete-level sample of n = 10, non-significant Shapiro–Wilk and Mauchly tests provide limited evidence for the assumptions. Effect sizes and their rankings are also potentially unstable at this sample size. We therefore interpreted standardised effects alongside raw differences, confidence intervals, FDR-adjusted p values, and the non-parametric sensitivity analysis. The corresponding Friedman and Wilcoxon sensitivity results are reported in Supplementary Tables S2A and S2B. AI-assisted tools were used only for grammar, language checking, and formulation support; all scientific content, analyses, interpretations, and revisions were manually verified and approved by the authors. Results Model performance and inter-individual variability Mean accuracy ranged from 63.3% to 85.5% across models, and the ensemble showed the highest mean accuracy (85.5 ± 10.3%). The 71.8% no-information rate demonstrates that overall accuracy alone can overstate performance in this imbalanced dataset. Across models, mean balanced accuracy ranged from 54.2% to 76.7%, macro-F1 from 50.6% to 77.8%, and Cohen’s kappa from 0.33 to 0.67 (Supplementary Table S3). Performance and uncertainty varied substantially between athletes. The source dataset contained 7,605 valid strokes and 7,545 overlapping three-stroke windows. Window counts by athlete ranged from 432 to 1,212 and are reported by condition in Supplementary Tables S1B and S1C. Athlete-level results showed substantial inter-individual variability (Fig. 2). Athlete 7 had lower performance across several models despite no obvious irregularity during visual reinspection of the raw IMU signals. This athlete contributed 692 test windows (Boat: 519, RP3: 83, C2: 90), which was within the observed athlete-level range of 432–1,212 windows. The lower performance therefore cannot be attributed simply to a very small test set, although segmentation uncertainty and lack of prior RP3 experience remain plausible contributors. Athlete-level accuracy by model. The heatmap shows accuracy for each held-out athlete and model using overlapping three-stroke windows under leave-one-subject-out validation. Colour intensity indicates accuracy from low to high. SVM, support vector machine; CNN, convolutional neural network; LSTM, long short-term memory; KNN, k-nearest neighbours; XGBoost, extreme gradient boosting Differences in classification accuracy across conditions The confusion matrices showed that classification accuracy differed systematically across the three conditions (Fig. 3). Figure 3 displays three representative models: the Ensemble, Logistic Regression, and the CNN. These were selected to provide an interpretable cross section of the modelling approaches, covering an overall best-performing approach (Ensemble), a strong and transparent linear baseline (Logistic Regression), and a raw-sequence-based deep learning model (CNN). The complete set of confusion matrices for all remaining models is provided in Supplementary Figure S1. Across all models, Boat was the most consistently identified condition, with high correct classification rates (up to 95.6%, Fig. 3A) and very low confusion with C2 (0.4%, Fig. 3A). In contrast, RP3 showed the lowest correct classification rates (about 50%, Fig. 3A-C) and was most frequently (about 40%, Fig. 3A-C) misclassified as Boat, indicating reduced separability between these two conditions. C2 showed intermediate performance (about 75%, Fig. 3A-C) and was primarily confused with RP3 (about 12%, Fig. 3A-C) and, depending on the model, to a lesser extent with Boat. In the ensemble (Fig. 3A) and logistic regression models (Fig. 3B), Boat was classified with high accuracy, whereas errors occurred more often for RP3 and, to a lesser extent, C2. The CNN (Fig. 3C) exhibited the same overall pattern, with particularly pronounced RP3 to Boat confusion, suggesting persistent overlap between these two conditions in the present dataset. Confusion matrices. Rows indicate the true class and columns indicate the predicted class. Values are row-normalised and reported as percentages, that is, each row sums to 100% and reflects the distribution of predictions given the true class. Classes correspond to the rowing condition: C2 (Concept2 static ergometer), RP3 (RP3 dynamic ergometer), and Boat (on-water rowing). The colour scale uses a blue gradient where darker shades indicate higher percentages and lighter shades indicate lower percentages. Panels are labelled (A) Ensemble, (B) Logistic Regression, and (C) CNN Condition-related differences in IMU features Condition-specific (C2, RP3, Boat) differences in selected IMU features are given in Table 1. At the athlete level, the rmANOVA revealed significant main effects of condition for nine out of ten magnitude features (F ≥ 8.36; FDR-adjusted p ≤ 0.003), with large effects (ηp² ≥ 0.48). The only non-significant feature was the gyroscope magnitude minimum (rmANOVA: F = 1.26; ηp² = 0.12; FDR-adjusted p = 0.307). Boat showed lower acceleration magnitude maxima (accmag_max) and ranges (accmag_range), but a higher acceleration magnitude minimum (accmag_min), than both ergometers (Fig. 4A). Acceleration magnitude mean (accmag_mean) was highest on C2, intermediate on RP3, and lowest on Boat, while acceleration magnitude median (accmag_median) separated C2 from RP3 and Boat. Gyroscope magnitude mean, median, maximum, and range were higher on both ergometers than on Boat. RP3 had the highest descriptive mean and median, whereas C2 had the highest descriptive maximum and range (Fig. 4B). Table 1Condition-specific differences in selected IMU magnitude features. Athlete-level values are shown as mean ± SD for each condition. Each repeated-measures analysis of variance tested the main effect of condition; its p value was Benjamini–Hochberg FDR-adjusted across ten omnibus tests. Pairwise paired t-tests report the mean difference (MD) with 95% confidence interval, the absolute paired-samples effect size |dz|, and the p value after Benjamini–Hochberg FDR correction across all 30 pairwise testsFeatureBoat (mean ± SD)RP3 (mean ± SD)C2 (mean ± SD)rmANOVA result (pFDR)Boat vs. RP3 (MD [95% CI]; |dz|; pFDR)Boat vs. C2 (MD [95% CI]; |dz|; pFDR)RP3 vs. C2 (MD [95% CI]; |dz|; pFDR) acc mag_min 8.24 ± 0.277.30 ± 0.527.46 ± 0.55F = 17.66; ηp² = 0.66; pFDR < 0.001MD = 0.94 [0.58, 1.29]; |dz| = 1.90; pFDR < 0.001MD = 0.78 [0.45, 1.11]; |dz| = 1.68; pFDR = 0.001MD = -0.16 [-0.61, 0.29]; |dz| = 0.25; pFDR = 0.560 acc mag_max 12.68 ± 0.7315.19 ± 1.7415.60 ± 1.86F = 16.31; ηp² = 0.64; pFDR < 0.001MD = -2.51 [-3.56, -1.46]; |dz| = 1.71; pFDR = 0.001MD = -2.91 [-4.13, -1.69]; |dz| = 1.71; pFDR = 0.001MD = -0.41 [-1.86, 1.04]; |dz| = 0.20; pFDR = 0.599 acc mag_range 4.44 ± 0.947.88 ± 2.208.14 ± 2.33F = 17.42; ηp² = 0.66; pFDR < 0.001MD = -3.44 [-4.82, -2.06]; |dz| = 1.79; pFDR < 0.001MD = -3.69 [-5.19, -2.19]; |dz| = 1.76; pFDR < 0.001MD = -0.25 [-2.08, 1.58]; |dz| = 0.10; pFDR = 0.764 acc mag_mean 10.07 ± 0.0510.19 ± 0.0810.45 ± 0.10F = 52.02; ηp² = 0.85; pFDR < 0.001MD = -0.12 [-0.17, -0.08]; |dz| = 2.12; pFDR < 0.001MD = -0.38 [-0.47, -0.29]; |dz| = 3.12; pFDR < 0.001MD = -0.26 [-0.37, -0.14]; |dz| = 1.62; pFDR = 0.001 acc mag_median 9.99 ± 0.059.98 ± 0.0810.25 ± 0.13F = 36.81; ηp² = 0.80; pFDR < 0.001MD = 0.01 [-0.04, 0.07]; |dz| = 0.18; pFDR = 0.613MD = -0.25 [-0.34, -0.17]; |dz| = 2.07; pFDR < 0.001MD = -0.27 [-0.36, -0.18]; |dz| = 2.12; pFDR < 0.001 gyr mag_min 0.09 ± 0.020.09 ± 0.010.08 ± 0.01F = 1.26; ηp² = 0.12; pFDR = 0.307MD = 0.00 [-0.01, 0.02]; |dz| = 0.18; pFDR = 0.613MD = 0.01 [-0.01, 0.03]; |dz| = 0.41; pFDR = 0.313MD = 0.01 [-0.01, 0.02]; |dz| = 0.43; pFDR = 0.292 gyr mag_max 1.99 ± 0.362.86 ± 0.703.09 ± 0.94F = 8.36; ηp² = 0.48; pFDR = 0.003MD = -0.87 [-1.26, -0.48]; |dz| = 1.60; pFDR = 0.001MD = -1.11 [-1.82, -0.39]; |dz| = 1.11; pFDR = 0.011MD = -0.24 [-1.00, 0.53]; |dz| = 0.22; pFDR = 0.583 gyr mag_range 1.89 ± 0.352.77 ± 0.693.01 ± 0.94F = 8.58; ηp² = 0.49; pFDR = 0.003MD = -0.88 [-1.27, -0.48]; |dz| = 1.58; pFDR = 0.001MD = -1.12 [-1.83, -0.41]; |dz| = 1.13; pFDR = 0.010MD = -0.24 [-1.01, 0.52]; |dz| = 0.23; pFDR = 0.583 gyr mag_mean 0.73 ± 0.101.01 ± 0.160.95 ± 0.08F = 35.61; ηp² = 0.80; pFDR < 0.001MD = -0.28 [-0.33, -0.22]; |dz| = 3.61; pFDR < 0.001MD = -0.23 [-0.29, -0.16]; |dz| = 2.41; pFDR < 0.001MD = 0.05 [-0.05, 0.16]; |dz| = 0.35; pFDR = 0.389 gyr mag_median 0.66 ± 0.100.90 ± 0.150.81 ± 0.07F = 33.19; ηp² = 0.79; pFDR < 0.001MD = -0.24 [-0.30, -0.19]; |dz| = 3.29; pFDR < 0.001MD = -0.15 [-0.21, -0.09]; |dz| = 1.82; pFDR < 0.001MD = 0.09 [0.00, 0.18]; |dz| = 0.74; pFDR = 0.066Abbreviations: IMU inertial measurement unit, C2 Concept2 ergometer, RP3 RP3 rowing ergometer, Boat on-water rowing, acc accelerometer, gyr gyroscope, mag Euclidean vector magnitude, min minimum, max maximum, mean arithmetic mean, median median, range maximum minus minimum, SD standard deviation, rmANOVA repeated-measures analysis of variance, MD mean difference, CI confidence interval, dz paired-samples Cohen’s d, FDR false discovery rateAcceleration magnitude features are expressed in m/s² and gyroscope magnitude features in rad/s; mean differences and confidence intervals use the corresponding units Condition-specific differences in selected IMU magnitude features. Athlete-level values are shown as mean ± SD for each condition. Each repeated-measures analysis of variance tested the main effect of condition; its p value was Benjamini–Hochberg FDR-adjusted across ten omnibus tests. Pairwise paired t-tests report the mean difference (MD) with 95% confidence interval, the absolute paired-samples effect size |dz|, and the p value after Benjamini–Hochberg FDR correction across all 30 pairwise tests Abbreviations: IMU inertial measurement unit, C2 Concept2 ergometer, RP3 RP3 rowing ergometer, Boat on-water rowing, acc accelerometer, gyr gyroscope, mag Euclidean vector magnitude, min minimum, max maximum, mean arithmetic mean, median median, range maximum minus minimum, SD standard deviation, rmANOVA repeated-measures analysis of variance, MD mean difference, CI confidence interval, dz paired-samples Cohen’s d, FDR false discovery rate Acceleration magnitude features are expressed in m/s² and gyroscope magnitude features in rad/s; mean differences and confidence intervals use the corresponding units Condition-specific differences in magnitude-based IMU features. Panel A shows z-standardised acceleration magnitude features and Panel B shows z-standardised gyroscope magnitude features for C2, RP3, and Boat. For each feature, z-scores for the small background points were computed across all individual strokes; the same feature-specific transformation was applied to the 30 athlete-condition means used for the summary points. Small background points represent individual strokes within the displayed axis range; larger points and horizontal error bars indicate the mean and standard error across the ten athlete-level condition means in each condition. Abbreviations: acc, accelerometer; gyr, gyroscope; mag, Euclidean vector magnitude; min, minimum; max, maximum; range, maximum minus minimum; median, median; mean, arithmetic mean; C2, static Concept2 ergometer; RP3, dynamic RP3 ergometer; Boat, on-water rowing Acceleration magnitude included the gravitational component. Absolute differences in accmag_mean were 0.12–0.38 m/s², despite large standardised effects. These effect sizes partly reflect the small between-athlete standard deviations and should not be interpreted as equally large practical biomechanical differences. A non-parametric Friedman sensitivity analysis confirmed the rmANOVA findings: the same nine of ten features were significant after FDR correction, while gyrmag_min remained non-significant; pairwise Wilcoxon signed-rank tests yielded the same FDR-level decisions as the paired t-tests (Supplementary Tables S2A and S2B). After FDR correction across all 30 pairwise tests, accmag_mean was the only feature that differed between all three conditions (|dz| ≥ 1.62; FDR-adjusted p ≤ 0.001). accmag_min, accmag_max, accmag_range, gyrmag_max, gyrmag_range, gyrmag_mean, and gyrmag_median differed between Boat and each ergometer (|dz| ≥ 1.11; FDR-adjusted p ≤ 0.011), but not between RP3 and C2 (FDR-adjusted p ≥ 0.066). accmag_median differed between C2 and both RP3 and Boat (|dz| ≥ 2.07; FDR-adjusted p < 0.001), but not between RP3 and Boat (FDR-adjusted p = 0.613). Figure 4 visualises the stroke-level distributions as small background points; larger points and horizontal error bars show the condition mean and standard error across the ten athlete-level condition means. Discussion This study examined whether trunk-mounted IMU signals can distinguish between C2, RP3, and Boat rowing conditions using classical machine learning on handcrafted features and deep learning on raw multichannel time series. The main finding was that models classified the three conditions above the balanced chance level, but class imbalance requires interpretation against the no-information rate and imbalance-robust metrics. The deep learning ensemble (Fig. 5) reached the highest mean accuracy of 85.5 ± 10.3%. However, model performance varied between athletes (Fig. 2), and athlete 7 showed clearly lower accuracy despite a visual recheck of the raw data without obvious recording irregularities. The confusion matrices (Fig. 3) showed that Boat was identified most reliably, whereas RP3 was frequently classified as Boat. These findings support condition-specific trunk movement signatures, but they do not prove mechanical equivalence between RP3 and Boat. Model comparison of condition-classification accuracy. All models used overlapping three-stroke windows. Classical models represented each window with concatenated stroke-level feature vectors, whereas deep-learning models used 600-time-step raw IMU sequences. Bars show mean athlete-level accuracy and error bars show the standard deviation. Brackets show the 95% confidence interval. Evaluation used leave-one-subject-out validation. SVM, support vector machine; CNN, convolutional neural network; LSTM, long short-term memory; KNN, k-nearest neighbours; XGBoost, extreme gradient boosting; CI, confidence interval; SD, standard deviation Group-level classification of rowing conditions The ensemble accuracy of 85.5 ± 10.3% (Fig. 5) indicates that C2, RP3, and Boat can be separated from trunk-mounted IMU data at the group level. This finding is plausible because rowing produces a highly structured cyclic signal, and (non-rowing) wearable sensor studies have repeatedly shown that acceleration and angular velocity signals contain enough temporal and magnitude-based information to distinguish movement classes [37, 44, 55]. The present result therefore extends general wearable activity recognition evidence to a rowing-specific question in which the target classes are not different activities, but different mechanical environments of the same sport movement. The magnitude of the observed accuracy aligns with previous inertial sensor classification work. Awais et al. [65] reported very high accuracy for both LSTM-based deep learning and feature-based support vector machine approaches when classifying common activities of daily living in older adults. Lee et al. [66] showed that a deep learning approach classified squat posture more accurately than conventional machine learning in an inertial sensor setup. In sport-specific applications, Link et al. [67] classified Ultimate Frisbee throwing techniques with convolutional neural networks, whereas Croteau et al. [68] used inertial sensors and machine learning to detect passes and shots in water polo. These studies differ from the present rowing design, but they support the core assumption that inertial signals can encode movement context and sport technique in a way that supervised models can learn. Importantly, the present findings should not be read as a simple demonstration that deep learning is generally superior to feature-based learning. Several classical models reached accuracies in the low to mid 80% range, which indicates that condition-specific information was already present in handcrafted time- and frequency-domain features. This agrees with Awais et al. [65], who did not show a universal advantage of deep learning over classical machine learning, and with Jalal et al. [37], who demonstrated that accelerometer and gyroscope feature sets can support robust activity recognition. The main conclusion is therefore that the movement signatures of the three rowing conditions were learnable at both the feature level and the raw time series level. The best performance of the probability-averaging ensemble is methodologically plausible, but it should be interpreted cautiously. Tsutsumi et al. [44] showed that model ensembles can improve sensor-based activity recognition, which supports the ensemble strategy used here. However, LSTM- and CNN-based references should primarily be used to justify temporal and local pattern learning rather than ensemble learning itself [53–55]. In the present study, the ensemble result is best interpreted as a robust group-level indicator that the three rowing environments produce separable signals, not as definitive evidence that one architecture is superior for rowing diagnostics. Inter-individual variability and individualised prediction Despite strong group-level performance, the athlete-level results revealed substantial inter-individual variability. Several athletes reached high accuracies across models, whereas athlete 7 remained clearly lower across approaches even though visual inspection did not reveal obvious sensor or data quality problems. Athlete 7 had no prior RP3 experience. This may have contributed to the lower performance, but it cannot explain errors in the other conditions or establish a causal relation. This pattern suggests that classification errors may reflect genuine differences in movement execution, equipment familiarity, anthropometry, developmental stage, or boat setup rather than only technical noise. Rowing depends on coordinated interaction between the lower limbs, trunk, and upper body, and previous biomechanical work has shown that changes in stroke rate, gearing, fatigue, and rowing context can alter movement and power-related characteristics of the rowing stroke [4–6, 18, 23, 69]. The athlete-specific variability also fits evidence from other cyclic sports. Martens et al. [70] showed high inter-individual variability in muscle activation patterns during front crawl swimming and cautioned against applying one overall reference pattern to individual technique guidance. Although swimming and rowing differ mechanically, this finding is relevant because both involve cyclic, technically refined movement patterns in which athletes may achieve similar task goals through individual coordination strategies. In the present study, a classifier trained on other athletes may have learned a dominant group pattern that did not fully represent the movement solution of athlete 7. This explanation remains speculative, but it is more consistent with the visually normal raw signals than an assumption of simple measurement failure. The relatively robust individual performance of LSTM-based models is also plausible. LSTM networks were developed to retain sequential dependencies over time, and hybrid LSTM and convolutional architectures are commonly used for multivariate time series classification because they can combine local feature extraction with temporal sequence modelling [53–55]. In rowing, where the timing of drive, finish, recovery, and catch is central to the stroke pattern, a model that preserves temporal dependencies may better accommodate athlete-specific timing differences than a model that relies only on summary features. Nevertheless, the observed variability shows that strong group-level classification does not automatically justify direct individual feedback. Applied use should therefore include athlete-specific calibration, repeated measurements, and prospective validation before the model informs technical or training decisions. Condition-specific classification patterns The confusion matrices (Fig. 3) showed a coherent condition-specific pattern. Boat was the most reliably identified condition, RP3 was the least separable and was most often confused with Boat, and C2 showed intermediate classification accuracy. This pattern supports the hypothesis that RP3 and Boat share more similar trunk movement signatures than C2 and Boat. It also argues against interpreting the RP3 to Boat errors as random model failure. Instead, the errors may partly reflect biomechanical overlap between dynamic ergometer rowing and on-water rowing. However, because only 40% of athletes had prior RP3 experience, unfamiliarity may also have altered the RP3 trunk signature and contributed to the observed confusion. The strongest direct support for this interpretation comes from Elliott et al. [17], who compared RP3 rowing with single scull rowing and reported high similarity in force curve patterns, with cross-correlations between combined on-water force curves and RowPerfect force curves above 0.90. This does not prove that trunk-mounted inertial measurement unit signals are similar between RP3 and Boat, but it shows that a dynamic ergometer can reproduce important aspects of the on-water force application pattern. Šarabon et al. [14] should be interpreted more narrowly. That study compared stable and compliant ergometer designs during simulated 2000 m rowing, not RP3 directly with on-water rowing. It is therefore useful as evidence that ergometer compliance alters biomechanical and physiological responses, but it should not be cited as direct proof of RP3 to Boat equivalence. The clearer separability of C2 from the other conditions is also biomechanically plausible. Benson et al. [20] reported that rowers used higher stroke rates and lower peak stroke forces on a dynamic ergometer than on a stationary C2 ergometer while maintaining similar power output. Greene et al. [23] further showed that fixed foot-stretcher ergometry increases inertial demands at the catch and affects knee moment, joint power, and coordination patterns. Vinther et al. [22] similarly reported lower peak force and altered neuromuscular activity during slide-based compared with stationary ergometer rowing. These findings support the interpretation that the static C2 imposes a distinct mechanical environment that can be detected by trunk inertial signatures. The fact that Boat was classified most reliably also deserves a cautious interpretation. On-water rowing differs from ergometer rowing not only because the rower and boat interact dynamically, but also because environmental and equipment factors are less standardised. Wilson et al. [18] observed different lumbar spine motion between ergometer and single scull rowing, and Vogler et al. [25] reported physiological differences between ergometer and on-water incremental rowing tests. In the present study, the Boat condition also involved individually adjusted boats and six boat types, whereas the ergometer conditions used standardised devices. The high Boat accuracy may therefore reflect true on-water movement characteristics and additional equipment-related variability. This does not invalidate the classification result, but it limits how strongly the result can be interpreted as a pure biomechanical difference between rowing modalities. Magnitude-based feature differences and biomechanical interpretation The feature analyses provide a mechanistic layer beneath the classification results. Nine of ten magnitude-based features differed between conditions after FDR correction, but accmag_mean was the only feature that separated all three conditions in the pairwise tests. gyrmag_median was descriptively highest on RP3, followed by C2 and Boat, although the RP3–C2 difference did not remain significant after FDR correction. The higher accmag_mean on C2 may reflect the repeated acceleration and deceleration demands placed on the rower when the flywheel and foot-stretcher remain fixed. This interpretation is consistent with fixed-ergometer studies reporting increased inertial demands, altered joint power, and higher force-related demands than in dynamic or slide-based systems [20, 22, 23]. Wilson et al. [18] also showed greater lumbar flexion during ergometer than single-scull rowing in a step test. The present IMU data extend these findings by showing condition-specific trunk acceleration differences, but the small absolute differences and single-sensor design preclude direct inference about local joint mechanics or technical quality. The higher gyrmag_median on RP3 is more difficult to interpret and should be framed cautiously. Because 60% of the athletes had no prior RP3 experience, less consolidated technique in this unfamiliar condition may also have increased trunk angular motion; the signal therefore cannot be attributed solely to mechanical stabilisation demands. Because the gyroscope magnitude reflects angular velocity at the trunk, the RP3 pattern may indicate greater angular movement or stabilisation demands when the flywheel and foot-stretcher unit can move relative to the athlete. This interpretation is compatible with reports that dynamic or slide-based ergometer rowing changes stroke rate, force production, and neuromuscular activity compared with stationary ergometer rowing [20–22]. However, the present single trunk sensor cannot identify the exact source of this angular signal. It cannot separate pelvic motion from thoracic motion, nor can it quantify local joint torques or muscular stabilisation directly. The lower acceleration and gyroscope magnitudes in Boat suggest smoother trunk movement signatures during on-water rowing than during both ergometer conditions. This interpretation is consistent with the idea that the athlete, boat, oars, and water form a coupled system rather than a body moving against a fixed resistance unit. Elliott et al. [17] showed that dynamic ergometry can approximate aspects of single scull force curves, whereas Wilson et al. [18] showed that single scull rowing differs from ergometer rowing in lumbar kinematics. The present results do not prove that lower trunk magnitudes are inherently better or more technically efficient. They show that the Boat condition generated a distinct inertial signature that classifiers could recognise and that magnitude features could quantify. Because all sensors were placed at the trunk, the feature results should be interpreted as global movement signatures rather than direct measures of local force production, joint kinematics, or segment-specific technique. Wearable sensor reviews in cyclic sports emphasize that field-based inertial measurements can provide useful estimates and movement descriptors, but they also require careful interpretation because sensor position, modelling assumptions, and validation context shape the meaning of derived variables [71]. In rowing, where force and power arise from coordinated lower limb, trunk, and upper body action, trunk signals capture only part of the full biomechanical chain [4–6]. Practical implications The main practical value of the IMU-only analysis is not identifying Boat, which GNSS can distinguish directly, but testing whether trunk movement separates C2 from RP3. Performance was weaker for this distinction than for Boat, so the present results support retrospective methodological analysis rather than immediate automated logging or real-time coaching. Wearable IMU systems can support technical feedback in other cyclic sports such as swimming [72], but rowing deployment still requires athlete-specific calibration and validation against independent biomechanical or coaching criteria. The condition-specific features may nevertheless help describe equipment-related movement signatures. Strengths and limitations A major strength is the direct comparison of static ergometer, dynamic ergometer, and on-water rowing within one measurement framework and under leave-one-subject-out evaluation. Several limitations remain. The sample comprised only ten youth rowers and was heterogeneous in experience, anthropometry, developmental stage, and RP3 familiarity; only 40% had prior RP3 experience. Boat was always assessed first, and Boat used six individually rigged boat types whereas the ergometers were device-standardised. Drag settings were self-selected. These factors may have contributed to class separation. One trunk-mounted IMU cannot resolve oar, pelvis, joint, or force mechanics. Stroke detection was checked against synchronised video and RowerUp outputs [36] for plausibility but was not validated quantitatively against frame-by-frame reference annotations. Finally, adjacent three-stroke windows overlapped and were not statistically independent, and only nine athletes were available for training in each fold. Future directions An important priority is multi-sensor validation. Additional sensors on the pelvis, boat shell, oar, foot-stretcher, or seat could identify whether RP3 to Boat confusion arises primarily from trunk timing, lower limb mechanics, equipment motion, or oar-related dynamics. This direction aligns with broader evidence that wearable sensors can estimate or approximate mechanical outputs in cyclic sports, but that sensor placement and modelling assumptions strongly influence interpretability [71]. Combining inertial sensors with force or power measurement would also connect automated condition recognition to rowing diagnostics that coaches already use [3, 25]. Finally, future machine learning work should test individualized and transfer learning strategies. Transfer learning has been explored in sport-specific wearable classification with small datasets, but its benefits should be demonstrated for rowing rather than assumed [67]. Athlete-specific LSTM models, calibrated over multiple sessions and validated against expert annotation, may eventually support on-water feedback. Before that step, however, researchers must determine whether the models remain accurate across fatigue states, stroke rates, training intensities, boats, and environmental conditions. Conclusion Trunk-mounted IMU data allowed classical machine-learning and deep-learning models to distinguish C2, RP3, and Boat at group level. RP3 and Boat showed the strongest confusion, but this pattern may reflect both biomechanical overlap and confounding from RP3 familiarity, measurement order, and asymmetric boat standardisation. C2 showed the highest mean acceleration magnitude and RP3 the descriptively highest median gyroscope magnitude. Athlete-level variability and class imbalance preclude immediate diagnostic use; broader prospective validation and athlete-specific calibration are required. This interpretation should be read as RP3-specific rather than as evidence for all dynamic ergometers. The present design tested one RP3 Model T device, and generalisation to other dynamic ergometers remains an inference. Supplementary Information Additional file 1: Supplementary Tables S1A-S3 report equipment settings, stroke and window counts, non-parametric sensitivity analyses, and classification metrics; Supplementary Figure S1 shows the remaining row-normalised confusion matrices. Additional file 2: Supplementary Code S1, illustrative Python workflow. Privacy-preserving illustrative code using deterministic synthetic signals to schematically represent the reported analysis stages and all ten model approaches. The file contains no study data, participant identifiers, original paths, fitted study models, or study predictions. Acknowledgements We thank the rowers from the regional training centres Breisgau-Hochrhein and Karlsruhe who participated in the study. Abbreviations Concept2 static ergometer Convolutional neural network False discovery rate Global navigation satellite system Inertial measurement unit K-nearest neighbours Leave-one-subject-out Long short-term memory RP3 Model T dynamic ergometer Repeated-measures analysis of variance Support vector machine Authors’ contributions A.L. and W.R. conceptualized the study and conducted the experiment. S.H. programmed the machine learning and deep learning models. S.H., T.S. and A.L. analysed and interpreted the data. S.H., A.L. and T.S. wrote the first manuscript draft. W.R., A.L., T.S. and S.H. edited and revised manuscript. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability Availability of data and materialsThe aggregate and derived results supporting the findings are reported in the manuscript and supplementary materials. Raw HDF5 and CSV data are not publicly available because the study involved minor athletes and the recordings contain potentially privacy-sensitive athlete-level information. Further data may be available from the corresponding author upon reasonable request, subject to applicable privacy, consent, and ethics constraints. Additional file 2 (Supplementary Code S1) provides privacy-preserving, executable illustrative code that schematically represents the reported analysis stages and all ten model approaches using deterministic synthetic signals. It contains no study data and is not the original study program or a reproduction package. Parameters explicitly reported in the manuscript or recoverable from the archived analysis are represented directly; remaining implementation details are transparent illustrative defaults. The code cannot recreate the reported numerical results without the protected source data, authoritative intermediate inputs, original fitted model states, and original execution environment. Declarations Ethics approval and consent to participate The study was conducted in accordance with the Declaration of Helsinki. Data collection took place on 10, 12, 13, and 28 October 2024. Written informed consent was obtained from the parents or legal guardians of all participants before participation, and all athletes and guardians received written and verbal information about the study aims, design, potential benefits, and risks. Following data collection, the Internal Ethics Committee of IST University of Applied Sciences granted positive ethical approval for the exact study procedures, measurements, Archinisis sensor system, instrumentation, data analyses, and consent arrangements reported in this manuscript (approval no. IR100726IST233, issued on 10 July 2026). Consent for publication Written informed consent for publication of the anonymised images in Fig. 1 was obtained from the parents or legal guardians of the individuals shown. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Additional file 1: Supplementary Tables S1A-S3 report equipment settings, stroke and window counts, non-parametric sensitivity analyses, and classification metrics; Supplementary Figure S1 shows the remaining row-normalised confusion matrices. Additional file 2: Supplementary Code S1, illustrative Python workflow. Privacy-preserving illustrative code using deterministic synthetic signals to schematically represent the reported analysis stages and all ten model approaches. The file contains no study data, participant identifiers, original paths, fitted study models, or study predictions. Data Availability Statement Availability of data and materialsThe aggregate and derived results supporting the findings are reported in the manuscript and supplementary materials. Raw HDF5 and CSV data are not publicly available because the study involved minor athletes and the recordings contain potentially privacy-sensitive athlete-level information. Further data may be available from the corresponding author upon reasonable request, subject to applicable privacy, consent, and ethics constraints. Additional file 2 (Supplementary Code S1) provides privacy-preserving, executable illustrative code that schematically represents the reported analysis stages and all ten model approaches using deterministic synthetic signals. It contains no study data and is not the original study program or a reproduction package. Parameters explicitly reported in the manuscript or recoverable from the archived analysis are represented directly; remaining implementation details are transparent illustrative defaults. The code cannot recreate the reported numerical results without the protected source data, authoritative intermediate inputs, original fitted model states, and original execution environment.