From intradiscal pressure to multimodal estimation of lumbar spinal loads Authors: Asghar Rezaei, Chih-Hsiu Cheng, Mohammad Nikkhoo, Babak Dashtdar, Lichun Lu, Kenton Kaufman Published: 2026 ABSTRACT BackgroundEstimation of lumbar spinal loads is important for understanding low back pain, guiding ergonomic interventions, and informing surgical and rehabilitation planning. Historically, intradiscal pressure (IDP) provided one of the few internal in vivo measures of disc loading; more recently, telemetry, musculoskeletal (MS) modeling, finite element (FE) analysis, hybrid MS–FE approaches, displacement/control-based methods, and AI surrogates have expanded the toolbox for estimating spinal loads.Material and methodsWe present a narrative perspective review based on a literature search in PubMed, Scopus, and Web of Science using terms related to spinal loads, IDP, telemeterized implants, MS modeling, FE analysis, hybrid MS–FE coupling, displacement/control-based methods, wearable/EMG-based approaches, and AI/machine learning surrogates. Human lumbar studies and methodological contributions relevant to load estimation or validation were included; animal models were excluded.ResultsInvasive approaches (needle-based IDP, discography, intra-abdominal pressure, and telemeterized implants) provide task-dependent internal pressures or forces in small, selected cohorts and now primarily serve as benchmarks for model validation. MS models estimate segmental compression, shear, and net moments from motion and EMG, with typical L4–L5 compressive forces of ∼1–2 kN in relaxed standing and ∼3–5 kN during common lifting tasks. FE and hybrid MS–FE simulations resolve how these loads are distributed across discs, facets, and ligaments and relate segmental forces to internal stresses. Displacement-driven/control-based models and emerging AI/wearable-based surrogates provide additional non-invasive pathways for task-specific lumbar load estimation.ConclusionsThis methods-focused synthesis outlines how invasive data support MS, FE, hybrid, and AI-based approaches and highlights recurring challenges in muscle redundancy, constitutive and parameter uncertainty, limited in vivo benchmarks, and heterogeneous model reporting. Within this framework, IDP is best regarded as an internal benchmark rather than a stand-alone metric of “spinal load” which is more fully described by compression, shear, moments, and internal stresses. Introduction Back pain is a prevalent clinical issue associated with significant economic, social, and psychological burdens, and it is also a leading reason for seeking emergency medical care [1]. Up to 80% of the population will experience at least one episode of low back pain (LBP) during their lifetime [1]. Although most LBP cases are classified as nonspecific, symptoms are typically provoked or relieved by mechanical factors such as posture, loading, and movement. Accordingly, quantifying lumbar spinal loads, typically expressed as joint compression, shear, and net moments, offers valuable insight into how mechanical exposure contributes to the onset and persistence of LBP [2]. Researchers have therefore been endeavoring for decades to measure spinal loads to better understand spinal biomechanics and the mechanisms underlying back pain [3].Spinal loads are transmitted through intervertebral discs (IVDs) and the posterior elements, with a substantial portion of the compressive load borne by the discs and the remainder shared by facet joints, ligaments, and bony structures [4,5]. Historically, many studies have measured intradiscal pressure (IDP) to probe disc behavior during physical activities [4]. IDP reflects the internal stress state within the nucleus and inner annulus and provides an important internal benchmark, but it is only one component of the overall load picture and does not, by itself, describe the full distribution of spinal loads across tissues. Due to the invasive and complex nature of direct pressure or force measurements, several non-invasive experimental and computational techniques have been developed to estimate spinal loads during various physical activities. Musculoskeletal (MS) models estimate segmental compression, shear, and net moments from motion, external forces, and EMG; finite element (FE) models resolve how these loads are distributed within discs, facets, and ligaments; and hybrid MS–FE, telemetry, and AI-based approaches extend these estimates to subject-specific and real-world contexts. Together, invasive measurements and non-invasive estimations provide complementary perspectives on how loads are transmitted across the spinal column and how spinal tissues respond under different postures and tasks. This review synthesizes prior work on lumbar spinal loads by outlining key invasive and non-invasive methods used to estimate and interpret lumbar spinal loading. The paper focuses on lumbar spinal loadings from in vitro, in vivo, and in silico studies to compare methodological approaches, identify their respective strengths and limitations, and highlight remaining challenges and knowledge gaps in spinal load estimation. Methods – scoping review approach We performed a narrative perspective review to summarize methodological advances in estimating lumbar spinal loads. Searches were conducted in PubMed, Scopus, and Web of Science up to September 2025 using keyword families centered on (1) spinal loads (e.g., spinal load, spinal loading, lumbar loading), (2) internal in-vivo measures (e.g., intradiscal pressure, telemeterized implant, instrumented vertebral body replacement), (3) computational approaches (e.g., MS model/modeling, inverse dynamics, EMG, optimization, FE, constitutive, hybrid, coupled, displacement-driven, control-based), and (4) emerging technologies (e.g., wearable, AI, machine learning, neural network). Human studies, methodological papers, and validation studies directly relevant to lumbar loading were included; animal models were excluded. Publications from 2010 onward and seminal earlier sources were prioritized, and reference lists of key articles and reviews were hand-screened to identify additional directly relevant methodological studies that might not be captured by the primary keyword combinations. As a perspective review, this work does not aim for PRISMA-style exhaustiveness; instead, it offers a structured synthesis of representative, up-to-date exemplars across invasive, model-based, and AI/technology-assisted methods for assessing spinal loading. Invasive spinal load measurements In vitro studies Many early studies of spinal loads were conducted in vitro because cadaveric testing allows controlled loading and direct instrumentation under conditions that are not feasible in vivo, despite the practical and ethical challenges associated with obtaining human specimens [6]. These studies have provided significant opportunities to investigate IDP up to the point of vertebral fractures or disc failures under multiple loading conditions, allowing for a thorough assessment of IVD mechanics and its failure [[6], [7], [8]]. Efforts to understand spinal loads and the origins of LBP date back to the 1950s [9]. Nachemson [9] was the first to use a mechano-electrical pressure sensor to measure the disc pressure by inserting an instrumented needle inside the nuclei of cadaveric IVDs. Most cadaveric studies have utilized needle sensors, which are percutaneously inserted into the nucleus pulposus (NP) to measure IDP [9]. These studies have measured pressure in both horizontal and vertical directions by rotating the needle transducer to assess if the pressure was hydrostatic [9,10]. Although pressure sensors have seen technological advancements, transitioning from mechano-electrical sensors to more advanced options like piezoelectric transducers [8,9,11], the method for locally measuring IDP in human or cadaveric subjects has remained largely unchanged. Several studies have affixed pressure sensors to a device that pulls a needle through the disc, enabling them to measure IDP across an anterior-posterior line and assess the pressure profile inside a disc [10]. The profilometry technique showed a uniform distribution of pressure (in both directions) inside a healthy nucleus and inner layer of the AF. The pressure gradually dropped to zero in the outer layer of the annulus. These profiles also changed with physical activities. The fluoroscopic-guided lumbar discography technique has also been employed in several studies to investigate IDP. Using an injecting pump, this technique introduces a contrast agent into the NP to measure the intrinsic pressure (IDP at which the contrast agent is just entered into the NP) and the maximum or failure pressure of the AF (IDP at which the contrast agent leaks out of the NP) [12]. These studies have shown similar pressure data compared with those studies using needle sensors and provide insight into the annulus failure depending on the condition of the disc. Pressure measurements in cadaveric studies have also been used to evaluate the performance of various clinical interventions such as nucleoplasty [13], spine manipulation [14], distraction therapy for LBP [15], and instrumentations and implants [16]. Despite several limitations such as ignoring muscle interactions or not accurately replicating physical activities, post-mortem changes of IVD materials, cadaveric studies have served as the foundation for IDP measurements in human subjects and LBP research. In vivo studies Several studies have expanded their investigation of spinal loads to human subjects to assess IVD reactions through needle pressure sensors during a wide range of physical activities including various lying, sitting, standing, walking, lifting, and bending movements [[17], [18], [19], [20], [21], [22], [23], [24]]. [Fig. 1] shows previously measured IDP as well as estimated intradiscal compressive load during various physical activities and postures. A simplified relationship has often been used to relate compressive force to IDP based on the following formula [25]:(1)F=IDP×A×kwhere A is the effective disc cross-sectional area and k is a proportionality factor accounting for the non-uniform transmission of load across the disc. Because A varies across spinal levels and between individuals, and because k was primarily established for healthy or only mildly degenerated discs, this relationship should be interpreted cautiously—particularly in aging or advanced degeneration, where the coupling between IDP and compressive load becomes less reliable. Using the ranges for A and k reported in literature [22,25], compressive loads in lying postures are typically low and generally do not exceed about 300 N, whereas relaxed standing corresponds to a compressive load of roughly 600 N. The highest reported load, approximately 2800 N, was associated with lifting a 20-kg weight while bending forward.Fig. 1Compressive loads estimation derived from IDP measured in IVDs during various physical activities and postures. Compressive loads estimation derived from IDP measured in IVDs during various physical activities and postures. Some in-vivo studies also assessed clinical interventions such as IDP changes during passive and active tractions [26], vertebral axial decompression on IDP [27], or spinal manipulation [28]. A number of studies used discography to assess the effect of degeneration on disc compliance [29] or measure internal (intrinsic) pressure during different postures [30]. Later studies also included instrumented implants that incorporated tiny load cells to record forces and moments applied to the implant as a measure of spinal loads [31]. While these instruments directly measure force and moment data, the technique is invasive and limited to patients who undergo spinal surgeries. In vivo IDP and force data associated with different postures and activities have provided insights into the physiological responses of IVDs and the spinal column under various loading conditions. Measuring IDP remains challenging due to its invasive nature and reliance on both the measurement technique and the individual's characteristics. However, most of these measurements were only performed on small numbers of subjects. Bashkuev et al. [32] highlighted that the pressure sensor can affect the measured values, creating significant artifacts. It was also reported that factors like lumbar shape, body mass, and height impact IDP measurements. Most importantly, while some studies have included individuals with LBP or spinal pathologies, these studies only focused on differentiating IDP in normal and degenerated discs, regardless of patient spinal pathologies. In addition to direct pressure recordings, some studies have adopted indirect in vivo surrogates to assess spinal loading. One approach uses intra-abdominal pressure (IAP) as a correlate of IDP or net spinal loads [33,34]. These studies reported significant correlations between IAP and IDP or spinal loads, suggesting that IAP can act as a global indicator of trunk loading. However, measurements of IAP (e.g., via gastrointestinal or bladder catheters) are themselves invasive and are highly sensitive to abdominal and respiratory muscle activity, including breathing pattern, Valsalva maneuvers, abdominal bracing or belt use, and posture [34]. Similar to IDP recordings, this muscle-dependent behavior limits the ability of IAP measurements to provide level-specific spinal loads or to disentangle contributions from different spinal structures [34]. A second approach uses diurnal stadiometry, in which changes in body height over the day are interpreted as cumulative disc compression due to sustained loading and fluid exudation [4]. Stadiometry is non-invasive and inexpensive, but accurate stature measurement is challenging and confounded by hydration, gait/activity patterns, and contributions from non-spinal joints. Moreover, it provides no regional or level-specific information and is insensitive to asymmetric/lateral loading. Beyond these indirect surrogates, in vivo spinal loading can also be characterized more directly using telemeterized implants [35]. One specific example is the use of instrumented vertebral body replacements, which provide direct measurements of implant-transmitted compressive loads [35]. These measurements reveal clear task- and posture-dependent load profiles at the operated level and complement IDP by capturing non-disc load paths routed through posterior elements and instrumentation. However, they quantify only the fraction of total spinal load carried by the device; residual spinal structures continue to bear load, study cohorts are necessarily postoperative, and device-specific calibration/uncertainty limits generalizability. Despite these constraints, implant telemetry remains the most direct in-vivo load measurement available and provides high-value benchmarks for validating non-invasive MS predictions and constraining their boundary conditions [35]. Additionally, some in vivo studies attempt to investigate how spinal loads affect IVDs using magnetic resonance imaging (MRI). In these studies, patients experience weight-bearing loads during MRI scans to investigate the lumbar spine deformation during loading and to assess spinal disorders such as spinal stenosis [36]. Related MRI-based techniques have also been developed to quantify internal disc deformation and strain non-invasively, for example by combining MRI with digital volume correlation [37]. While these studies have provided valuable insights into the responses of the IVD and the lumbar spine under loading conditions, none of these studies directly measure spinal loading. Non-invasive spinal load estimations The concerns related to invasiveness and cost associated with in-vivo and in-vitro methods for quantifying spinal loads have prompted the adoption of biomechanical modeling approaches. These models can simulate conditions and address inquiries that are challenging or even unfeasible to tackle using experimental techniques [3]. MS modeling is a commonly employed tool for obtaining a macroscopic understanding of both the normal and abnormal biomechanics of the spine. MS models analyze segmental joint reaction forces (compression and shear), net moments, muscle forces, and muscle activation patterns in conjunction with corresponding movements or positions of the spine. These models offer a versatile platform to explore a wide range of scenarios, enabling a better understanding of the intricate interactions within the spine's MS system and contributing to advancements in the field of spinal biomechanics [3]. Within these models, forces acting on both active (trunk muscles) and passive (ligamentous spine) structures are computed by maintaining a dynamic equilibrium between external moments (such as those caused by inertia, gravity, or external loads) and internal moments generated by the trunk's active and passive components [38]. This equilibrium can be established at a single level or simultaneously at multiple spinal levels. When analyzing the equilibrium of forces acting on the spine, we encounter significant challenges due to the multitude of unknown muscle forces. These forces exceed the number of available equilibrium equations, leading to a highly redundant problem. As a result, it becomes impossible to achieve deterministic solutions for the equilibrium equations [39]. This complexity emphasizes the necessity for alternative modeling strategies and load-related metrics to characterize how muscle recruitment and co-contraction influence spinal loading [39]. To address this obstacle, researchers have introduced a variety of biomechanical methodologies for resolving muscle redundancy. These approaches can be broadly grouped into: (1) single-equivalent muscle approaches, (2) optimization-driven models, (3) electromyography (EMG)-driven models, and (4) hybrid methods that combine EMG information with optimization (EMGAO). In practice, however, these categories are not completely discrete, because some optimization-based musculoskeletal frameworks can incorporate muscle activation patterns or activation constraints to improve physiological realism while preserving computational efficiency. Each approach has its own advantages and limitations, making it more suitable for some applications than others [3,38,[40], [41], [42], [43], [44]]. Single-equivalent muscle models simplify muscle recruitment by aggregating agonist muscles into a single representative force while often disregarding antagonists, making them computationally efficient and suitable for ergonomic or occupational studies, though limited in accuracy because they fail to capture individual activation patterns, co-contraction, or detailed load-sharing [40,41]. Optimization-based models resolve redundancy by formulating an optimization problem in which the central nervous system is assumed to minimize physiological criteria such as energy expenditure, muscle stress, or force variability [3,45]. Depending on the implementation, these models may also incorporate activation-pattern constraints or EMG-informed weighting to improve plausibility without requiring fully EMG-driven calibration. While widely applied and capable of producing detailed spinal load estimations, these models can be computationally demanding, particularly for dynamic tasks involving many muscles [38,45]. EMG-driven models provide an alternative by directly using recorded activation signals to generate musculotendon forces, offering subject-specific estimations of neuromuscular recruitment without relying on assumed cost functions [46]. However, they require careful calibration, are sensitive to signal noise and electrode placement, and cannot reliably capture deep muscle activity [47]. To overcome these limitations, EMGAO models integrate EMG information into optimization frameworks, enhancing physiological realism while retaining computational efficiency [48,49]. Notably, Samadi et al. [49] developed a stability/EMG-driven (SEMG) model of the thoracolumbar spine, which combines EMG information with additional constraints to refine muscle force and spinal load estimates during static tasks. This hybrid approach refines force predictions by constraining recruitment patterns with neural activation data, thereby reducing reliance on assumed cost functions. Overall, EMGAO and SEMG models have been shown to provide more accurate and individualized estimations of spinal loads than either optimization- or EMG-based methods alone, and they have been successfully applied to lifting, walking, and static postures, demonstrating their potential for broader clinical and occupational applications [[42], [43], [44]]. In addition to full MS models, several groups have derived reduced-order, data-driven predictive equations for spinal loads from large sets of MS simulations or experimental measurements. These models typically express lumbar compression and shear forces, or net segmental/lumbar moments, as functions of a small set of task variables such as trunk flexion angle, hand-load magnitude and distance, lifting asymmetry, and basic anthropometrics (e.g., body mass, stature) [50]. Once calibrated, such regression-based equations enable rapid estimation of spinal loads for common lifting or material-handling tasks without re-running full simulations, making them attractive for ergonomic risk assessment, workplace design, or screening [51,52]. However, because they are usually developed for specific postures, load ranges, and populations, their validity is task- and context-dependent, and they cannot easily capture complex three-dimensional movements, muscle co-contraction strategies, or individual variations in neuromuscular control [53,54]. These reduced-order equations therefore complement, rather than replace, full MS/FE models and newer ML-based approaches in the toolbox for spinal load estimation. In several studies, these model-predicted loads have been additionally mapped to disc pressure via simplified load–pressure relationships or embedded FE disc submodels solely to benchmark the non-invasive load estimates against classic invasive IDP data. [Fig. 2] therefore summarizes L4–L5 shear and compressive forces as well as IDP values derived from MS-based spinal load models (or coupled MS–FE approaches) and compares them with the in vivo IDP measurements reported by Wilke et al. [55]. In this context, the primary quantities of interest are segmental compression, shear, and net moments; IDP is used as an internal validation signal rather than a clinical endpoint. This comparison illustrates how different non-invasive load estimation pipelines reproduce (or deviate from) the canonical invasive IDP benchmark, thereby linking model-based spinal load predictions to internal experimental reference data. Across the studies, tasks that generate higher IDP also correspond to higher compressive loads and moments, reinforcing the plausibility of the predicted loading patterns. Recent model predictions reveal that increased external loads and moments necessarily require elevated muscle activities and increased trunk stiffness to maintain equilibrium [56]. These neuromuscular adaptations, which are consistent with classic telemeterized-implant and ergonomic findings, counterbalance the additional forces but also increase segmental compression and shear, thereby elevating tissue loading and fatigue risk [56,57]. A strong association between abdominal muscle co-activity (force coactivity ratio) and task-related load margins at the hands underscores the important role of co-contraction in modulating spinal loading. Consequently, co-activity indices and trunk stiffness can serve as useful load-related metrics when interpreting model-based spinal load estimates [58]. Neuromuscular adaptations influence the postural behaviors of workers during lifting tasks in response to fatigue. These adaptations lead to notable variations in kinematics, kinetics, and EMG-based indices, shedding light on the adaptive changes in lifting behavior under dynamic fatiguing conditions [59].Fig. 2Comparison of L4–L5 compression force, shear force, and IDP values derived from musculoskeletal-based spinal load models (and coupled MS–FE approaches) with invasive in vivo IDP measurements reported by Wilke et al. [55]. In the modeling studies, segmental compression, shear, and net moments were first estimated from task kinematics and external loads; IDP was then computed as a derived internal variable (via load–pressure relationships or FE disc sub-models) solely to benchmark these non-invasive load estimates against the invasive IDP data. Tasks: (1) relaxed upright stance with no hand load, hands ≈15 cm anterior to L5–S1; (2) upright stance holding 19.8 kg at 25 cm anterior to L5–S1; (3) upright stance holding 19.8 kg at 55 cm anterior to L5–S1; (4) 50° forward flexion with hands ≈25 cm anterior to L5–S1; (5) 70° forward flexion holding 19.8 kg, hands 32 cm anterior to L5–S1; (6) peak voluntary flexion of 110° with hands 40 cm anterior to L5–S1; (7) asymmetric one-handed lift of 19.8 kg on the left side, ≈34 cm lateral and 0 cm posterior to L5–S1 in upright stance. Comparison of L4–L5 compression force, shear force, and IDP values derived from musculoskeletal-based spinal load models (and coupled MS–FE approaches) with invasive in vivo IDP measurements reported by Wilke et al. [55]. In the modeling studies, segmental compression, shear, and net moments were first estimated from task kinematics and external loads; IDP was then computed as a derived internal variable (via load–pressure relationships or FE disc sub-models) solely to benchmark these non-invasive load estimates against the invasive IDP data. Tasks: (1) relaxed upright stance with no hand load, hands ≈15 cm anterior to L5–S1; (2) upright stance holding 19.8 kg at 25 cm anterior to L5–S1; (3) upright stance holding 19.8 kg at 55 cm anterior to L5–S1; (4) 50° forward flexion with hands ≈25 cm anterior to L5–S1; (5) 70° forward flexion holding 19.8 kg, hands 32 cm anterior to L5–S1; (6) peak voluntary flexion of 110° with hands 40 cm anterior to L5–S1; (7) asymmetric one-handed lift of 19.8 kg on the left side, ≈34 cm lateral and 0 cm posterior to L5–S1 in upright stance. Displacement-driven models, which estimate spinal loads using vertebral displacements measured from medical imaging, avoid the need for muscle force calculations via optimization or EMG-based approaches. These models use direct kinematic input, making them distinct from conventional force-controlled methods. A study using Monte Carlo simulations found that small errors in measured vertebral translations (0.1–0.3 mm) significantly impacted spinal load predictions, particularly IDP and shear forces at L5-S1, suggesting that current imaging accuracy may be insufficient for reliable biomechanical modeling [60]. A more recent study introduced a force–displacement control FE model, incorporating in vivo vertebral rotations and follower loads. This approach improved predictions of disc compression forces, intradiscal pressure, and ligament forces compared to traditional passive FE models but remained limited in estimating shear and facet joint forces [61]. While displacement-driven models reduce reliance on assumed muscle activation patterns, their accuracy depends on imaging precision and physiological load integration. They offer a promising alternative for spinal load estimation, especially when muscle force data is unavailable. FE spine models aim to replicate spinal deformation and internal stress distributions and can represent the whole spine, specific regions, or individual motion segments. Their geometry may be obtained in different ways: some models are reconstructed from medical imaging for subject-specific applications, whereas many others use idealized or parametric representations derived from imaging datasets, anatomical literature, or simplifying assumptions. Material properties of spinal tissues are defined through constitutive formulations: hyperelastic models capture nonlinear elastic behavior, while both viscoelastic and poroelastic formulations describe time-dependent responses. Viscoelastic models representing the time-dependent behavior of the solid matrix, and poroelastic models additionally capturing coupled solid–fluid interactions and load-driven fluid transport. Hybrid approaches are often used for more complete disc modeling [[62], [63], [64]]. Validation of these models is generally conducted through comparison with experimental in vitro measurements, such as IDP, range of motion, or stiffness curves, as well as in vivo clinical observations [65,66]. Accordingly, the choice of constitutive laws, boundary conditions, and validation metrics directly influences the accuracy of predicted compressive and shear forces, facet loads, and internal stress states (including IDP when calculated), and thus determines the clinical and ergonomic utility of FE models. FE simulations hold the potential to discern and elucidate the distinct contributions of various lumbar spine elements to the load distribution [67]. In recent years, this technique has offered valuable insights into the intricate mechanics of load sharing among different structural elements. These studies highlight the critical role of discs in resisting muscle-generated moments during standing, while ligaments exert a relatively lesser influence on moment distribution [68]. Facet joints typically carry only a small fraction of compressive load in neutral standing, but FE and hybrid models show that facet contact forces can increase substantially in extension, axial rotation, and with disc height loss, shifting part of the compressive and shear load from the disc to the posterior elements [65,[68], [69], [70]]. Ligaments (e.g., posterior longitudinal, capsular, interspinous and supraspinous) remain minimally loaded near neutral but develop appreciable tensile forces as the motion segment approaches the limits of flexion, extension, or lateral bending, thereby constraining motion and contributing to overall load sharing [67,68,71]. These findings highlight that “spinal load” is distributed among discs, facets, ligaments, and bony structures, and that the relative contribution of each structure depends on posture, loading direction, and degeneration. When computed, IDP is one component of the FE-predicted internal state and can be used, together with segmental forces, facet contact forces, and ligament tensions, to assess load-bearing capacity and local stress patterns under different loading scenarios [66,72]. Multiple FE models have revealed that the highest IDP values occur during flexion, ranging from approximately 1 to 1.58 MPa, while IDP values during extension, lateral bending, and axial rotation do not exceed 1 MPa [25,69,[73], [74], [75], [76], [77], [78], [79]]. [Fig. 3] summarizes the estimated IDP in the L4-L5 level using various FE models, highlighting the significant disparities in the reported values. One potential solution to gain reliable insights from a specific model is to employ statistical techniques, such as factorial and probabilistic designs, to assess the model's sensitivity and resilience to variations in input parameters and their interactions [74]. FE models are commonly used to calculate and analyze IDP, facet joint forces, ligament forces, tissue stresses, and related intradiscal forces with spatial resolution determined by the model geometry, mesh density, element formulation, material properties, and boundary/loading conditions [25]. In principle, compressive force within the disc can be approximated from FE-derived IDP using Equation (1); however, this conversion depends on the effective disc cross-sectional area and on assumptions that are most applicable to healthy or mildly degenerated discs. [Table 1] therefore presents representative intradiscal compressive forces inferred from IDP values together with the corresponding facet forces [69,79]. Because disc cross-sectional area varies across spinal levels and between individuals, these values should be interpreted as approximate, level-specific estimates rather than fixed conversions [80].Fig. 3Comparison of the L4-L5 IDP values calculated by different FE spine models with invasive in vivo IDP measurements reported by Wilke et al. [55]. In these FE studies, IDP is one component of the predicted internal stress state and is shown here as a common benchmark to relate model-based estimates of disc loading to the canonical in vivo reference data.Table 1Representative lumbar IDP values (MPa) and the corresponding approximate intradiscal compressive forces and facet forces (N) reported or inferred for each lumbar level during flexion, extension, side bending, and axial rotation [69,79]. The intradiscal forces derived from IDP should be interpreted cautiously, as the conversion depends on level- and subject-specific disc area as well as assumptions that are most applicable to healthy or mildly degenerated discs.Table 1MovementLevelL1-L2L2-L3L3-L4L4-L5FlexionIDP1.6–1.81.35–1.701.30–1.551.20–1.50Estimated force1480–16601460–18401450–17301310–1640ExtensionIDP0.50–0.620.45–0.570.45–0.500.40–0.45Estimated force460–570490–620500–560440–490Facet force4–9030–8715–8645–110SidebendingIDP0.77–0.830.67–0.730.57–0.650.45–0.60Estimated force710–770730–790640–730490–660Facet force4–2411–3010–3130–38Axial rotationIDP0.70–10.64–0.940.61–0.890.60–0.80Estimated force650–920690–1020680–990660–880Facet force55–7560–6858–6353–65 Comparison of the L4-L5 IDP values calculated by different FE spine models with invasive in vivo IDP measurements reported by Wilke et al. [55]. In these FE studies, IDP is one component of the predicted internal stress state and is shown here as a common benchmark to relate model-based estimates of disc loading to the canonical in vivo reference data. Table 1Representative lumbar IDP values (MPa) and the corresponding approximate intradiscal compressive forces and facet forces (N) reported or inferred for each lumbar level during flexion, extension, side bending, and axial rotation [69,79]. The intradiscal forces derived from IDP should be interpreted cautiously, as the conversion depends on level- and subject-specific disc area as well as assumptions that are most applicable to healthy or mildly degenerated discs.Table 1MovementLevelL1-L2L2-L3L3-L4L4-L5FlexionIDP1.6–1.81.35–1.701.30–1.551.20–1.50Estimated force1480–16601460–18401450–17301310–1640ExtensionIDP0.50–0.620.45–0.570.45–0.500.40–0.45Estimated force460–570490–620500–560440–490Facet force4–9030–8715–8645–110SidebendingIDP0.77–0.830.67–0.730.57–0.650.45–0.60Estimated force710–770730–790640–730490–660Facet force4–2411–3010–3130–38Axial rotationIDP0.70–10.64–0.940.61–0.890.60–0.80Estimated force650–920690–1020680–990660–880Facet force55–7560–6858–6353–65 Representative lumbar IDP values (MPa) and the corresponding approximate intradiscal compressive forces and facet forces (N) reported or inferred for each lumbar level during flexion, extension, side bending, and axial rotation [69,79]. The intradiscal forces derived from IDP should be interpreted cautiously, as the conversion depends on level- and subject-specific disc area as well as assumptions that are most applicable to healthy or mildly degenerated discs. The integration of MS and FE models leverages the unique strengths of both approaches and enables simultaneous evaluation of spinal kinematics, kinetics, tissue stresses, and deformations. Active hybrid MS–FE models inherently allow combined consideration of muscle forces and passive tissue responses in shaping lumbar loading patterns [81]. Numerous co-simulation models of the spine typically employ a one-way coupling approach, where data are transferred sequentially between simulations [82]. When compared with in vivo benchmarks, coupled model predictions for IDP and centers of rotation under various tasks have shown close agreement, supporting their use for approximating internal spinal loading [68,70]. These coupled models present significant potential for enhancing surgical and treatment planning, mitigating the risk of MS injuries, and enabling subject-specific simulations [82]. For instance, MS spine models can benefit from incorporating joint stiffness parameters derived from FE simulations of IVDs, while FE models of the spine or its components can be refined by applying muscle or ligament forces and moments from MS as boundary conditions [70]. Traditional MS models simplify intervertebral joints, whereas passive FE models neglect musculature, leading to incomplete representations of in vivo spinal mechanics. Hybrid models that import MS-estimated muscle forces into passive FE models improve the realism of predicted IDP, ligament forces, and centers of rotation, and highlight the limitations of purely in vitro loading conditions [71]. Across the methods discussed above, a consistent order-of-magnitude picture emerges for lumbar loading: representative studies report L4–L5 compressive forces of ∼1–2 kN during relaxed standing and low-demand activities, ∼3–5 kN during common symmetric lifting tasks with moderate trunk flexion and 15–30 kg hand loads, and higher values in asymmetric, high-flexion, or fatiguing conditions [3,47,57,68,70,83]. Shear forces at the same level are generally reported in the range of ∼100–400 N for typical activities and can approach ∼800–1000 N in deep flexion or with large horizontal load distances [47,60,83] [Table 2].Table 2Spinal loads during certain physical activities estimated by MS models.Table 2CategorySpinal LevelActivityLoad rangeCitationCompressionL4–L5Standing/walking∼1–2 kN[3,47,57,68,70,83]CompressionL4–L5Symmetric lifting∼3–5 kN[3,47,57,68,70,83]ShearL4–L5Typical activities∼100–400 N[47,60,83]ShearL4–L5Deep flexion/Holding weights far away from the body∼500–800 N[47,60,83] Table 2Spinal loads during certain physical activities estimated by MS models.Table 2CategorySpinal LevelActivityLoad rangeCitationCompressionL4–L5Standing/walking∼1–2 kN[3,47,57,68,70,83]CompressionL4–L5Symmetric lifting∼3–5 kN[3,47,57,68,70,83]ShearL4–L5Typical activities∼100–400 N[47,60,83]ShearL4–L5Deep flexion/Holding weights far away from the body∼500–800 N[47,60,83] Spinal loads during certain physical activities estimated by MS models. Telemeterized implants measuring instrumented vertebral body replacement or fixation loads report activity-dependent peak forces in broadly similar ranges (on the order of several hundred to a few thousand newtons during everyday tasks) [34,35,65], providing an important in vivo benchmark. FE and hybrid MS–FE models complement these estimates by resolving how these externally applied loads are distributed across discs, facets, and ligaments, and by linking segmental forces to local stresses and IDP. Together, these approaches frame spinal loads not as a single quantity but as a set of related metrics (compression, shear, moment, and internal stress) that must be interpreted jointly when assessing mechanical risk and planning interventions. Emerging technologies are reshaping spinal load assessment by offering practical, non-invasive, and potentially real-time alternatives to traditional methods. Wearable sensors, such as inertial measurement units (IMUs), pressure insoles, and surface electromyography (sEMG), are increasingly used to monitor trunk kinematics, ground reaction forces, and muscle activity during occupational and clinical tasks [84]. These tools enable continuous field-based assessments that are more representative of everyday activities than laboratory-based motion capture systems. However, their accuracy depends on sensor placement, calibration, soft-tissue artifacts, signal drift, and whether the tested movement pattern resembles the conditions under which the method was developed. AI techniques, in particular machine learning and deep neural networks, have also become promising approaches for predicting spinal loads. Their main advantage is speed and once trained, they can rapidly estimate posture, muscle activation, internal forces, or spinal-load-related quantities from non-invasive features such as posture, hand-load position, anthropometrics, EMG, or wearable-sensor signals [[85], [86], [87]]. For example, Aghazadeh et al. [88] demonstrated that ANN models could predict whole-body posture, segmental orientations, and lumbosacral moments during load-handling activities, with derived L4–L5 IDP values showing agreement with in vivo measurements. Hosseini et al. [89] extended this concept to dynamic lifting and showed that ANN-predicted postures produced spinal load estimates in agreement with motion-capture-based MS models. These posture-prediction models are useful when full laboratory motion capture is unavailable, but their precision depends on the similarity between the training data and the target task. They may be less reliable for unusual movement strategies, pathological gait, assistive-device use, prosthetic limbs, extreme postures, or tasks outside the training distribution. In addition, AI can support upstream motion prediction and neuromuscular estimation. Biomechanically informed recurrent neural networks such as Bio-LSTM have been used to predict gait and 3D body-pose dynamics with high temporal consistency, showing that learned motion patterns can improve biomechanically plausible kinematic inputs for downstream load estimation [90]. Nevertheless, gait-oriented networks may not transfer directly to lifting, asymmetric trunk motion, or patient populations unless retrained or externally validated for those conditions. Deep-learning models have also been trained to predict musculoskeletal forces directly from kinematic, kinetic, and EMG data, providing fast approximations of internal force magnitudes with good agreement to conventional MS modeling and enabling near-real-time implementation [91]. However, such models inherit the assumptions and possible errors of the MS models used as labels and may lose accuracy when applied to movements, anthropometries, or clinical populations under-represented in the training dataset. Similarly, InverseMuscleNET showed that recurrent neural networks can estimate muscle activation patterns as an alternative to static optimization and inverse muscle modeling [92]. This approach can reduce computational cost and avoid iterative optimization, but its accuracy is limited by EMG quality, task specificity, electrode placement, and the limited observability of deep muscles. Overall, AI-based methods should be viewed as fast surrogate or augmentation tools rather than replacements for validated biomechanical models. They are most appropriate when rapid estimation, field deployment, or large-scale screening is needed, whereas conventional MS/FE models remain preferable when mechanistic interpretation, tissue-level stresses, implant assessment, or extrapolation to new surgical scenarios is required. The main limitations of current AI approaches are dependence on training-data diversity, limited external validation, reduced interpretability, uncertainty in extrapolation beyond trained conditions, and scarcity of direct in vivo spinal-load labels. Therefore, future work should focus on larger and more diverse datasets, explainable AI techniques, explicit uncertainty quantification, external validation in pathological and non-standard populations, and hybrid AI–biomechanical frameworks that preserve anatomical and physical constraints while enabling rapid subject-specific estimation. To provide a clearer overview of the methodological approaches discussed above, [Table 3] summarizes the main outputs, advantages, limitations, and most appropriate applications of the principal approaches used for lumbar spinal load estimation.Table 3Summary of methodological approaches for estimating lumbar spinal loads.Table 3Methodological approachMain outputsAdvantagesLimitationsMost appropriate applicationsIDP measurementLocal intradiscal pressure; estimated compressive load when converted using disc area and correction factorsDirect internal measurement; historically important in vivo benchmark; useful for validating models in relatively intact discsInvasive; small cohorts; sensitive to sensor artifacts and disc condition; does not capture total spinal load or load sharing across facets/ligamentsBenchmarking compressive disc loading; validating MS/FE modelsTelemeterized implantsImplant-transmitted forces and momentsMost direct in vivo measurement of spinal loading; captures task- and posture-dependent forcesLimited to postoperative/surgical cohorts; measures only the load path through the implant; device-specific calibration and altered load sharingModel validation; postoperative activity/load assessmentMS modelsMuscle forces, joint reaction forces, compression, shear, net momentsNon-invasive; task-specific; suitable for ergonomics and movement analysis; can evaluate many postures/tasksDepends on motion/force data quality, muscle recruitment assumptions, scaling, and redundancy resolutionErgonomic assessment; lifting-task analysis; comparison of movement strategiesEMG-driven/EMG-assisted modelsSubject-specific muscle activation, muscle forces, spinal compression/shearIncorporates measured neuromuscular activity; better captures co-contraction than purely optimization-based modelsSensitive to EMG quality, electrode placement, normalization, and inability to measure deep muscles reliablySubject-specific analysis; rehabilitation; studies of muscle recruitment and co-contractionFE modelsTissue stresses/strains, IDP, facet forces, ligament forces, local load sharingHigh spatial resolution; evaluates internal tissue mechanics; useful for implant/surgical 'what-if' scenariosSensitive to geometry, mesh, material laws, boundary conditions, and validation data; computationally demandingImplant design; surgical planning; tissue-level stress/load-sharing analysisHybrid MS-FE modelsMS-derived loads combined with FE tissue-level outputsCombines realistic muscle/joint loading with local tissue mechanics; improves physiological relevanceComplex workflow; high computational cost; limited validation datasetsSubject-specific simulations; detailed surgical/implant biomechanicsDisplacement-driven/control-based modelsLoads inferred from measured kinematics or neuromuscular control assumptionsAvoids explicit optimization of muscle forces in some implementations; useful when kinematics are availableSensitive to imaging/kinematic errors; may not directly capture true muscle recruitmentImaging-based load estimation; studies of posture-control assumptionsReduced-order/regression modelsApproximate compression, shear, or moments from task variablesFast, simple, and practical for screening; useful when full MS/FE simulations are not feasibleValid only within trained task ranges and populations; limited extrapolation; simplified representation of co-contraction and 3D motionWorkplace screening; rapid ergonomic assessmentAI/ML-based surrogatesPredicted posture, activation, forces, moments, or spinal-load estimatesFast after training; compatible with wearable/field data; potential for real-time estimationRequires large and diverse datasets; limited interpretability; weak extrapolation outside training domain; often depends on MS/FE labels rather than direct in vivo loadsReal-time monitoring; field-based assessment; large-scale screening; AI-biomechanical hybrid workflows* Abbreviations: MS: musculoskeletal; FE: finite element; IDP: intradiscal pressure; EMG: electromyography; ML: machine learning. Table 3Summary of methodological approaches for estimating lumbar spinal loads.Table 3Methodological approachMain outputsAdvantagesLimitationsMost appropriate applicationsIDP measurementLocal intradiscal pressure; estimated compressive load when converted using disc area and correction factorsDirect internal measurement; historically important in vivo benchmark; useful for validating models in relatively intact discsInvasive; small cohorts; sensitive to sensor artifacts and disc condition; does not capture total spinal load or load sharing across facets/ligamentsBenchmarking compressive disc loading; validating MS/FE modelsTelemeterized implantsImplant-transmitted forces and momentsMost direct in vivo measurement of spinal loading; captures task- and posture-dependent forcesLimited to postoperative/surgical cohorts; measures only the load path through the implant; device-specific calibration and altered load sharingModel validation; postoperative activity/load assessmentMS modelsMuscle forces, joint reaction forces, compression, shear, net momentsNon-invasive; task-specific; suitable for ergonomics and movement analysis; can evaluate many postures/tasksDepends on motion/force data quality, muscle recruitment assumptions, scaling, and redundancy resolutionErgonomic assessment; lifting-task analysis; comparison of movement strategiesEMG-driven/EMG-assisted modelsSubject-specific muscle activation, muscle forces, spinal compression/shearIncorporates measured neuromuscular activity; better captures co-contraction than purely optimization-based modelsSensitive to EMG quality, electrode placement, normalization, and inability to measure deep muscles reliablySubject-specific analysis; rehabilitation; studies of muscle recruitment and co-contractionFE modelsTissue stresses/strains, IDP, facet forces, ligament forces, local load sharingHigh spatial resolution; evaluates internal tissue mechanics; useful for implant/surgical 'what-if' scenariosSensitive to geometry, mesh, material laws, boundary conditions, and validation data; computationally demandingImplant design; surgical planning; tissue-level stress/load-sharing analysisHybrid MS-FE modelsMS-derived loads combined with FE tissue-level outputsCombines realistic muscle/joint loading with local tissue mechanics; improves physiological relevanceComplex workflow; high computational cost; limited validation datasetsSubject-specific simulations; detailed surgical/implant biomechanicsDisplacement-driven/control-based modelsLoads inferred from measured kinematics or neuromuscular control assumptionsAvoids explicit optimization of muscle forces in some implementations; useful when kinematics are availableSensitive to imaging/kinematic errors; may not directly capture true muscle recruitmentImaging-based load estimation; studies of posture-control assumptionsReduced-order/regression modelsApproximate compression, shear, or moments from task variablesFast, simple, and practical for screening; useful when full MS/FE simulations are not feasibleValid only within trained task ranges and populations; limited extrapolation; simplified representation of co-contraction and 3D motionWorkplace screening; rapid ergonomic assessmentAI/ML-based surrogatesPredicted posture, activation, forces, moments, or spinal-load estimatesFast after training; compatible with wearable/field data; potential for real-time estimationRequires large and diverse datasets; limited interpretability; weak extrapolation outside training domain; often depends on MS/FE labels rather than direct in vivo loadsReal-time monitoring; field-based assessment; large-scale screening; AI-biomechanical hybrid workflows* Abbreviations: MS: musculoskeletal; FE: finite element; IDP: intradiscal pressure; EMG: electromyography; ML: machine learning. Summary of methodological approaches for estimating lumbar spinal loads. * Abbreviations: MS: musculoskeletal; FE: finite element; IDP: intradiscal pressure; EMG: electromyography; ML: machine learning. From a methodological standpoint, each approach occupies a distinct niche in spinal load estimation. Invasive telemetry provides gold-standard, activity-dependent segmental forces, but only in small surgical cohorts. MS models are most appropriate when motion and/or EMG data are available and comparative evaluation of multiple tasks or interventions is required, whereas FE models are indicated when tissue-level stress distributions, implant design, or local load sharing between disc, facets, and ligaments are of primary interest. Displacement-driven and reduced-order or AI-based surrogates are particularly useful when only limited kinematic or task descriptors are available, or when near real-time feedback is needed in field or clinical settings. In practice, hybrid pipelines that combine these methods often yield the most robust and interpretable estimates of spinal loading across ergonomic, rehabilitative, and surgical applications. Limitations, clinical implication and future directions One common trigger of LBP is altered spinal mechanics, meaning that changes in spinal loading can influence symptom development and persistence. Understanding how aging affects spinal loading transmitted across the spinal column is crucial, especially considering that the elderly population experiences the most cases of LBP. Unfortunately, using techniques from previous publications to measure IDP may not yield reliable results for assessing spinal loading due to the degeneration of most IVDs with age. While several studies have measured IDP in degenerated discs, these measurements do not accurately reflect the spinal load transferred to the neighboring vertebrae because the nature of pressure within a degenerated nucleus is not hydrostatic [93,94]. Furthermore, only a limited number of studies have investigated the roles of the posterior elements in the load-bearing capacity of the spinal column concerning age, the level of disc degeneration, or patients with low back pain. Therefore, our understanding of how degeneration alters spinal load transmission, and the load-sharing environment remains very limited. Although universal pain thresholds for spinal loads or degeneration grades are not established, current methods can inform several clinical and occupational decisions when used together. IDP provides task-specific compressive benchmarks in relatively intact discs, assisting with activity counseling and postoperative comparisons, but it should not be over-interpreted in degeneration. Telemeterized implants offer the most direct in-vivo loads along the implant path. They can personalize restrictions (e.g., early lifting limits), help evaluate technique/posture effects, and calibrate model inputs. MS models estimate level-specific compression/shear and muscle demands during real tasks, supporting ergonomic redesign (load height/reach, tool placement), return-to-work planning, and prehabilitation (targeting co-contraction patterns or trunk endurance). FE models can resolve tissue-level stresses and facet/contact forces under “what-if” scenarios (cage size/position, rod construct, osteotomy), helping anticipate adjacent-segment risks and guide implant choices. Hybrid MS–FE and AI-assisted pipelines can deliver near-real-time, patient-specific guidance (e.g., lifting technique coaching, home-based monitoring) when paired with robust validation and uncertainty reporting. Practically, we recommend reporting task-mapped load profiles (compression/shear/facet forces, with uncertainty bands) and focusing on within-patient trends across activities or time rather than hard thresholds, enabling individualized decision-making in surgery, rehabilitation, and workplace design. EMG-driven and FE models are promising tools for estimating spinal loads and predicting spinal pathologies. However, some obstacles limit their clinical applications currently. Most importantly, FE predictions depend on patient-specific geometry, constitutive properties, and boundary/loading conditions, and should be calibrated and validated against comparable benchmarks. At present, high-quality in vivo reference data and reliable material parameters, particularly for degenerated discs, remain sparse, constraining model fidelity and generalizability. Accordingly, future work should prioritize structured personalization workflows that (1) derive subject-specific parameters from measurable inputs (e.g., imaging and motion/EMG), (2) perform model calibration to subject-specific targets, (3) conduct task-level external validation on held-out activities, and (4) report uncertainty and sensitivity analyses alongside standardized documentation of geometry sources, constitutive laws, and boundary conditions. AI-based models hold substantial promise for rapid, noninvasive spinal load estimation in occupational risk assessment, rehabilitation, and clinical decision-making. To be clinically credible, next-generation AI should emphasize diverse, well-annotated datasets, pre-registered external validation, explainable/traceable outputs, and uncertainty quantification, and, where feasible, integrate with mechanistic pipelines (hybrid AI–biomechanical models) to appropriately account for the underlying physics and anatomy. Progress along these lines will enable reproducible, patient-specific assessments that are actionable for ergonomic design, prehabilitation, and surgical planning. Declaration of competing interest The authors declare no conflicts of interest.