Source-linked AI summary
Real-Time Musculoskeletal Surrogates for Pediatric Cerebral Palsy: a Credibility Pilot
Mohammad Arif Ul Alam
TL;DR
Pediatric CP rehabilitation needs credible real-time MSK surrogates that generalize across children and quantify uncertainty. The paper develops a subject-conditioned causal neural surrogate with OpenSim inputs and evaluates it using leakage-free subject-held-out validation. MT-length prediction is accurate and fast, whereas small-cohort force estimation and input-only uncertainty propagation remain unresolved.
Problem
Existing neural-surrogate evidence is limited across heterogeneous pediatric CP gait, while conventional MSK simulations are too computationally expensive for interactive rehabilitation and muscle force is an underdetermined model estimate.
Method
A subject-conditioned causal surrogate combines temporal joint kinematics, OpenSim-derived static parameters and true muscle capacities, training-only perturbations, and leave-one-subject-out plus locked-test evaluation.
Results
MT-length R2 was 0.9245 in development LOSO validation and 0.9477 on locked subjects, with 7.69% locked-test nRMSE; nominal 90% intervals covered only 4.35% of force and 0.70% of MT-length observations.
Takeaways & Limitations
The pipeline provides a credible preliminary foundation for real-time pediatric MSK surrogate evaluation while identifying force modeling and epistemic uncertainty as central next challenges.
Takeaways & Limitations
The small, heterogeneous cohort makes direct raw muscle-force regression insufficient, and CMC force and activation labels are physics-based model estimates rather than in-vivo measurements.
Abstract
from arXiv · showhide
Real-time musculoskeletal (MSK) surrogates could support personalized rehabilitation for children with cerebral palsy (CP), but their credibility depends on subject-wise evaluation, low inference latency, and calibrated uncertainty. We develop a subject-conditioned causal neural surrogate using OpenSim-derived static parameters, temporal joint kinematics, true muscle capacities, and training-only perturbations. On a real pediatric CP gait dataset comprising nine children, we use leave-one-subject-out validation on six development subjects and evaluate a frozen configuration once on three locked test subjects. The surrogate accurately reproduces musculotendon lengths (R-square = 0.92 in development validation and approximately 0.95 on locked subjects; nRMSE < 8%) while requiring only sub-millisecond to few-millisecond neural inference, well below a 100 ms interactive-rehabilitation target. In contrast, direct muscle-force estimation remains unstable at this small, heterogeneous scale: pooled metrics can overstate within-subject, per-muscle accuracy. A Monte Carlo credibility pilot further shows that propagating only +/-5% anthropometry and muscle-capacity variation produces severely overconfident nominal 90% intervals (approximately 4% force coverage and below 1% MT-length coverage). These results establish a leakage-free evaluation and credibility framework for pediatric MSK surrogates, while identifying force modeling and epistemic uncertainty as the central next challenges for clinically credible digital twins.
1. Introduction
Pediatric CP gait analysis requires patient-specific, computationally efficient MSK surrogates because movement impairments are heterogeneous and conventional simulations are too slow for interactive use. The study therefore emphasizes subject-held-out evaluation, transparent partitioning, and uncertainty assessment as prerequisites for credible clinical digital twins.
- CP causes heterogeneous movement impairments, making patient-specific gait analysis important for clinical decision-making.
- Conventional muscle-driven MSK simulations estimate otherwise unmeasurable quantities but remain too computationally expensive for interactive rehabilitation or continuously updated digital twins.
- Neural surrogates provide a practical route to real-time MSK analysis, but existing evidence often comes from healthy adults, few muscles, or task-specific datasets.
- Pediatric CP is especially challenging because gait patterns, body dimensions, strength, and neuromuscular control vary substantially between children.
- Credible evaluation requires leakage-free partitioning, validation on independent subjects, and explicit uncertainty assessment rather than point predictions alone.
2. Method
The method combines a causal temporal surrogate with subject-specific OpenSim parameters and a leakage-free subject-level evaluation protocol. Training-only perturbations and physics-informed multi-task supervision are used while keeping validation and locked-test data separated from model selection.
- Subject-Conditioned Causal Surrogate: Approximately 0.9 million trainable parameters define a causal temporal convolutional network paired with a static-parameter multilayer perceptron.
- Subject-Conditioned Causal Surrogate: A 251-frame (≈250 ms) history of 23 joint coordinates and speeds is fused with subject-specific OpenSim parameters.
- Subject-Conditioned Causal Surrogate: The fused model predicts 92 muscle forces, activations, and musculotendon lengths, plus moment arms for 23 independent coordinates.
- Supervision and Physics-Informed Objective: The multi-task objective uses normalized mean-squared-error terms for force, activation, MT length, moment arm, and joint moment.
- Supervision and Physics-Informed Objective: Predicted forces are coupled to joint-level mechanics through a moment-arm relation, but the torque term is evaluated only on available lower-limb coordinates.
- Robustness and Evaluation: Synthetic Fmax and anthropometry perturbations, kinematic noise, and channel occlusion were applied exclusively to training windows.
- Leakage-Free Evaluation Protocol: Six development subjects used leave-one-subject-out validation, while three additional children formed a locked test set evaluated once after architecture and training duration were frozen.
- Leakage-Free Evaluation Protocol: Within-subject and mean-per-muscle R2 were primary measures because frame-pooled/global R2 can be inflated by between-subject force-scale differences.
3. Results
The causal subject-conditioned surrogate accurately reproduced musculotendon lengths and met the real-time latency target, while force and torque predictions remained unreliable and uncertainty intervals were severely overconfident.
- 3.1. Musculotendon Geometry Surrogate: R2 = 0.9245 in development LOSO validation and R2 = 0.9477 on locked subjects for MT-length prediction.Locked-test MT-length nRMSE was 7.69% overall, with 6.51%, 6.38%, and 9.51% for mild, moderate, and severe subjects.
- 3.2. Real-Time Inference: 7.41 ms on a single-thread CPU and 3.16 ms on an NVIDIA A100 GPU for one causal input window, below the 100 ms criterion.This establishes real-time feasibility for the neural surrogate component, without claiming a measured speedup over OpenSim.
- 3.3. Direct Muscle-Force Estimation Remains Difficult: 0.2165 pooled force R2 and 34.55% nRMSE across development LOSO folds, with subject-level force R2 ranging from −0.635 to 0.486.On locked subjects, pooled force R2 was 0.6056, while subject-wise values ranged from −1.052 to 0.733, showing pronounced heterogeneity.
- 3.3. Direct Muscle-Force Estimation Remains Difficult: Pooled force metrics can be inflated by between-subject force-scale differences, so subject-wise results should be primary in this heterogeneous cohort.The study recommends normalized force, f/Fmax, as a future scale-robust secondary outcome, but did not compute it here.
- 3.4. Torque-Consistency Diagnostic: −0.0055 pooled joint-moment R2 in development LOSO, with subject-wise values ranging from −0.024 to 0.000.The masked torque-consistency term was empirically inert at the available sample size and label quality.
4. A Credibility (VVUQ) Pilot
The credibility pilot tested whether limited input perturbations produce calibrated uncertainty intervals for the frozen surrogate. Input-only propagation yielded severely overconfident intervals, indicating that additional uncertainty sources require explicit treatment.
- Pilot design: 100 Monte Carlo samples per held-out subject independently perturbed muscle-specific Fmax values and anthropometry-related static inputs by ±5%.Prediction intervals were evaluated over randomly sampled causal windows using empirical coverage and mean interval width.
- Uncertainty calibration: 4.35% empirical coverage for force and 0.70% for MT length fell far below the nominal 90% intervals.These pooled locked-cohort results demonstrate severe overconfidence.
- Uncertainty calibration: 8.68 N mean force interval width and 0.00062 m mean MT-length interval width accompanied the low empirical coverage.The narrow intervals show that the propagated input variation captured only a small component of observed prediction error.
- Uncertainty calibration: 3.94%–4.72% force coverage and 0.37%–0.88% MT-length coverage showed the same overconfidence pattern for every held-out subject.The experiment was a sensitivity and calibration diagnostic rather than a complete credibility assessment.
- Interpretation: Model-form and epistemic uncertainty associated with limited, heterogeneous pediatric training data likely dominates the narrow input variation tested here.The study therefore identifies input-only propagation as insufficient for clinically consequential or regulatory-grade credibility claims.
5. Discussion and Limitations
The discussion frames the pipeline as a rigorous preliminary evaluation substrate, while emphasizing that force modeling and credibility remain constrained by cohort size, label provenance, and incomplete uncertainty treatment.
- Force evaluation: Pooled force metrics should not replace within-subject, per-muscle performance because body size and force scale can inflate frame-pooled agreement.This is especially important when interpreting heterogeneous force results at exploratory scale.
- Discussion: Direct regression of raw muscle force from kinematics remains insufficient at this exploratory scale, despite the pipeline providing a rigorous evaluation substrate.The conclusion is bounded to the preliminary pediatric CP setting described by the study.
- Limitations: Nine children and one lower-limb gait trial per subject constrain modeling of inter-subject heterogeneity and reliable subgroup analyses.The authors identify larger and more diverse pediatric datasets as required next steps.
- Uncertainty: 4.35% force coverage and 0.70% MT-length coverage show that plausible input uncertainty is not equivalent to calibrated predictive uncertainty.Regulatory-grade credibility requires explicit treatment of model-form and epistemic uncertainty.
- Force modeling: CMC-derived force and activation labels are physics-based model estimates rather than in-vivo measurements.Future work is directed toward normalized force, Hill-type muscle dynamics, and sensitivity to alternative modeling assumptions.
6. Conclusion
The study provides a credible preliminary foundation for real-time pediatric MSK surrogate evaluation while identifying muscle-force estimation and input-only uncertainty propagation as unresolved challenges.
- Conclusion: The leakage-free, subject-conditioned pipeline supports reproducible, real-time surrogate evaluation on real pediatric CP gait data.Its focused technical agenda targets in-silico pediatric neurorehabilitation.