MuscleLens

Abstract

Parkinson’s disease (PD) gait assessment remains dominated by subjective clinical rating, while the instrumented gait laboratories that could substitute for it are inaccessible to most patients.

We present MuscleLens, a pipeline that lifts monocular video into a phase-locked \(30 \times 80\) muscle activation code on a unified musculoskeletal model. The stack combines SMPL recovery, GMR retargeting to MyoFullBody, and a frozen MuscleMimic policy that emits 80-dimensional muscle activations at 100 Hz.

Evaluated on roughly 4 800 subject-level samples from CARE-PD, augmenting SMPL-H kinematics with muscle features (i) raises 5-fold PD balanced accuracy from \(0.807\) to \(0.814\) at matched AUC (\(0.937\)); (ii) consistently lowers cohort-normalised UPDRS MAE across PCA dimensions, reaching \(0.344\) versus \(0.349\) at 128 dimensions; and (iii) improves leave-one-cohort-out balanced accuracy on six of seven cohorts.

Beyond accuracy, the actuator-space code exposes co-activation patterns that joint kinematics alone cannot encode, providing a biomechanically interpretable view of PD gait. The monocular branch demonstrates the end-to-end deployment route: about 1 minute from a 15-second clip to a complete muscle code on a single NVIDIA RTX 5090, bringing muscle-level analysis within reach of routine clinical-style acquisition.

Key supervised results

5-fold logistic regression on 4 669 PD-labelled (resp.\ 2 559 UPDRS-labelled) subject samples; mean ± std across folds. The last column is the cohort-normalised PCA setting at 128 dimensions — the strongest evidence that muscle features contribute complementary signal beyond joint kinematics, with the improvement holding consistently across 8 / 16 / 32 / 64 / 128 PCA dimensions.

Leave-one-cohort-out transfer

Fusion improves balanced accuracy on 6 of 7 cohorts (mean Δ = +0.020). Pooled LOCO PD AUC looks alarmingly low for both modalities (kinematics 0.282, fusion 0.340) but is a between-cohort probability-calibration artifact — four of seven cohorts are 100 % PD, so per-cohort AUC is undefined and the pool is driven by cross-cohort calibration shift rather than ranking failure inside any single cohort. BAcc, which is invariant to that shift, is the meaningful summary on this corpus.

Fusion embedding overview

Video → SMPL → Muscle visualisations

A single ``walk-forward’’ clip carried through the full MuscleLens pipeline. The three stages below correspond left-to-right to the blocks in Fig. 1 of the paper.

Compute & runtime

The full inference stack runs on a single NVIDIA RTX 5090. The MuscleMimic policy emits activations at roughly 12 s per SMPL clip; on the video branch, end-to-end processing of a 15-second clip (GVHMR → SMPL → GMR → MuscleMimic) completes in about 1 minute, returning both kinematics and muscle activations in a single pass. The headline number is the end-to-end budget a downstream user would experience — placing MuscleLens within reach of routine clinical-style acquisition without specialised gait laboratories.

Conclusion & limitations

MuscleLens establishes a shared route from monocular video or SMPL-H motion to phase-locked muscle activation codes, and provides initial evidence that the resulting actuator-space representation carries information complementary to joint kinematics: five-fold PD balanced accuracy improves, cohort-normalised UPDRS regression improves at every PCA dimension tested, and LOCO balanced accuracy improves on six of seven cohorts. The activation code also surfaces co-activation structure that joint kinematics alone cannot encode — a step toward more interpretable PD gait analysis.

Three limitations bound the present claim:

1. Monocular branch is a deployment demonstration, not yet a clinically validated front end. Large-scale patient-video evaluation remains future work; the qualitative video → SMPL → muscle clips above document the current state.
2. The musculoskeletal model and tracking policy are not calibrated to older adults with PD-specific movement strategies, which likely caps co-activation fidelity.
3. The residual cohort-transfer gap on pooled LOCO PD AUC reflects a between-cohort probability-calibration problem more than a within-cohort ranking failure. Domain-invariant training and cohort-calibrated scoring are natural next steps and the most direct route to strengthening the present claim.

Citation

@inproceedings{musclelens2026,
  title     = {MuscleLens: A Shared Muscle-Space Pipeline for Parkinsonian Gait Analysis},
  author    = {Anonymous Author(s)},
  booktitle = {ICRA 2026 Workshop on Enabling Autonomy and Independence in
               Aging Societies through Advanced Robotics and AI},
  year      = {2026}
}

Acknowledgments

Acknowledgments — TBD.