MuscleLens:
A Shared Muscle-Space Pipeline for Parkinsonian Gait Analysis

MuscleLens accepts either monocular video or SMPL motion. For video, we recover SMPL-H with GVHMR; for CARE-PD we use the released SMPL-H sequences directly. The motion is then retargeted to the MyoFullBody musculoskeletal model via GMR, and replayed under the frozen mm-fullbody-base MuscleMimic policy, which emits 80-dimensional muscle activations at $100$ Hz. A lightweight coordinate-frame harmonisation precedes retargeting on the CARE-PD branch to keep cross-source processing consistent.

Activations are segmented into gait cycles, resampled to 30 phase bins per cycle, and averaged across valid cycles to form a code $\mathbf{A}\in\mathbb{R}^{30\times 80}$, where $A_{p,m}=\mathbb{E}_k[a_m(t_p^{(k)})]$. We compare a 550-dimensional NMF summary of this code, a 5 310-dimensional kinematic descriptor, and their concatenation.

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.

All quantitative results below use the same SMPL-H input, isolating the effect of muscle augmentation from any upstream reconstruction error. Kinematics remain the strongest single modality; muscle features add a small but consistent gain on top of kinematics, most defensibly once cohort structure and feature dimensionality are controlled. The gain appears in three independent settings: random folds, cohort-normalised PCA, and LOCO balanced accuracy.

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, with the improvement holding across 8 / 16 / 32 / 64 / 128 PCA dimensions.

Modality comparison. (a) Random five-fold PD classification; (b) Random five-fold UPDRS-gait prediction; (c) After cohort-wise z-scoring and equal-dimensional PCA, fusion yields consistently lower UPDRS MAE than kinematics across PCA dimensions.

Fusion improves balanced accuracy on 6 of 7 cohorts (mean $\Delta=+0.020$). Pooled LOCO PD AUC looks alarmingly low for both modalities (kinematics $0.282$, fusion $0.340$) but is a between-cohort probability-calibration artefact: 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, invariant to that shift, is the meaningful summary on this corpus.

(a) HDBSCAN finds one dominant cluster and a small side cluster. (b) Cohort labels still retain source structure. (c) UPDRS-gait labels show a weak ordering rather than clean separation.

The fused unsupervised embedding shows one dominant cluster and a smaller side cluster, with severity labels forming only a weak gradient. Weighted UPDRS purity is $0.472$ for muscle, $0.448$ for kinematics, and $0.447$ for fusion; PD purity sits at ≈ 0.96 across all modalities, essentially the labelled PD base rate of 95.1 %. The unsupervised view is therefore not diagnostic on its own; the supervised analyses above remain where the muscle-space contribution is most credible.

MuscleLens establishes a shared route from monocular video or SMPL-H motion to phase-locked muscle activation codes, and provides initial evidence that the 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. The 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 yet 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.