A Non-Invasive 3D Gait Analysis Framework for Quantifying Psychomotor Retardation in Major Depressive Disorder

Predicting the status of Major Depressive Disorder (MDD) from objective, non-invasive methods is an active research field. Yet, extracting automatically objective, interpretable features for a detailed analysis of the patient state remains largely unexplored. Among MDD’s symptoms, Psychomotor retardation (PMR) is a core item, yet its clinical assessment remains largely subjective. While 3D motion capture offers an objective alternative, its reliance on specialized hardware often precludes routine clinical use. In this paper, we propose a non-invasive computational framework that transforms monocular RGB video into clinically relevant 3D gait kinematics. Our pipeline uses Gravity-View Coordinates along with a novel trajectory-correction algorithm that leverages the closed-loop topology of our adapted Timed Up and Go (TUG) protocol to mitigate monocular depth errors. This novel pipeline enables the extraction of 297 explicit gait biomechanical biomarkers from a single camera capture. To address the challenges of small clinical datasets, we introduce a stability-based machine learning framework that identifies robust motor signatures while preventing overfitting. Validated on the CALYPSO dataset, our method achieves an 83.3% accuracy in detecting PMR and explains 64% of the variance in overall depression severity (R^2=0.64). Notably, our study reveals a strong link between reduced ankle propulsion and restricted pelvic mobility to the depressive motor phenotype. These results demonstrate that physical movement serves as a robust proxy for the cognitive state, offering a transparent and scalable tool for the objective monitoring of depression in standard clinical environments.

💡 Research Summary

The paper presents a fully non‑invasive framework for quantifying psychomotor retardation (PMR), a core symptom of major depressive disorder (MDD), using only a single RGB video camera. The authors adapt a Timed Up and Go (TUG) protocol, record patients with a standard smartphone (Samsung Galaxy S21 FE) and process the footage with a state‑of‑the‑art monocular 3D human mesh recovery model called Gravity‑View Human Mesh Recovery (GVHMR). GVHMR outputs SMPL parameters (shape β, pose θ, global orientation ϕ, and translation t) in a gravity‑aligned coordinate system, which already provides a world‑grounded floor plane.

A major challenge with monocular reconstruction is cumulative drift of the global trajectory. The authors exploit the closed‑loop nature of the TUG test (the subject ends where they started) and formulate a loop‑closure optimization that enforces start‑to‑end consistency while preserving local velocity profiles. After drift correction, they align each subject’s walking direction to a common global Z‑axis using principal component analysis on the horizontal trajectory, thereby removing any dependence on the participant’s heading relative to the camera.

From the corrected 3D skeleton sequence (45 joints per frame) the pipeline automatically segments the linear walking phases by detecting low‑velocity turn regions. Within the steady‑state walking segments, 297 explicit biomechanical features are extracted. These include spatio‑temporal metrics (stride length, gait speed, symmetry), joint‑level kinematics (maximum joint angles, angular velocities, ankle propulsion), pelvic and spinal mobility (tilt, rotation), and arm‑swing characteristics. Each feature is chosen because it maps directly onto clinical descriptors of PMR, such as reduced initiation, slowed gait, and postural rigidity.

Because the CALYPSO dataset contains only 42 participants (22 asymptomatic, 20 symptomatic for PMR), the authors introduce a stability‑based feature selection method to avoid over‑fitting. They perform extensive bootstrapping (1,000 resamples), train simple classifiers (logistic regression, random forest) on each resample, and compute feature importance rankings. Features that consistently rank high across resamples are retained, yielding a robust set of motor signatures that are both predictive and interpretable.

Experimental results show that the binary PMR classifier reaches 83.3 % accuracy (AUC ≈ 0.78, F1 ≈ 0.81). For regression of the total Hamilton Depression Rating Scale (HDRS) score, the model explains 64 % of the variance (R² = 0.64, MAE ≈ 2.1 points). The most stable and predictive biomarkers are ankle dorsiflexion propulsion, pelvic anterior‑posterior tilt variability, and gait symmetry indices—findings that align with clinical observations that depressed patients exhibit reduced push‑off power and constrained pelvic motion.

Compared with prior work that relies on multi‑camera setups, depth sensors (e.g., Kinect), or wearable inertial units, this approach requires only a single off‑the‑shelf smartphone, dramatically lowering cost, setup time, and patient burden. Moreover, unlike end‑to‑end deep‑learning classifiers that act as black boxes, the explicit biomechanical markers provide transparent decision support that clinicians can readily interpret.

The authors acknowledge limitations: the modest sample size, the need for external validation on larger, multi‑site cohorts, and potential sensitivity to lighting or background variations not fully explored. Future directions include real‑time processing pipelines for immediate feedback during clinical visits and integration with electronic health records to track longitudinal changes in motor signatures.

In summary, the study demonstrates that high‑quality 3D gait kinematics can be derived from monocular RGB video, that loop‑closure drift correction makes long‑duration recordings reliable, and that a stability‑driven machine‑learning framework can extract robust, clinically meaningful motor biomarkers of psychomotor retardation. This work paves the way for scalable, objective monitoring of depression severity in routine psychiatric practice.