Abstract
Dynamic prediction of locomotor capacity after stroke could enable more individualized rehabilitation, yet current assessments largely provide static impairment scores and do not indicate whether patients can perform specific tasks such as slope walking or stair climbing. Here, we present a wearable-informed data-physics hybrid generative framework that reconstructs a stroke survivor's locomotor control from wearable inertial sensing and predicts task-conditioned post-stroke locomotion in new environments. From a single 20 m level-ground walking trial recorded by five IMUs, the framework personalizes a physics-based digital avatar using a healthy-motion prior and hybrid imitation learning, generating dynamically feasible, patient-specific movements for inclined walking and stair negotiation. Across 11 stroke inpatients, predicted postures reached 82.2% similarity for slopes and 69.9% for stairs, substantially exceeding a physics-only baseline. In a multicentre pilot randomized study (n = 21; 28 days), access to scenario-specific locomotion predictions to support task selection and difficulty titration was associated with larger gains in Fugl-Meyer lower-extremity scores than standard care (mean change 6.0 vs 3.7 points; p < 0.05). These results suggest that wearable-informed generative digital avatars may augment individualized gait rehabilitation planning and provide a pathway toward dynamically personalized post-stroke motor recovery strategies.
Introduction: Beyond Static Clinical Scores
Stroke remains a leading cause of long-term disability, frequently resulting in hemiplegia and significant gait impairments. In current clinical practice, rehabilitation progress is typically measured through "static" metrics—standardized assessments like the Fugl-Meyer Assessment or kinematic snapshots taken during level-ground walking. While these scores provide a baseline of impairment, they lack predictive power. A clinician can observe a patient walking on a flat surface but cannot easily determine if that patient is ready to safely navigate a 1:12 incline or climb a flight of stairs.
This gap often leads to a "trial-and-error" approach in physical therapy. If a task is too easy, the patient’s recovery may plateau; if it is too difficult, the patient faces an increased risk of falls or injury. To address this, we developed a generative framework that creates a "digital twin" of a stroke survivor, capable of forecasting performance in unobserved, challenging environments using minimal wearable sensor data.
A Data-Physics Hybrid Framework
The proposed framework bridges biomechanical physics and deep learning through a three-stage generative process:
- Healthy Motion Atlas (HMA): Using Deep Reinforcement Learning (specifically Proximal Policy Optimization), a physics-based digital avatar is trained within a Unity 3D environment to master normative gait. This stage establishes a "population prior," representing the fundamental mechanics of human movement.
- Atlas-to-Patient Adaptation: The model employs transfer learning to personalize the HMA. By processing data from five Inertial Measurement Units (IMUs) worn during a brief 20-meter walk, the avatar learns the patient's unique "control signature." This includes critical compensatory strategies such as pelvic hiking or circumduction, which are often necessary for the patient to maintain balance.
- Task-Conditioned Prediction: To forecast movement on slopes and stairs, the framework utilizes Generative Adversarial Imitation Learning (GAIL). The avatar explores a physics simulator (PhysX) to find a movement solution for the new task. The model is penalized if the resulting motion deviates too far from the patient’s observed level-ground "style," ensuring the prediction remains patient-specific while respecting physical laws like gravity and joint limits.
Clinical Validation and Results
The framework was validated using a combination of clinical data from 21 post-stroke inpatients and 19 healthy adults, supplemented by the CMU Motion Capture database to ensure a robust representation of gait cycles.
Prediction Accuracy
The personalized avatars demonstrated high fidelity in predicting unobserved tasks. For slope navigation, the predicted postures reached 82.2% similarity to ground truth data, while stair negotiation reached 69.9% similarity. These results significantly outperformed traditional physics-only baselines, which often fail to account for the idiosyncratic compensatory movements of stroke survivors.
Impact on Rehabilitation
In a 28-day multicentre pilot randomized study (n=21), the clinical utility of these predictions was tested. Therapists in the experimental group used the avatar’s predictions to support task selection and difficulty titration. Patients in this "model-assisted" group showed significantly higher gains in Fugl-Meyer lower-extremity scores—an average increase of 6.0 points compared to 3.7 points in the standard care control group (p < 0.05).
Discussion and Future Directions
A key insight of this work is that "abnormal" gait features are often functional necessities for stroke survivors. By using physics as a regularizer, the generative model ensures that predicted movements are dynamically feasible without trying to "correct" the patient's necessary compensations. Furthermore, the ability to generate these complex predictions from a single 20-meter walk makes the technology highly scalable for clinical environments.
However, limitations remain. The current study focused on patients capable of independent walking. Future iterations must account for assistive devices like canes or walkers, which require modeling complex human-device interactions. Additionally, expanding the predictive scope to include turning and obstacle avoidance will be essential for full real-world application.
Ultimately, this framework represents a shift toward the "Digital Twin" in rehabilitative medicine—a virtual environment where clinicians can safely test the limits of recovery and tailor interventions to the specific biomechanical reality of each patient.
Summary updated