A multimodal machine learning approach to forecast upper limb motor recovery after stroke using kinematic and electromyographic data

Background: Forecasting post-stroke rehabilitation outcome is essential to personalize therapeutic strategies. Traditional approaches rely on clinical assessment scales, which, while essential, may benefit from complementary objective measures. In this direction, robot-assisted assessment offers the unprecedented possibility of precisely collecting patients’ physiological signals, enabling a data-driven approach to recovery assessments. Leveraging the use of multimodal features collected during rehabilitation, we developed a machine learning approach to forecast the upper limb (UL) motor recovery in stroke survivors after rehabilitation. Methods: This study evaluated a 4-week rehabilitation program, using both standard physical therapy and the ALEx robot to promote UL motor recovery in 11 subacute stroke survivors, compared to 6 healthy individuals. Kinematic measures (Kin) and surface electromyography (sEMG) were collected during a 3D reaching task involving six target points. From these tasks, 19 kinematic and 76 sEMG features were extracted. To forecast the UL motor recovery post intervention, a two-step machine learning approach was devised: a machine learning regression model was developed and validated to predict the Fugl-Meyer Assessment for UL (FMA-UL) post rehabilitation, whereas an anomaly detection algorithm identified patients who exhibited limited or no motor recovery post intervention. The anomaly detection approach used a fully-connected autoencoder that identified patients with reduced recovery likelihood. The regression models, optimized via a nested Leave-One-Subject-Out approach, guided feature selection and refined hyperparameters to predict FMA-UL scores post intervention. Results: The optimized regression model achieved an RMSE in predicting the FMA-UL post intervention of 5.45. The autoencoder effectively identified patients with reduced recovery potential, showing a higher distribution of reconstruction errors for these individuals. Conclusions: The findings confirm that combining kinematic and sEMG data improves motor recovery assessment. The proposed machine learning approach holds potential for aiding clinicians and therapists in identifying patients who are more likely to recover UL motor functions before rehabilitation begins. By accurately predicting recovery outcomes, this method can help guide the development of personalized therapeutic strategies, optimizing treatment planning in advance.