High-resolution automated Fugl-Meyer Assessment using sensor data and regression model

Motor function assessment is a critical component in post-stroke rehabilitation program. Fugl-Meyer Assessment (FMA) is regarded as the most comprehensive tool to describe post-stroke patient's motor function and is widely utilized. However, the FMA scoring system classifies patient's motor function within only three levels (0, 1, or 2) for each assessment item. Hence, it's difficult to observe minor improvement in patient's motor function. To address this issue, a high-resolution scoring system was proposed in this research. Six regression models were built for six upper-extremity FMA items. These models were trained using data gathered from Microsoft Kinect sensor and glove sensor. The trained models could predict FMA scores with the resolution of 14 fractional digits. The representativeness of the predicted scores was evaluated by calculating their Pearson's correlation with the actual kinematic variables. All predicted scores represent patient's motor function better than standard FMA score. In addition, the highest correlation's mean of 0.58 was found on shoulder elevation item which utilized Neural Network Regression algorithm. This algorithm also outperformed the other in most assessment items. The ability to observe the changes in patient's motor function in detail helps therapists providing more responsive treatment and could likely increase the patient's adherence toward the given treatment.