Patient Similarity and Joint Features for Rehabilitation Outcome Prediction

Rehabilitation outcome prediction can be useful for clinicians providing therapy services for patients undergoing rehabilitation. Machine learning models trained with medical record information available at admission can predict rehabilitation outcomes fairly well. In our earlier work we found rehabilitation outcome prediction accuracy can be improved by also including sensor-based features that objectively quantify patient movement abilities during therapy tasks. In this paper, we hypothesize an even greater improvement in rehabilitation outcome prediction accuracy can be obtained with our proposed algorithm, called Joint Patient Prediction (JPP), which is based on patient similarity and joint prediction. To validate this hypothesis, we collected wearable inertial sensor data from 27 patients as they performed ambulatory tasks. We utilize features derived from the collected wearable sensor data and clinical information obtained from medical records at admission as inputs to our JPP algorithm. Results indicate we are able to achieve higher prediction accuracy when utilizing our proposed JPP approach over the baseline, nonJPP models. The proposed JPP method is a useful technique for mapping wearable sensor data collected during therapy tasks into rehabilitation outcome measures.