Data Science and Analytics for Precision Rehabilitation (DAPR) Center - Research Project

University Of Southern California, Los Angeles CA

Investigators

Abstract

Modified Project Summary/Abstract Section Precision rehabilitation seeks to tailor rehabilitation therapy by considering each person’s unique biological, personal, and/or socioeconomic factors, aiming to deliver more effective clinical care. This approach relies on artificial intelligence (AI) and machine learning (ML) methods to create accurate predictive models, which require large, heterogeneous datasets. However, rehabilitation datasets are often small and varied, while electronic health record (EHR) datasets are large but lack necessary rehabilitation variables. To address these challenges, the Data Science and Analytics for Precision Rehabilitation (DAPR) Center Research Project will develop comprehensive data schemas and user-friendly informatics tools to aggregate existing rehabilitation data, creating large datasets for predictive modeling. DAPR will build extensive data dictionaries, develop tools for dataset annotation, and enable federated queries to access harmonized data globally, thus advancing AI/ML algorithms for predictive modeling. The research project has three specific aims. Aim 1 will aggregate and harmonize different rehabilitation datasets within stroke by developing a comprehensive data schema (DAPR- Stroke) using the ENIGMA Stroke Recovery dataset (N>2000), aligning with existing common data models and developing specifications for new rehabilitation data elements as needed. Existing informatics tools for web- based annotation and harmonization that have been successfully applied in neuroimaging (Nipoppy, Neurobagel, led by Co-Is Poline, Kennedy) will be adapted to annotate the dataset and the schema will be expanded by annotating additional stroke datasets (N>1500). Aim 2 will develop and apply precision rehabilitation algorithms to predict outcomes after stroke using harmonized large datasets identified via federated query across DAPR data nodes and applying AI/ML for predicting outcomes. Black-box models like recurrent neural networks, known for high accuracy but low explainability, will be compared with gray-box models like Hierarchical Bayesian algorithms, which offer better explainability but lower accuracy. Final models utilizing multimodal data will be developed and shared. Aim 3 will harmonize and apply precision rehabilitation algorithms to varied rehabilitation conditions. DAPR data schemas, informatics tools, and precision rehabilitation algorithms developed in Aims 1 and 2 will be adapted for Parkinson’s Disease and Cerebral Palsy, ensuring their flexibility across different conditions. Enhancements will be made to easily adapt DAPR’s tools to other conditions, linking efforts to ongoing big data initiatives for maximum impact. Successful completion of this work will modernize medical rehabilitation research by: (1) creating tools for harmonizing rehabilitation datatypes, (2) increasing the availability of large AI/ML-ready datasets, (3) advancing infrastructure for federated query and processing, and (4) developing algorithms for personalized precision rehabilitation. All activities will be informed by advisory committees, including people with lived experiences (PWLE); harmonized datasets and tools will be disseminated via our others cores to promote community uptake of these tools and maximize impact.