1441 - Machine learning-based platform for identification of sub-acute post-stroke recovery trials: The Spaulding recovery evidence platform (Spaulding-REP)

Objectives: To develop a machine learning system (using natural language processing) for living evidence mapping and identification of clinical trials testing interventions for sub-acute post-stroke recovery.

Design: We performed a living systematic review (Pubmed, Central, and Embase, searched by February 2023) using an open-access machine learning algorithm to assign probabilities of included studies (active learning method). This method used human input ( 2 expert researchers on systematic reviews) to “teach” the system the characteristics of included studies. First, the researchers labeled 100 abstracts of studies as included or excluded. Then, using text analysis and neural networks, the machine learning algorithm detected patterns in the abstracts associated with the included studies (keywords and word relationships). After proper calibration and model optimization, we fed 15,562 abstracts of potential studies, which were organized by likelihood of being included. We included only randomized clinical trials considering four domains: motor impairment, sensory disturbances, cognitive impairment, and emotional alterations.

Results: We identified 835 clinical trials (21,750 participants), including 28 different interventions. The identification process was accelerated by a factor of 10 compared to manual selection (180 days versus 18 days). The accuracy of our system was high (AUC of 0.88). The most common interventions were physical therapy, occupational therapy, and neuromodulation. The motor impairment domain was the most studied. The majority of studies had some concerns about the risk of bias assessments, mainly due to blinding and allocation concealment uncertainties.

Conclusions: Our findings demonstrate the potential of machine learning techniques to accelerate evidence synthesis (meta-analysis) and translation to rehabilitation practice. Similarly, it can help systematically identify the research gaps in the stroke recovery field, which future primary studies can improve.