Medical Student University of Miami Department of Neurological Surgery Miami, FL, US
Introduction: NeuroBridge is a machine-learning (ML) algorithm developed to evaluate spinal cord compression injury, providing community health centers in low- and middle-income countries (LMICs) with the decision-making capacity of multi-tier institutions. Data collected by community workers is computed into triage scores, enabling evidence-based decisions regarding patient transport to higher-level care facilities. NeuroBridge integrates variables predicting severity of injury, deficit in neurological function and likelihood of long-term impairment, addressing critical gaps in predictive tools like the Injury Severity Score (ISS).
Methods: Retrospective analysis of the National Spinal Cord Injury Model Systems dataset was conducted, which included n=35,675 observations, n=658 variables from 1972-2021. ML techniques were utilized to compute two key metrics: Deficit Severity Score (DSS), reflecting immediate sensorimotor deficits, and Global Severity Score (GSS), incorporating injury mechanism, patient demographics, and clinical outcomes to assess long-term risk. Random forest models and Shapley Additive exPlanations were used for algorithm development and predictive variable identification. K-fold cross-validation ensured stable predictive performance. Metrics were computed from variables routinely collected by community healthcare workers.
Results: The NeuroBridge demonstrated strong performance with overall accuracy of 93.1%, and Area Under the Receiver Operating Characteristic Curves of 0.79 and 0.74 for the DSS and GSS, respectively. Overall sensitivity was 76.8%; specificity was 96.5%. The ASIA Motor Index and long-term clinical outcomes, including rehospitalization rates and mobility scores, were highly predictive. The DSS captured immediate functional deficits while the GSS broadly assessed long-term morbidity and clinical outcomes, addressing limitations of traditional tools.
Conclusion : NeuroBridge provides a scalable, data-driven solution for triaging spinal cord injuries in LMICs, equipping non-specialist healthcare providers to make informed decisions on patient care and resource allocation. By integrating immediate injury severity with key outcome factors, the algorithm surpasses traditional models, providing practical measures of patient urgency. Future efforts will focus on refining the model’s predictive accuracy and expanding its applicability to new regions and injury types, with the goal of transforming healthcare delivery and improving outcomes in LMICs.
.jpg)
