Machine Learning/artificial Intelligence Algorithms for Predicting Outcomes of Traumatic Spinal Injury: A Systematic Review

Introduction: In recent years, machine learning (ML) and artificial intelligence (AI) algorithms have shown promise in predicting outcomes for individuals with traumatic spinal cord injury (SCI). This systematic review synthesizes existing literature to assess the efficacy of ML/AI in forecasting various clinical outcomes for SCI patients and to identify ML/AI models with the greatest efficacy in this task.

Methods: A comprehensive search was conducted across PubMed, Embase, Scopus, and Web of Science without applying filters beyond the search criteria.

Results: Our search yielded 2,568 studies, with 19 meeting inclusion criteria. The selected studies, primarily published from 2020-2024, encompassed models that examined all spinal levels, unspecified spinal levels, as well as specific regions, including cervical, thoracic, thoracolumbar, lumbar, and lumbosacral areas. Outcomes were grouped into five categories; only studies assessing the most frequent metric in each category were included in subgroups: AIS score prediction (5 studies, 21,464 patients) with an average accuracy of 0.762; prolonged length of stay (LOS) (3 studies, 222,617 patients), average AUROC of 0.738; prolonged ICU-LOS (3 studies, 50,576 patients), average AUROC of 0.716; ambulatory outcomes (4 studies, 10,712 patients), average AUROC of 0.893; and functional independence (3 studies, 4,081 patients), average R² of 0.721.

Among the ML/AI models utilized, the most frequently applied algorithms included Random Forest (9 studies), XGBoost (8 studies), Decision Tree (6 studies), and LightGBM and Support Vector Machines (5 studies each). Other algorithms, like CatBoost, Artificial Neural Networks (ANN), Gradient Boosting, Elastic Net, and k-Nearest Neighbors, were used less frequently. Ensemble models, including deep learning methods like Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), were also represented, albeit less commonly.

Conclusion : This review highlights the potential of ML/AI tools in SCI outcome prediction, with applications that could enhance clinical decision-making and personalize rehabilitation strategies. Future research should focus on validating these models prospectively and integrating them into clinical workflows to optimize patient care and outcomes.