Machine Learning-Based Predictive Model for Functional Independence in Spinal Cord Injury: Protocol for a Predictive...

Spinal cord injury (SCI) causes substantial disability by disrupting spinal pathways, making functional independence a central rehabilitation goal. In 2024, approximately 15.4 million people worldwide were living with SCI. Despite its clinical relevance, traditional prognostic tools, notably the International Standards for Neurological Classification of Spinal Cord Injury (ISNCSCI), have limitations. Furthermore, access to advanced diagnostics constrains prediction. Existing models often fail to estimate independence, limiting goal setting and resource planning.

This study aims to develop and validate machine learning (ML)-based rules that predict functional independence, as measured by the Spinal Cord Independence Measure, version III (SCIM-III), at 3, 6, and 12 months post injury in individuals with SCI. The model combines clinical admission predictors, readily available to enhance predictive performance: age, sex, time since injury, upper and lower extremity strength, ISNCSCI data (American Spinal Injury Association Impairment Scale grade and neurologic level of injury), rehabilitation type, and the Trunk Control Scale score validated by the research team.

Using retrospective electronic clinical records (2015-2026) from Mexico's Nacional de Rehabilitación Luis Guillermo Ibarra Ibarra (INR LGII; a tertiary national rehabilitation reference center) for model development and a prospective independent cohort (October 2026-October 2027) for external validation, the study will develop ML models to accurately and generally predict SCIM-III outcomes. Eligible participants are adults (≥18 years of age) with subacute or chronic SCI. Six architectures are compared: linear regression, classification and regression tree, categorical boosting (gradient boosting algorithm), light gradient boosting machine, multilayer perceptron, and Gaussian process regression, using 10-fold stratified cross-validation for internal validation. Performance is assessed by root-mean-square error, mean absolute error, R², area under the curve, calibration plots, and decision curve analysis. The approach prioritizes reproducibility, interpretability, and clinical applicability, addressing prior models' limitations (small sample sizes, limited validation, and impractical input requirements). Key steps include data preprocessing, feature selection, model training with appropriate algorithms, and robust evaluation against existing prognostic benchmarks. External validation and ethical considerations are integrated, with commitments to data sharing where feasible.

The study was approved and funded in October 2025. The retrospective development cohort targets 500 registries (from 2025-2026) collected between October 2025 and June 2026. A total of 119 records were processed by March 2026. ML development is planned between June 2026 and March 2027. The external prospective validation cohort (n=100) is planned for October 2026-October 2027, and validation for June 2027-March 2028. Results are expected by January 2028.

A reliable, clinically actionable ML-based prediction tool that estimates SCIM-III trajectories to support goal setting, rehabilitation planning, and resource allocation in SCI care will be developed. Model performance will be benchmarked against published SCI prediction rules, and it will be considered clinically useful if it outperforms the mean-score baseline and existing rules on the prospective validation cohort.