Machine Learning-Based Prognostic Models for Functional Outcomes in Spinal Cord Injury: Systematic Review

Machine learning is increasingly used to develop prognostic prediction models for spinal cord injury. Nevertheless, current studies exhibit heterogeneity in outcome measures, predictors, modeling strategies, and validation methods. Moreover, the reporting quality, risk of bias, and clinical applicability of these models have not been systematically evaluated using assessment tools specific to prediction models.

This review aimed to assess the reporting quality and risk of bias of machine learning-based prognostic models for spinal cord injury, and evaluate their clinical applicability, model features, validation, and implementation barriers.

We searched the China National Knowledge Infrastructure (CNKI), Wanfang Data, VIP Database, Sinomed, PubMed, Web of Science, Embase, and Scopus databases from their inception up to May 14, 2025. Two investigators independently screened studies, extracted data, and assessed risk of bias. Reporting quality and risk of bias were evaluated using the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement and the Prediction Model Risk of Bias Assessment Tool (PROBAST). Descriptive statistics and narrative synthesis were used to summarize the findings.

In total, 19 cohort studies were included. TRIPOD adherence ranged from 54.8% (17/31) to 81.1% (30/37), with a median of 74.2% (IQR 64.5%-77.4%). Overall, all 19 studies were judged to have a high risk of bias, mainly because of limitations in the analysis domain. Only 1 (5.3%) study included external validation, while 16 (84.2%) studies used internal validation and 2 (10.5%) studies reported model development only. No study justified the sample size; 6 (31.6%) studies reported imputation or other methods for handling missing data, and calibration was rarely reported.

Machine learning shows potential for spinal cord injury prognostic modeling, especially when complex clinical or imaging data are available. However, existing evidence is limited by incomplete reporting, high risk of bias, substantial heterogeneity, and limited external validation. Larger, methodologically robust studies with standardized outcomes, external validation, and evaluation of clinical usefulness are necessary before these models can be implemented in routine clinical practice.

PROSPERO CRD420251071502; https://www.crd.york.ac.uk/PROSPERO/view/1071502.