JMIR Publications

ABSTRACT

Background:

Spinal cord injury (SCI) is a severe disorder of the central nervous system caused by trauma, neoplasms, or inflammatory processes, characterized by complete or incomplete impairment of motor and sensory functions accompanied by dysfunction of the autonomic nervous system. Globally, the incidence of SCI ranges from 40 to 80 cases per million population, predominantly affecting males aged 15–35 years. Current management strategies for SCI comprise surgical interventions, pharmacological therapies, emerging technologies, and rehabilitative regimens incorporating assistive technologies. Nevertheless, optimizing treatment strategies and enhancing therapeutic efficacy remain critical challenges requiring in-depth investigation in clinical practice. To address these challenges, predictive models integrating multifaceted clinical factors may facilitate risk assessment and prognostic probability estimation, thereby guiding clinicians in establishing precise rehabilitation goals and discharge plans. Machine learning, a subfield of artificial intelligence, constructs predictive models by deciphering intrinsic correlations within clinical data. Its value in developing clinical decision support systems has catalyzed interdisciplinary research. While machine learning has been extensively applied to SCI diagnosis and prognostication, its practical implementation faces limitations. Heterogeneity in patient characteristics, therapeutic strategies, and imaging manifestations poses challenges for developing robust neurofunctional outcome prediction models Moreover, inconsistencies in prediction targets (e.g., ASIA Impairment Scale (AIS), Spinal Cord Independence Measure (SCIM)) and the absence of standardized evaluation frameworks compromise model comparability. Currently, evidence regarding the reporting quality and risk of bias in studies using machine learning for SCI prognostic modeling remains limited.

Objective:

To systematically evaluate the reporting quality and risk of bias in existing studies on machine learning (ML)-based prognostic prediction models for spinal cord injury (SCI) through a comprehensive literature analysis, thereby exploring the performance of ML algorithms in predicting SCI outcomes.

Methods:

We searched CNKI, Wanfang Database, VIP Database, SinoMed, PubMed, Web of Science, Embase, and Scopus from inception to May 14, 2025. Two investigators independently screened studies, extracted data, and assessed bias risk. The reporting quality and risk of bias were evaluated using the TRIPOD statement and the PROBAST. Descriptive statistics and visualization tools were employed to analyze the results.

Results:

Nineteen studies were included. The adherence to TRIPOD items ranged from 54% to 82% (median: 68%). Critical methodological details were frequently underreported, including treatment information, blinded assessment of predictors, handling of missing data, risk stratification criteria, sample size justification, full model presentation, model interpretation, and availability of supplementary materials. All studies were judged as high risk of bias, particularly in statistical analysis domains, primarily due to small sample sizes that increased risks of overfitting or underfitting, potentially leading to systematically overestimated predictive performance (e.g., accuracy, AUC). The applicability of all studies was rated as low.

Conclusions:

ML demonstrates significant methodological value in developing prognostic prediction models for SCI patients, often outperforming traditional models. It provides a novel decision-making tool for functional recovery assessment and personalized intervention planning. However, future investigations necessitate larger sample sizes and rigorous adherence to TRIPOD and PROBAST guidelines to address methodological biases. Multicenter collaborative databases and standardized data collection protocols should be established to reduce inter-study heterogeneity.