Reporting completeness of prognostic prediction models using machine learning methods in physical therapy: A methodological study based on the TRIPOD+AI statement

Machine learning (ML) methods are increasingly applied to develop prognostic clinical prediction models (CPMs) in physical therapy, yet concerns remain regarding inadequate reporting transparency and limited reproducibility. The TRIPOD+AI guideline, published in 2024, provides updated and comprehensive reporting standards for regression- and ML-based prediction models. This methodological study aimed to evaluate the completeness of reporting of ML-based prognostic CPMs in physical therapy using the TRIPOD+AI statement. A systematic search of MEDLINE, Epistemonikos, and reference lists identified studies published between 2015 and 2024 that developed and/or validated ML-based prognostic models within the physical therapy discipline. Eligible full texts were screened, and data were extracted on study characteristics and adherence to all TRIPOD+AI items. Adherence was summarized descriptively at both item and study level, and subgroup analyses compared reporting completeness by model population (single vs. multiple models), use of external validation, and journal impact factor. Thirty-three studies met the inclusion criteria, the majority being development-only designs (91%). Overall reporting completeness was suboptimal, with adherence ranging from 26% to 58% (median 42%). Key deficiencies included the absence of sample size justification (reported in 9.1% of studies), insufficient reporting of missing data handling (11.5%), limited availability of full model details (12.1%), absence of reporting of fairness considerations (0%), and minimal adoption of open science practices such as registration, protocol availability, or sharing of data and code. Notably, no study reported patient or public involvement. Studies performing external validation demonstrated significantly lower adherence than development-only studies, while journal impact factor and number of models developed showed no association with reporting quality. In conclusion, reporting of ML-based prognostic CPMs in physical therapy is poor, lacking transparency, limiting reproducibility and hindering clinical translation. Greater adherence to TRIPOD+AI is essential to improve transparency, facilitate external validation, and support the safe implementation of ML methods in physical therapy.