Development and internal validation of an explainable machine learning model for predicting functional disability trajectories in hospitalized older patients with chronic low back pain.

OBJECTIVE: Functional disability poses a critical challenge to healthy aging in older patients with chronic low back pain (CLBP). However, understanding of heterogeneous functional disability trajectories in this population remains limited, and effective risk stratification tools are lacking. This study aimed to develop and internally validate an explainable machine learning (ML)-based model for predicting heterogeneous functional disability trajectories and to construct an exploratory trajectory-based risk stratification pathway. METHODS: In this prospective cohort study, data on general characteristics, functional disability, pain level, physical activity, depression, and frailty were collected from older patients with CLBP. Growth mixture model was employed to identify heterogeneous functional disability trajectories. Ten explainable ML-based models were developed to predict these trajectories, and exploratory classification thresholds were determined using the Youden index. RESULTS: Four distinct functional disability trajectories were identified: persistent mild (PM), moderate and improving (MI), moderate and stable (MS), and persistent severe (PS). LightGBM achieved an AUC of 0.895 (95% CI: 0.855-0.941) in the validation set, with a calibration slope of 1.832 (95% CI: 1.625-2.021), intercept of 0.522 (95% CI: 0.317-0.795), and Brier score of 0.114 (95% CI: 0.098-0.137). Exploratory classification thresholds were 0.407, 0.308, and 0.320 for PM, MI, and MS, respectively, and 0.209 for the PS trajectory. CONCLUSION: This study developed an explainable ML-based prediction model for functional disability trajectories in older patients with CLBP and preliminarily explored a trajectory-based risk stratification pathway. The findings provide an interpretable, data-driven approach for estimating trajectory-specific risk within this dataset and may help inform future research on early risk identification. Further external validation, recalibration, and impact analyses are required before clinical implementation.