Discharge Mobility After Osteoporotic Thoracolumbar Fractures Using a 2-Stage Machine Learning Prediction Model: Prospective Cohort Study

Abstract Background Discharge mobility after osteoporotic vertebral fractures is heterogeneous and difficult to predict at admission. Machine learning offers structured approaches for multiclass functional outcome prediction, but existing models predominantly address binary end points and report global performance metrics without explicit analysis of the error structure. Objective This study aimed to develop and evaluate a 2-stage machine learning framework with a predefined decision policy designed to reduce extreme (LOW-HIGH) misclassification for graded discharge mobility prediction in hospitalized patients with osteoporotic thoracolumbar fractures. Methods This prospective, single-center cohort study enrolled consecutive patients aged 65 years or older admitted for acute osteoporotic thoracic or lumbar fractures (N=158). Discharge mobility was categorized as LOW (<50 m or nonambulatory), MID (50‐500 m), or HIGH (>500 m or stair climbing). Clinical associations were analyzed using proportional odds ordinal regression. A predefined 2-stage architecture combined gradient-boosted trees (stage A: HIGH vs non-HIGH) and a random forest classifier (stage L: LOW vs non-LOW). The final 3-class prediction was assigned by a predefined decision policy designed to reduce clinically relevant extreme (LOW-HIGH) misclassification. All preprocessing steps, including fold-wise multiple imputation, were performed within training partitions only. The decision policy was additionally reevaluated in a nested cross-validation with fold-internal threshold selection across 10 repeated partitions. Results Prefracture HIGH mobility was the only significant independent predictor of discharge mobility (odds ratio [OR] 3.38, 95% CI 1.68‐6.80; P <.001); age, comorbidity burden, fracture grade, and treatment were not. In the prespecified analysis with fixed thresholds, performance was moderate (accuracy 0.494; balanced accuracy 0.450; macro- F 1 0.444). Binary discrimination yielded an area under the curve of 0.693 (HIGH vs non-HIGH) and 0.625 (LOW vs non-LOW). Using the predefined decision policy, major LOW-HIGH errors were less frequent than with the prefracture mobility–only baseline (5 vs 22), and no HIGH-to-LOW misclassification occurred in the prespecified analysis. In the leakage-free nested cross-validation, macro- F 1 was 0.388 (SD 0.031) and major errors had a mean of 8.3 (SD 3.4) vs 22.5 (SD 1.1) for the baseline, with fewer major errors in all 10 partitions. HIGH-to-LOW errors had a mean of 2.6 (SD 2.2) in this analysis. Conclusions Discharge mobility after osteoporotic thoracolumbar fractures is primarily determined by prefracture functional reserve rather than fracture severity or comorbidity. A 2-stage machine learning framework achieved moderate multiclass discrimination and reduced extreme LOW-HIGH misclassification relative to a mobility-only baseline; this reduction persisted under nested cross-validation, whereas the absence of HIGH-to-LOW errors was specific to the prespecified partition. External validation is required before clinical deployment.