SpineTrace: Auditable Video-Based Scoliosis Screening with Temporal Pose Statistics

Abstract

Video-derived gait representations may support scalable pre-referral screening for scoliosis, but their evaluation must retain borderline cases, quantify model vulnerability, and distinguish measured performance from clinical or deployment readiness. This study evaluates SpineTrace as an auditable compact engineering baseline rather than a novel learning algorithm or state-of-the-art classifier. The public Scoliosis1K release comprised 1,493 gait sequences reportedly obtained from 1,050 adolescents, including 493 positive, 200 neutral, and 800 negative sequences. The supplied partition of 744 training and 749 test sequences was reproduced exactly. Feature scaling and hyperparameter selection were restricted to five stratified folds within the training partition, with three-class classification designated as the primary task. Among the reproduced RBF-SVM configurations, the highest macro-F1 was obtained using 102 temporal pose features derived from keypoint means and standard deviations. The model achieved 0.737 accuracy, 0.740 balanced accuracy, 0.702 macro-F1, and 0.890 macro one-versus-rest ROC-AUC. Recall was 0.917, 0.619, and 0.685 for negative, neutral, and positive sequences, respectively, demonstrating the greater difficulty of borderline and positive cases. Adding three silhouette and five postural descriptors reduced macro-F1 to 0.656, indicating that semantically plausible handcrafted features did not provide complementary predictive information. Independently retuned PAV-SVM and DTW-kNN baselines achieved macro-F1 scores of 0.545 and 0.473, while a compact ST-GCN adaptation achieved 0.645±0.055 across three optimization seeds. Removing keypoint-confidence features from the full model reduced macro-F1 from 0.656 to 0.618, and coordinate jitter equivalent to 2% of torso length reduced it to 0.287, identifying pose quality as a critical vulnerability. Fixed-test uncertainty was estimated using 2,000 class-stratified sequence bootstraps. Because the local release lacks participant mappings and raw RGB videos, participant independence, cross-estimator transfer, end-to-end latency, and cross-dataset validity could not be established. SpineTrace therefore provides a reproducible sequence-level baseline for auditing feature utility and failure boundaries, not a diagnostic or deployment-ready screening system.