FAIR-Coach: Biomechanically Informed Evidence Representation for Movement Diagnosis with Offline Squat-Based Evaluation

Abstract

Accurate classification is not sufficient for reliable movement diagnosis. The system must also identify and evaluate the biomechanical evidence underlying its decisions. Domain-shift experiments support this distinction: full or partial classifier collapse occurred in 77.8% of out-of-domain transfer cases.

A central gap is that changes in evidence representation often come with changes in reasoning complexity. Predictive accuracy alone does not establish whether a diagnosis relies on biomechanical evidence as intended.

This thesis investigates these questions using FAIR-Coach and a structured per-repetition representation called RepEvidence. The core empirical evaluation is offline and squat-based; Path A and live agentic interaction belong to the broader system architecture but are not part of the core diagnosis experiments. The evaluation compares generic, globally aggregated evidence with diagnosis-aligned biomechanical evidence and tests four relational reasoning configurations against corresponding non-relational controls. The primary comparison follows a pre-specified, frozen protocol on a held-out TEST split of five participants with paired bootstrap confidence intervals, followed by multi-seed and subject-level robustness analyses on the same TEST cohort.

Diagnosis-aligned evidence outperformed generic evidence at all seven seeds under an architecture-fixed Reasoner v2 comparison (mean Δ = +0.038, SD = 0.010). A separate protocol-matched Concat+MLP rerun showed the same direction of effect (mean Δ = +0.052, SD = 0.010), and the advantage was positive for all five TEST participants. Because the diagnosis-aligned configuration also adds ten measurements, these results do not isolate alignment alone, and absolute performance remained modest (TEST micro-F1 0.3435). In contrast, none of the four relational configurations showed a statistically confirmed advantage over their corresponding non-relational comparisons on the TEST split, and one performed significantly worse. In a corrected intervention-based evaluation of the final diagnosis model, only one of three tested labels satisfied the grounding criterion, so diagnosis-level grounding was not demonstrated. At the language layer, LoRA-DPO trained on template-constructed preference pairs improved diagnosis-to-language grounding under an automatic evaluation protocol on 40 held-out examples. Unsupported claims decreased from 75% to 17.5%, evidence specificity increased from 75% to 92.5%, and consistency with the model's predicted diagnosis increased from 22.5% to 77.5%.

In this offline squat-diagnosis setting, the full diagnosis-aligned evidence configuration had a greater impact on predictive performance than increasing reasoning-architecture complexity. The results therefore support diagnosis-aligned evidence representation while also identifying clear limits to claims of diagnosis-level biomechanical grounding.