Details

Automated Analysis of Achilles Reflex Response Using Signal Processing and Supervised Learning

Year: 2026

Term: Winter

Student Name: Maria Villagomez

Supervisor: Matthew Holden

Abstract: The Achilles tendon reflex half-relaxation time (HRT) is a clinically established marker of thyroid function, with hypothyroid patients displaying lengthened reflex responses compared to euthyroid individuals. Despite its diagnostic value, obtaining quantifiable metrics remains limited by the need for manual annotation: while physicians can visually assess reflex delay at the bedside, reliable HRT measurement requires precise marking of reflex onset and offset from a recorded signal. This thesis presents a fully automated pipeline for HRT extraction from smartphone videos of the Achilles reflex, requiring no manual labelling at inference time. The pipeline operates in three stages: (1) frame-wise segmentation of reflex intervals from the heel-displacement signal using U-Time; (2) complex-level validity classification using MiniROCKET, which filters out reflex complexes unsuitable for HRT measurement; and (3) HRT computation via a velocity-based onset detection method. All components are evaluated under leave-one-subject-out cross-validation across 18 healthy volunteers, each contributing one left-leg and one right-leg recording. An ablation over 24 pipeline combinations identified velocity-based onset detection as the dominant factor, reducing mean absolute error (MAE) by approximately 5× relative to segment-boundary onset across all pipeline modes. The best full-ML pipeline (peak de- tection combined with MiniROCKET classification and velocity-based onset) achieved an MAE of 21.8 ms across 30 recordings. A one-sided non-inferiority test confirmed that model- rater differences fall within the ±30 ms clinical equivalence margin (p = 0.0008), confirming that the pipeline is a non-inferior substitute for a human rater for the purpose of euthyroid/hypothyroid screening. Inter-rater MAE among four clinicians was 14.7 ms, providing context for the model’s 21.8 ms error. These results demonstrate that accurate, fully automated Achilles reflex HRT extraction from consumer-grade smartphone video is clinically feasible, and motivate future validation in cohorts including diagnosed hypo/hyperthyroid patients.