HRL on Real Data

wearable accelerometer · body-position labeling

Real clinician-labelled accelerometer recordings (the Huxley body-position study). Each second's gravity vector is a reading; the hrl.RelaxationLabeler (affinity prior + temporal smoothing + a noise label) assigns a posture. Postures are calibrated per subject (sensor mounting varies), then matched to the clinician's codes and scored. No synthetic data.

mean accuracy across subjects vs. clinician ground truth

Per-subject results

subject	seconds	positions	accuracy	HRL flips	clinician flips	noise

HRL flips > clinician flips: the engine slightly over-segments (flags more posture changes than the clinician marked) — an honest artifact of per-second labeling without heavier temporal damping.

A recording, second by second

Confusion matrix — clinician (rows) vs HRL (cols)

The gravity vectors in 3D — drag to spin; coloured by HRL posture

each point a one-second reading; postures cluster where the torso held still. The clusters are the per-subject "canonical directions".

← Gospel-HRL · the same engine that reads scripture and grows bodies, here on real wearable sensor logs. · the full portfolio →