Each horizontal track shares the same time axis. The discomfort track at the top is your label layer — every dot is a logged event, height encodes severity. Below it, the position and sleep-stage tracks are the most likely mediators; the heart rate, HRV, and pacing-mix tracks are your candidate predictors. Vertical ochre rules mark interventions (programming changes, medication, behavioral experiments) — these are the moments the model's training distribution shifts, and you should expect feature relationships to differ on either side of them.
Look first for whether discomfort dots cluster on supine bands and within high-AP-VP windows. If they do, AV-timing mismatch at LRL is your live mechanism. If they cluster on supine bands but pacing mix is unremarkable, suspect atrial undersensing or a non-pacing diastolic mechanism. If they don't cluster on supine at all, your priors were wrong.