Heart Rate Memory: QT and T-Wave Dynamics
How the T wave follows heart rate through exercise and recovery
The heart’s electrical recovery adapts gradually to changes in heart rate. This delay creates distinct QT–RR and T-wave amplitude–RR paths during exercise and recovery. Explore these dynamics in the synchronized plots below. Our research uses personalized models to characterize this adaptation and improve detection of prior myocardial infarction.
What is on this page: recorded data only. Two de-identified stress-test recordings are played back beat by beat, and every curve below is measured from those recordings. The personalized models described in the papers are a separate step in the research: they are driven by a standardized heart-rate input so that responses can be compared across people. No model-generated response appears here.
Display options
Space play · ←→ ±5 s · Shift+←→ ±30 s · HomeEnd · 12 subject · M measurements · L loop
Methods & data
Two de-identified single-lead Holter recordings from a cardiac stress test, each covering the ten minutes either side of peak heart rate. Time is abstract — there are no dates or clock offsets. ECG is baseline-corrected and decimated to 100 Hz.
Beat detection, quality filtering and robust outlier rejection were applied upstream; the page plots what it is given and does not re-filter. The smoothed series driving the loops is a 41-beat centred moving mean, so it uses beats either side of the current one and is not a causal, real-time estimate. Between beats the smoothed series is linearly interpolated so the point glides; raw per-beat values are never interpolated.
Marker positions come from the delineator's sample indices on a 10 ms grid, while the QT and T-amplitude figures quoted are the source measurements. The two disagree slightly by construction — reconstructing QT from the marked endpoints differs from the measured QT by up to 8 ms. Where the delineator gave no point, or returned landmarks out of order, the T markers and QT shading for that beat are suppressed rather than repaired; the beat still contributes to the smoothed loops.
The averaged beats view takes every beat within ±30 s of the current moment, aligns them on the R-peak and averages them. Each curve spans −25 % to +75 % of the average RR of the minute it sits in, so it ends three quarters of the way through the cycle. That keeps the next beat out of the average: always its QRS, and its P wave too wherever diastole is long enough. Near peak heart rate it is not — there the T wave itself runs past where the next P wave begins, so the two genuinely overlap and no choice of window can separate them.
The time axis is locked for the whole record at the largest smoothed RR, so curves stay comparable from the first to the last; a curve recorded at a fast heart rate is simply shorter. Earlier curves stay on screen in grey, and the T-peak and T-end markers are drawn in green so they read against the orange, the blue and the grey alike.
Exercise is shown in orange, recovery in blue, split at peak heart rate.
Recorded, not modelled. Everything drawn on this page is measured from the two recordings: the beats, the intervals, the loops and the averaged beats are all observed data from a real stress test. The research goes a step further and fits a personalized model per participant, then drives every model with the same standardized rest–exercise–recovery heart-rate input so that the resulting responses are comparable between people and can be used as features for classification. Those standardized, model-generated responses are described in the papers below and are not shown here.
Read more
- Karimi S, Koscova Z, Li Q, Clifford GD, Vaccarino V, Shah AJ, Sameni R. A System Identification Approach to Subject-Specific QT-RR Dynamics in ECG-Based Myocardial Infarction Classification. Computing in Cardiology, 2026.
- Karimi S, Koscova Z, Li Q, Clifford GD, Vaccarino V, Shah AJ, Sameni R. A System Identification Approach to Analyzing T-Wave Amplitude Heart Rate Adaptation: A Case Study in Myocardial Infarction Detection.