From waveform landmarks to model features AI
The extraction stage creates 46 candidate inputs for each original heartbeat: ten timing measurements, eight descriptors for each of ECG, PPG, Film0 and LC_BCG0, and four original Film0 regression features. These measurements stay aligned to participant, beat ID and sample time.
Timing
The ten timings are ECG-to-PPG rise and peak; ECG-to-Film0 J; Film0 J-to-PPG rise; Film0 J-to-load-cell J; I–J and J–K intervals for each BCG channel; and ECG cycle duration. Delays use the original 1 kHz indices. A PPG rise marker is maximum positive slope, not the pulse foot.
The original four Film0 features are natural-log I–J and J–K durations in seconds, plus J-minus-I and J-minus-K amplitudes in volts. For example, an I–J interval of 100 ms enters the logarithmic feature as ln(0.1), not ln(100). See BCG and the source-study comparison.
Magnitude and shape
Each waveform contributes peak-to-trough span, median offset, median-centered RMS, maximum positive slope, 10–90% rise time, half-height width, and two profile residuals. Rise and width require the relevant crossings; absent crossings remain missing. A unit-range profile subtracts the cycle minimum and divides by its span. Shape residuals measure RMS distance from a person’s median normalized profile under cycle or own-landmark alignment.
For prediction, templates are learned within the training window only. Descriptive bands in the explorer use the observed excerpt and are not model templates. See variability and alignment.
Why 42 features remain
A calibration-input-only redundancy screen groups strongly related candidates using Pearson and Spearman correlations within people. Pairs require at least 20 paired windows and correlations in at least 30 participants. Complete-link groups require the 20th percentile of absolute correlation to reach 0.90 for both measures. Representatives prioritize original BCG features and observed coverage.
Four candidates are removed: untransformed Film0 I–J and J–K durations (log versions remain), PPG centered RMS (span remains), and ECG own-landmark shape residual (cycle-aligned residual remains). This panel is frozen for the reported models; selection is not repeated inside each tuning split. Descriptive across-beat variability summaries are not extra predictive features. See personal ridge.