What My HRV Pipeline Got Wrong: Three Leakage Modes, Their Corrections, and an Honest Re-test of Damped-Oscillator ECG Features

I previously reported 99.7% accuracy for heart rate variability (HRV)–based stress classification and 99.9% sensitivity and specificity for cardiovascular risk. This preprint audits and corrects those results. On the public MIT-BIH Arrhythmia Database (3,522 overlapping 3-minute HRV windows, 36 patients), three leakage modes each produce near-perfect accuracy: target leakage (a stress label, LF/HF ≥ 1.5, computed from an input), identity leakage (SDNN excluded but exactly recoverable from SD1 and SD2), and evaluation-unit leakage (patient sex predicted at 0.986 record-wise but 0.633 with patients held out). On SWELL-KW, the dataset behind the original claims, the LF/HF ≥ 1.5 label is true for 100% of windows; the experimental stress condition is predicted at 1.000 on the released train/test files but at 0.471 with participants held out, below the 0.558 majority baseline. Damped-oscillator features on raw ECG beats, the one component that appeared to work, were re-tested at the true 360 Hz rate under the standard inter-patient split: macro-F1 0.673, versus 0.671 for the baseline and 0.672 for a wavelet comparator. The earlier gains came from a patient-overlapping split and a sampling-rate error. The paper provides a corrections table for every earlier claim and a seven-item pre-publication checklist that catches each failure. Analysis code and exact outputs are included.