Music-Inspired MK545Pat for AF/PAF/Non-AF Classification Using ECG Signals

Abstract

Paroxysmal atrial fibrillation (PAF) may be missed on intermittent electrocardiographic (ECG) recordings. We evaluated MK545Pat, a deterministic one-dimensional descriptor derived from the first-appearance order of the notes in Mozart’s Piano Sonata K.545. The CPSC 2021 dataset contains 1436 records from 105 patients, from which 116,042 15 s segments were analyzed. Each segment produced 2304 histogram features from 2942 unit-stride 59-sample blocks. Segment-level 10-fold cross-validation achieved 97.49 ± 0.18% accuracy (range, 97.16–97.71%), and a separate 90:10 segment-level holdout achieved 97.49% accuracy and 95.85% macro-F1. A separate patient-wise leave-one-subject-out (LOSO) analysis held out all records and segments from one patient at a time. The unweighted mean of the 105 held-out patient-specific segment accuracies was 71.29% (median, 78.17%; SD, 24.22%; range, 0.00–96.00%). Pattern controls were lower for identity, reverse, and interleaved assignments, while a within-group permutation retained 97.49%, as predicted by the invariance analysis. The results support MK545Pat as a reproducible feature-engineering approach and show the expected performance gap between segment-level and cross-patient evaluation.