Speech Signal Preprocessing and Feature Extraction for Biomarker Identification in Acute Heart Failure: A Pilot Study

This study presents a configurable speech signal preprocessing and feature-extraction workflow for identifying candidate acoustic biomarkers in acute heart failure. The workflow was evaluated in 12 patients hospitalized with acute heart failure. Short recordings of repeated vowels /a/, /i/, and /o/ were acquired shortly after admission and again before discharge following treatment and clinical stabilization. The pipeline included active-RMS normalization, automatic segmentation of repeated vowels, optional edge trimming, alternative pitch-estimation variants, and extraction of three feature families: phonatory and temporal measures, spectral-shape descriptors, and MFCC-based cepstral features. Within-patient admission-to-discharge differences were evaluated using two-sided Wilcoxon signed-rank tests, with nominal p-values interpreted as exploratory. Phonatory and temporal measures produced the most consistent exploratory findings. The pause-duration trend for /a/ decreased between admission and discharge and was the most configuration-stable individual candidate. CPP maximum and CPP range for /o/ increased consistently across the evaluated phonatory configurations, indicating systematic changes in cepstral prominence. Shimmer-related measures provided additional exploratory findings. MFCC measures showed complementary changes, particularly in MFCC11 variability for /o/, whereas spectral-shape effects were generally weaker and less consistent. Because the study involved a small, single-center cohort without a control group or external validation, the findings should be regarded as hypothesis-generating candidate acoustic measures rather than clinically validated biomarkers.