Voice as a Biomarker to Detect Acute Decompensated Heart Failure: Pilot Study for the Analysis of Voice Using Deep Learning Models

Abstract

AbstractBackgroundAcute decompensated heart failure (ADHF) is a systemic congestion state requiring timely management. Admission for ADHF is closely related to the readmission and post-discharge mortality in patients, which makes it imperative to detect ADHF in its early stage.MethodsPatients with ADHF needed admission were eligible for enrollment, and those with respiratory infection, sepsis, lung/vocal cord disease, acute coronary syndrome, or serum creatinine>3mg/dL were excluded. A total of 112 patients were enrolled between July, 2020 and December, 2022. Voice was recorded two times: at admission for ADHF, and at discharge. Patients were asked to phonate five Korean vowels (‘a/e/i/o/u’) for 3 seconds each, and then to repeat the sentence ‘daehan minkook manse’ five times. Low-level audio features were extracted for classification. Then, Mel-Spectrogram was extracted from waveform and used as input features of the deep learning-based classification models. Two kinds of the deep learning-based classification models, convolutional neural networks and Transformer, were adapted for the further analysis.ResultsFor 100 patients in the final analysis, we randomized patients into two mutually exclusive groups: a training group (n=88) and a test group (n=12). In the analysis with low-level audio features, harmonics-to-noise ratio and Shimmer showed classification potential. Then, deep learning models were trained to classify whether certain voice belongs to ADHF state or recovered state. We treated it as a binary classification task, and the best performing model achieved a classification accuracy of 85.11% with DenseNet201. The classification accuracy was improved as 92.76% with ViT-16-large after inputting additional classic features of heart failure. With adding the low-level audio features in a training process, classification task accuracy was improved in DenseNet201 for about 2%.ConclusionsOur results proposed the clinical possibility of voice as a useful and noninvasive biomarker to detect ADHF in its early stage.