Abstract 13566: Voice as a Biomarker to Detect Acute Decompensated Heart Failure: Pilot Study for the Analysis of Voice Using Deep Learning Models

Abstract

Acute 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. Body fluid overload due to ADHF can change voice characteristics by causing edema of lung, larynx, and vocal cord. Voice is a noninvasive and accessible biomarker which can be monitored even in telemedicine practices, but there are few studies in the area. We hypothesized that there would be certain differences of voice characteristics in patients with ADHF when admitted to the hospital and at discharge. This was a prospective study conducted in a tertiary hospital in Korea. Patients 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 1, 2020 and December 31, 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. Voice was classified by Convolutional Neural Networks (CNNs). Mel-Spectrogram was conducted to convert the audio from waveform to the spectrogram, and then, transfer learning was done to further classify the voice characteristics. Deep learning models including DenseNet, InceptionNet, and ResNet were conducted to classify voice samples and identify potential characteristics associated with ADHF. We trained the models 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 79% with DenseNet. Interestingly, the classification accuracy conducted with CNNs was highest for ‘u’ achieving 72%, compared to other vowels. To the best of our knowledge, this is the first study to try to detect ADHF with voice using deep learning models, especially with CNNs. We achieved a best classification accuracy of 79%. Our results proposed the possibility of voice as a useful biomarker to detect ADHF in its early stage.