Feasibility of vocal biomarkers in heart failure: a multidisciplinary study report

Abstract

Introduction Vocal biomarkers have emerged as a promising non-invasive, low-cost approach for monitoring congestive status of heart failure (HF).Voice-based technologies may expand access to specialized care, especially in resource-limited settings. Yet, existing real-world challenges related to accessibility, usability, and health equity compromise clinical applicability.Purpose:To report the initial feasibility experience of an ongoing multidisciplinary study aimed to develop an AI algorithm capable of monitoring congestive status of HF using voice signal analysis, with focus on testing the decision-tree logic and the intended AI model type.Methods:This is the initial phase of a multi-center research endeavor considering conceptualization, context assessment, AI-mediated care to develop an accessible, inclusive and cutting-edge HF tele-monitoring technological tool combined with vocal biomarkers. A multidisciplinary international team including AI researchers, cardiologists, advanced practice nurses in cardiology and speech-language specialists from Africa, North America, South America developed an AI-based patient follow-up conversational agent using voice recordings collected through telephone calls.The proposed system leverages Large Language Models(LLM) based conversational and clinical workflow-inspired flowcharts to automate patient check-ups. A tele-monitoring protocol used at a large HF referral center was adapted into a clinical decision tree aligned with current guidelines. For exposure testing, two clinically stable adult outpatients with HF (ACC/AHA stages B–D) were included to evaluate the decision tree, voice acquisition process and suitability of the AI model rather than clinical performance.Results:The first patient, with higher health literacy, completed the telephone monitoring demonstrating good feasibility and fluidity of the decision-tree-guided interaction.In contrast, the second patient experienced substantial difficulty following the decision tree due to limited health literacy, requiring repeated explanations and rephrasing of questions to obtain valid responses. Applicability challenges observed were poor call quality, background noise, telephone device. Digital and health literacy affected voice data quality and AI-based voice technology requiring new approaches. Multidisciplinary and international collaboration was essential for protocol adaptation to reduce potential biases affecting equitable deployment of the proposed technology.Conclusions: Structured voice collection integrated into routine tele-monitoring workflows has variable applicability depending on patient context and local conditions. Successful design and implementation of an AI-based voice technology depends on addressing accessibility, usability, equity-related challenges from the earliest stages of development. Incorporating multidisciplinary perspectives and real-world care contexts favors real impact to the global burden of HF.Initial AI-based patient follow-upFor image description, please refer to the figure legend and surrounding text.