Vocal Prognostic Digital Biomarkers in Monitoring Chronic Heart Failure: A Longitudinal Observational Study

Abstract: Objective: This study aimed to evaluate which voice features can predict health deterioration in patients with chronic HF. Background: Heart failure (HF) is a chronic condition with progressive deterioration and acute decompensations, often requiring hospitalization and imposing substantial healthcare and economic burdens. Current standard-of-care (SoC) home monitoring, such as weight tracking, lacks predictive accuracy and requires high patient engagement. Voice is a promising non-invasive biomarker, though prior studies have mainly focused on acute HF stages. Methods: In a 2-month longitudinal study, 32 patients with HF collected daily voice recordings and SoC measures of weight and blood pressure at home, with biweekly questionnaires for health status. Acoustic analysis generated detailed vowel and speech features. Time-series features were extracted from aggregated lookback windows (e.g., 7 days) to predict next-day health status. Explainable machine learning with nested cross-validation identified top vocal biomarkers, and a case study illustrated model application. Results: A total of 21,863 recordings were analyzed. Acoustic vowel features showed strong correlations with health status. Time-series voice features within the lookback window outperformed corresponding standard care measures, achieving peak sensitivity and specificity of 0.826 and 0.782 versus 0.783 and 0.567 for SoC metrics. Key prognostic voice features identifying deterioration included delayed energy shift, low energy variability, and higher shimmer variability in vowels, along with reduced speaking and articulation rate, lower phonation ratio, decreased voice quality, and increased formant variability in speech. Conclusion: Voice-based monitoring offers a non-invasive approach to detect early health changes in chronic HF, supporting proactive and personalized care.

• The paper introduces an AI-driven methodology using daily mobile-acquired voice recordings to detect early health deterioration in chronic heart failure, outperforming traditional metrics.
• It employs advanced acoustic feature extraction, time-series aggregation, and machine learning models to correlate vocal changes with clinical outcomes.
• The study highlights that integrating vocal biomarkers into remote monitoring enables personalized, proactive interventions for managing chronic heart failure.

Vocal Prognostic Digital Biomarkers for Chronic Heart Failure: Longitudinal Observational Insights

Introduction and Motivation

The study targets the advancement of digital health monitoring in chronic heart failure (HF) via non-invasive, explainable AI-driven analysis of vocal biomarkers. With HF posing persistent burdens on clinical resources and patient quality of life, there is a critical need for effective, low-burden early warning mechanisms for health deterioration. Current standard-of-care (SoC) home monitoring, relying on daily weight and blood pressure, provides limited predictive accuracy and is hampered by poor adherence. Voice, as a composite biomarker of physiological states, presents a compelling alternative, yet prior literature has chiefly focused on acute HF stages, cross-sectional designs, and lab-acquired voice data. The paper addresses these gaps by deploying an intensive, two-month, longitudinal, at-home monitoring protocol in chronic HF patients, leveraging mobile-based voice acquisition and high-dimensional feature extraction.

Methods

A cohort of 32 chronic HF patients conducted daily at-home voice recordings (vowels and structured speech tasks) and SoC physiological measurements. Alongside, symptom tracking and biweekly Kansas City Cardiomyopathy Questionnaire (KCCQ) scores were collected. Acoustic analysis pipeline encapsulated rigorous preprocessing, comprehensive feature extraction via OpenSMILE, SenseLab, and DisVoice (yielding >5000 vowel and 25 speech features), and time-series aggregation over 2–14 day lookback windows. High-dimensional time-series features were then subjected to repeated-measures correlation with KCCQ outcomes, guiding aggressive feature selection. Classification leveraged Random Forest and XGBoost models, including recursive feature elimination, subject-stratified nested cross-validation, and SHAP-based explainability. Statistical differentiation and a patient case study further validated the discriminative power of selected biomarkers.

Numerical Performance and Biomarker Characteristics

Voice-based features consistently outperformed SoC (weight and blood pressure) signals for prospective detection of health deterioration. The salient quantitative results include:

• Sensitivity and specificity for voice-based models were 0.826 and 0.782, compared to 0.783 and 0.567 for SoC.
• Voice features achieved Area Under the ROC Curve (AUC) of 0.77 and AUPRC of 0.80, outperforming SoC (AUC 0.65, AUPRC 0.68).
• Multimodal integration (Voice + SoC + Symptoms) further improved sensitivity (0.837), specificity (0.876), and Matthews correlation coefficient (0.703).
• Vowel time-series features, such as increased shimmer variability, delayed energy shift, and decreased dynamic spectral characteristics, were highly correlated with worsening HF status.
• Speech task features—reduced speaking/articulation rate, lower phonation ratio, diminished voice quality, and increased formant variability—were salient for deterioration.

Statistical comparisons confirmed large effect sizes ($r > 0.5$) for top-ranked features, including speaking/phonation rate and shimmer variability. In a fully withheld case (acute HF/hospitalization), voice-based predictors accurately tracked health decline, in contrast to SoC features, which remained uninformative.

Technical Implications and AI Aspects

The study demonstrates that mobile-acquired, passively collected voice can serve as a sensitive longitudinal biomarker for HF progression. Unlike traditional metrics, vocal features encode integrated aspects of respiratory, phonatory, and neurophysiological status. The feature engineering and explainability pipeline is methodologically robust, addressing typical pitfalls in digital biomarker research such as overfitting, redundancy, and interpretability. The use of time-series descriptors captures temporal dynamics critical in chronic disease evolution.

From a technical perspective, the evidence questions the clinical dogma that fluid status (weight surge) or blood pressure are sufficient early indicators in HF home management. Integration of objective, explainable AI analysis of digital voice data offers a concrete pathway for proactive, personalized interventions in remote monitoring.

Limitations and Future Directions

Limitations include the relatively small, demographically homogeneous cohort (primarily male, German-speaking), variable home recording environments, and short monitoring window. These may restrict generalizability across patient populations and languages. Feature drift and uncontrolled confounders (e.g., upper respiratory tract infection, device positioning) are potential sources of noise.

Future work should prioritize large-scale, real-world validation, adaptability to multiple languages, integration with multi-modal digital biomarkers (e.g., wearable activity, sleep, ECG), and prospective outcome trials. Deep learning architectures could further enhance feature representation, though explainability should remain central for trust and clinical translation. AI systems derived from this approach may generalize toward other cardiorespiratory and neurodegenerative pathologies.

Conclusion

This study establishes that time-series vocal biomarkers, as extracted from mobile-acquired daily voice data, offer higher prognostic accuracy for detecting incipient health deterioration in chronic heart failure than current standard-of-care home monitoring metrics. The identification of interpretable, physiologically grounded acoustic-spoken markers demonstrates substantial promise for scalable, non-invasive, patient-centric remote management—constituting a step change in digital health strategies for chronic disease (2604.00308).