A Pilot Study on AI-Based Voice Analysis for Monitoring Patients Hospitalized with Acute Decompensated Heart Failure

Figures

• 

  Figure 1. Discrimination of admission and discharge states using vowel- and breathing-based voice analysis. (A) Receiver operating characteristic (ROC) curves for the best-performing model (vowel /i/ + Breathing segments) on all patients (N = 77) depicting the ability of the model to distinguish between admission and discharge samples.

• 

  Figure 2. Generalization performance of the speech-based model to unseen intermediate hospitalization days using a leave-one-patient-out evaluation on all patients. (A) Mean transformed patient-level prediction scores (sigmoid z-scores) across patient-specific normalized hospitalization duration (0 = admission, 1 = discharge), smoothed with a Gaussian filter. The prediction scores continuously increase throughout the hospitalization stay. Shaded area represents ±1 standard deviation across patients. (B) Confusion matrix comparing true vs. predicted hospitalization days for held-out patients. The model was trained exclusively on recordings from admission and discharge days and then evaluated on intermediate days for the left-out patient (N = 77).

• 

  Figure 3. Respiratory acoustic features comparing hospital admission and discharge recordings. (A) Mean inspiratory airflow dynamics at admission (orange) and discharge (teal). Solid lines depict the mean linear slope of inspiratory airflow rise, defined as the interval during inspiration where the amplitude envelope increases from 10% to 90% of its peak. Dashed lines represent the mean envelope amplitude across all patients at each timepoint, normalized to the individual recording’s amplitude range. Shaded regions illustrate ±1 standard deviation (see Supplementary material online, Appendix A3). (B) Mean total breathing duration (seconds) derived from standardized poem recordings at admission and discharge. (C) Mean breathing rate (number of breaths per minute) derived from the same recordings at admission and discharge.

• 

  Figure 4. Sensitivity analysis—effect of used vowel: (A–D) box plots show the normalized model-predicted scores at admission (orange) and discharge (teal) across all patients (all vowels: N = 79, vowel/i/: N = 77, vowel/o/: N = 75 and vowel/a/: N = 79). Per patient scores are computed as an average of all daily recordings. Statistical significance between admission and discharge distributions as well as F1-score for the models is indicated.