Dr. David P. Kao is a cardiologist connected with UCHealth University of Colorado Hospital in Aurora, Colorado. He graduated from Johns Hopkins University School of Medicine with a medical degree. In this video Dr. Kao discusses extracting Vocal Biomarkers for Pulmonary Congestion With a Smartphone App.
Link to the abstract
In patients with acute decompensated heart failure (ADHF), pulmonary congestion is the most common reason for admission, and chronic pulmonary congestion predicts poor post-hospitalization outcomes (1). Tracking fluid retention over the course of HF could help to avoid ADHF hospitalization and enhance inpatient care quality. However, objectively assessing congestion in a cost-effective, regular, and accurate manner remains difficult.
Machine learning (ML) has sped up "nonclinical" adoption of automated and functional speech recognition systems, but it's still unclear whether these systems can be used to track physiologic factors like lung congestion for clinical application. A number of studies have made use of the smartphone, which is a ubiquitous and noninvasive instrument for collecting patient data, and studies have also revealed a link between fluid retention and vocal cord vibration (2,3). Amir et al. (4) investigate the idea of utilizing a smartphone app to detect pulmonary congestion using voice analysis. In this editorial, we examine the significance of this study for HF management, potential alternative techniques, and some cautionary notes on equity in the use of such technologies with vulnerable groups.
Vocal Biomarker Tracking and Alerting Challenges
Because the spectrum of congestion is lower dimensional, it may be easier for doctors to evaluate, understand, and intervene on in a focused manner than a global indicator of HF condition such as New York Heart Association functional class (3). According to Amir and et al (4), the HearO software requires a patient to read specified sentences into their smartphone. The app reports disparities between a patient's present speech features and those in a reference state, such as euvolemia, using a set of fixed analytic measurements. The authors reveal that the app's voice analysis findings differ significantly between hospital admission and discharge.
This strategy, however, has a number of significant drawbacks. For the app to be useful, patients must, for example, engage with it and read the relevant sentence. Many nonpharmacologic self-care advice for people with HF are inconsistent, suggesting that involvement with a system like the HearO app may be unsatisfactory.
Another significant drawback of the HearO app is that it may be difficult to achieve genuine euvolemia during hospitalization. ADHF can show in a variety of ways, and lung congestion may not be the best clinical reference point in many cases (1). Given these considerations, a more holistic monitoring approach, such as leveraging always-on virtual assistants like Alexa or Siri to personalize clinical warnings, may be more beneficial. However, developing and training efficient ML models would necessitate vast, rich data sets (5), and the use of passive, continuous listening, of course, poses further privacy problems. For example, there are still difficulties in limiting language and dialect prejudice, therefore the use of smartphone technology in the clinic necessitates careful planning during development.
Expanding the Clinical Applications of Vocal Biomarker Tracking
Individuals with ADHF frequently appear with identical symptoms in clinical practice. Speech alterations related with ADHF could be learned over time for a single patient, thereby improving alarm accuracy and advance notification timing. Instead than depending on speech metrics to identify congestion just at admission or discharge, as the HearO app does, speech measures might be made more granular and diverse. By measuring distinct symptom kinds, phrase length and actual word content, for example, could be used to enhance vocal cord vibration. However, whereas a single worrisome signal, such as vocal cord vibration, may be as easy as triggering a formal clinical assessment, an abundance of concerning signals may complicate already complicated clinical operations. If smartphone-based ADHF management, such as the HearO app, is to be scaled, it will be necessary to determine what steps to take based on the nature of the warning signal in order to allow for cheap, frequent, and accurate volume assessments on a specific schedule and on an ad hoc basis without compromising quality of care or contributing to provider burnout.
A Word of Caution on Technology in Medicine
Bias is a problem in artificial intelligence, which has just recently received a lot of attention (3,6). Algorithms rely on developers to make modeling and data decisions, which both enable for bias to be introduced and systemic prejudice to be reinforced. Fairness, operational tolerances, and unfavorable effects on vulnerable populations should all be carefully considered while developing software, especially in safety-critical contexts.
Voice biomarker detection can be hampered by a variety of biases, including representational, measurement, omitted variable, algorithmic, and social biases (6). Any patient-monitoring device should assure demographic parity, equalized odds, equal opportunity, treatment equality, and fairness, in addition to standard efficacy standards. Because speech recording quality could be stratified according to socioeconomic level, the HearO app emphasizes how limited access to modern smartphones (e.g., the latest iPhone or Android devices) may effect future quality of care. As additional sectors of medicine use these devices to collect data, the HF community should consider how discrepancies in data could harm our patients.
Dr. Amir et al. (4) have highlighted active speech analysis as a significant step forward in increasing the techniques available to assess patients with HF. Although still in its infancy, the use of widely available mobile technologies shows that it has the potential to be widely adopted, as opposed to extremely invasive tactics that require dedicated gear. Before clinical application, more research and validation is required, but success in a use case like HearO could open the way for even more convenient and generalizable solutions.
Author Disclosures and Funding Support
Dr. Kao has served as a consultant to Codex Health, Inc. Dr. Ravindra has indicated that he has no ties to disclose that are relevant to the content of this study.
The authors certify that they complied with the authors' institutions' human research committees and animal welfare rules, as well as Food and Drug Administration procedures, including patient permission where relevant. Visit the Author Center for further information.