Current asthma management strategies rely heavily on regular symptom and lung function monitoring to prevent exacerbations and optimize treatment [1]. Traditionally, spirometry and peak flow meter readings have been used to assess lung function; however, these require clinical visits or patient-operated devices (eg, peak flow meters), which may not be readily accessible or optimally used [2]. Furthermore, these methods provide only episodic snapshots of lung function, potentially missing the subtle day-to-day fluctuations that can precede an exacerbation. Research indicates that regular monitoring combined with asthma education are of primary importance in asthma management [3].

Advancements in digital health have opened up avenues for more continuous and patient-centered monitoring approaches. Among digital technologies, vocal biomarkers have emerged as noninvasive, accessible solutions [4-6] leveraging changes in voice characteristics that can reflect underlying pathophysiological changes, including respiratory conditions. With the ubiquity of smartphones and other voice-enabled devices, vocal biomarkers can be captured and analyzed for real-time feedback on respiratory status. For asthma applications, such research has focused on analysis of cough, breath, wheeze sounds [7-11], and speech [12-14], attempting to identify or characterize asthma with high sensitivity and specificity. However, these studies have historically faced limitations such as small sample sizes, brief study durations, retrospective or laboratory-based testing, and limited real-world validation of automated scoring systems [11]. Our study design advances the field in several of these areas (duration, use of a predesign and automated scoring algorithm, ecological validity, and transparent index test comparators).

This study aimed to address these shortcomings by building on a previously developed and validated respiratory-responsive vocal biomarker (RRVB) that operates on user devices in real-world settings [15]. The RRVB tool analyzes 6-second held vowel elicitations measured on smartphones to extract and interpret vocal features correlating with respiratory function. The RRVB has been shown to provide approximately 70% sensitivity and specificity in distinguishing diagnoses of asthma, chronic obstructive pulmonary disease (COPD), interstitial lung disease, persistent cough, and COVID-19 from healthy controls [15]. This performance was demonstrated in 2 independent studies from India and the United States including English and 5 Indian languages.

Our study linked the need for improved asthma management strategies with the potential of the RRVB to provide real-time respiratory function assessments. It piloted the RRVB’s utility as a reliable asthma monitoring tool by correlating RRVB scores with exacerbations. In addition, this study explored participant engagement with the RRVB tool and collected feedback on its usability and perceived benefits, thereby contributing to a broader understanding of vocal biomarker tools in asthma management. The hypothesis was that the RRVB scores could accurately differentiate between exacerbations and normal respiratory states in patients with asthma as defined by combined peak flow and symptom measures and that participants would report favorably on its potential as an asthma management tool. We combined peak flow (objective) and symptom (subjective) measures to ensure a robust comparison of RRVB scores to asthma state, reflecting the common clinical practice of assessing asthma holistically and recognizing that neither peak flow nor symptoms alone fully capture a patient’s condition.

The study app was downloaded and used on participant smartphones to collect voice samples and questionnaire responses and was compatible with iOS and Android devices. Selected screens illustrating the app user interface and flow are shown in Figure 1. Voice samples (6-second held [ɑ] vowels, as in father) were processed to generate RRVB scores, which were displayed to the user on a result screen in real time (5-10–second latency). Peak flow and subjective asthma reports were collected through questionnaires on the app. Participants were encouraged to use the app as part of their daily routines, with push notifications reminding them to use it. A study session was defined as the successful completion of 1 full interaction with the study app, including both voice samples and questionnaire responses. If the session was not completed (ie, if the participant exited the app before finishing the voice or questionnaire components), it was treated as incomplete and not recorded in the study database. Incomplete study sessions were not penalized, but persistent low completion rates could lead to participant withdrawal, as explained in the Study Site and Recruitment section.

Development and validation of the RRVB on asthma and other respiratory conditions included various cohorts in multiple geographies [15]. The RRVB model generated a composite score from 0 to 100 indicating the relative risk of respiratory symptoms. Using a threshold score of 65, the model differentiated patients with respiratory symptoms from healthy controls with approximately 70% sensitivity and specificity. In addition, 2 quality control algorithms operated on voice samples and score results to ensure robust and valid outcomes (described in the Elicitation Check and 2-Sample Scoring section).

The study design followed Standards for Reporting of Diagnostic Accuracy Studies guidelines [18] and was approved by the Institutional Review Board of the Albert Einstein College of Medicine (078142). It was retrospectively registered on ClinicalTrials.gov (NCT05850390) on April 28, 2023, after study initiation (December 14, 2021) but before completion of participant enrollment. Participants were informed about incentives, including US $25 for completing monthly ACT questionnaires and US $15 weekly if conducting ≥4 study app sessions, and that these were not contingent on specific outcomes. Study data were deidentified through the use of numeric participant identifiers that were used for all data collection and analysis. App data were stored in a secure cloud repository hosted on Amazon Web Services. Informed consent was provided electronically through the study app during the onboarding and training session in the presence of the research associate, who was available to answer study-related questions.

As illustrated in Figure 2, the enrollment cohort included all participants who began the study and was used for analyzing engagement and retention. The analysis cohort, defined as participants with at least 4 study sessions, was used to analyze correlations among the RRVB, respiratory state, and ACT. The normalized cohort, a subset of participants in the analysis cohort with at least 5 normal function sessions, enabled estimation of reference RRVB and normalized scores.

This study enrolled 84 participants (n=68, 81% in person and n=16, 19% remotely) from December 2021 to August 2023, forming the enrollment cohort (Figure 2). Of these 84 participants, 12 (14%) completed <4 sessions, leaving 72 (86%) in the analysis cohort. Of these 72 participants, 15 (21%) had <5 normal function sessions, resulting in 57 (79%) participants in the normalized cohort. The enrollment cohort included 23% (19/84) of participants who were withdrawn due to lack of engagement (data were retained for analysis), and additional participants were enrolled to increase data availability.

Figure 3 (left) shows mean RRVB scores across respiratory states, from normal function (feeling good and normal PEF; mean score ±S.E. 61.7±1.4) to mild events (either not feeling good or reduced PEF; mean score 64.8±0.99) to exacerbations (not feeling good and reduced PEF; mean score 68.0±1.4). As expected, the scores increased from normal function to exacerbation by 6.3 points. This pattern was also present for normalized RRVB scores, as indicated in Figure 3 (right), with more pronounced changes—mean normalized scores increased from −1.0±0.31 at normal function to 1.5±0.62 for mild events and 7.1±0.92 for exacerbations, for a total increase of 8.1 points. Mean reference RRVB scores varied minimally (<2.5 points) as a function of level of asthma control (data not shown).

The RRVB scores’ ability to differentiate respiratory states is shown in Table 4, indicating the RR for exacerbations as 2.15 (95% CI 1.62-2.85; P<.001). For mild events, the RR was 1.42 (95% CI 1.23-1.65; P<.001). These results confirm the effectiveness of RRVB scores in distinguishing respiratory states, especially for exacerbations.

Following normalization, RRVB scores were segmented into high- and low-risk categories using the calculated threshold of 3.7. Table 4 shows that the RR for exacerbations was 3.57 (95% CI 2.70-4.73; P<.001). For mild events, the RR was 1.31 (95% CI 1.13-1.53; P<.001). Normalization significantly improved risk stratification for exacerbations by reducing score variability unrelated to respiratory function. Normalization did not improve risk stratification for mild events.

Table 5 shows the RRVB score changes for respiratory state transition assessed within individuals. Most transitions involved participants remaining in the same state, whereas direct transitions from normal function to exacerbation or vice versa were uncommon. The data reveal that RRVB scores remained stable when respiratory states did not change. When transitions occurred, the magnitude of RRVB score change varied—approximately 3.5 points for shifts between normal function and mild events and 2 to 3 points for changes between mild events and exacerbations. Shifts between normal function and exacerbations resulted in an RRVB score change of approximately 6 points, suggesting a cumulative effect of worsening respiratory conditions on RRVB scores.

Of the 234 ACT assessments conducted during the study, 201 (85.9%) were eligible for correlation analysis with RRVB scores based on the availability of at least one RRVB result within 2 weeks before or after the ACT assessment date. As expected, RRVB scores increased as asthma control (ACT) decreased (Figure 4). RRVB score increase as a function of asthma control was less pronounced compared to RRVB score increases as a function of momentary respiratory state, in particular for normalized RRVB scores, which only increased by approximately 2 points from well-controlled to poorly controlled asthma. Thus, normalization may obscure underlying differences in asthma control apparent in unadjusted scores. RR estimates for poorly controlled or not well controlled asthma were 1.17 and 1.15, respectively, with CIs that included 1.0 (Table 6). This suggests insufficient statistical evidence to support the effectiveness of time-averaged RRVB scores in stratifying levels of asthma control.

Engagement and RRVB performance across demographic and clinical subgroups illustrate how the RRVB’s predictive accuracy varied within the cohort (Table S1 in Multimedia Appendix 1). The prevalence of exacerbations highlights differences in asthma characteristics within these subgroups.

Engagement differences were analyzed by comparing mean app session estimates and identifying significant effects through nonoverlapping CIs. Substantial effects were found for gender and age range. Lower engagement was observed in male individuals and participants aged <30 years, indicating challenges in consistently engaging these individuals over time. Ethnicity did not significantly affect engagement, indicating good potential to serve diverse populations. Lower engagement approached significance in Black or African American participants. Asthma disease severity (medical record) and asthma control (ACT) were not significantly associated with engagement.

An increasing number of comorbidities was positively correlated with engagement, particularly among participants with ≥4 diagnosed conditions, who engaged at significantly higher rates than those with 0 to 1 or 2 to 3 comorbidities. This trend was also observed for mental health diagnoses, where participants with conditions such as depression or anxiety demonstrated significantly higher engagement than those without these conditions. When analyzing specific comorbidities (not included in Table S1 in Multimedia Appendix 1), 7 out of 10 conditions surveyed were associated with significantly higher engagement. This may suggest that individuals with multiple or complex health conditions may be more motivated to engage with digital health tools, possibly due to a greater perceived need for consistent asthma monitoring.

Exacerbation detection performance, calculated via RRs and comparing raw versus normalized RRVB score analysis, provided several insights. Normalized RRVB scores reduced subgroup differences in RRs, eliminating the statistically significant differences for gender and age observed in the raw RRVB score analysis. This suggests that certain subgroup factors may correlate with RRVB scores independently of respiratory function. Normalized RRVB scores improved RR estimates in 17 of 24 subgroup variables, and where it did not, the CIs overlapped. Nearly all subgroups yielded statistically significant RRs in the normalized analysis except in cases in which low exacerbation prevalence led to wide CIs.

Table 7 shows the relationship between PEF zones and asthma well-being responses from sessions with valid RRVB scores in the analysis cohort. The table includes 10.71% (224/2091) of sessions classified as exacerbations (yellow-red PEF zones and not good–bad well-being), averaging 3.1 events per participant or 1.0 per participant per month. In addition, 61.45% (1285/2091) of the sessions were recorded under the green PEF zone and feeling good category, indicating normal function conditions. The remaining 27.83% (582/2091) of the sessions were classified as mild respiratory events. It should be noted that the actual prevalence of exacerbations may differ from these numbers as use of the study app was self-directed by participants.

This study advances the application of vocal biomarker technology in asthma management via a smartphone-based tool in a prospective real-world study. RRVB scores, derived from 6-second held [ɑ] vowel (as in father) voice recordings, provided immediate respiratory function feedback and significantly correlated with exacerbations, including in demographic and clinical subgroups. Exacerbation risk was more than twice as large in the high versus low RRVB score range, and this RR increased to approximately 3.5 using normalized RRVB scores. The tool also identified milder respiratory events, although with lower performance. Many participants remained engaged over the 3-month study, reporting favorable opinions and benefits. Unlike previous studies that have explored speech-based vocal biomarkers and home-based monitoring in limited capacities [11], our research uniquely integrated an existing vocal biomarker algorithm that provided real-time feedback to participants in real-world settings over an extended duration. This operationalization of voice-based exacerbation monitoring using everyday technology without requiring additional devices demonstrates the feasibility of patient self-monitoring and enhances asthma management accessibility. Building on the initial RRVB validation in cross-sectional studies [15], these results extend the RRVB’s use cases to longitudinal monitoring.

The RRVB normalization method successfully improved exacerbation detection and reduced the subgroup differences observed using raw RRVB scores. This suggests that the normalized scores are more accurate in capturing true respiratory dynamics versus the possible confounding impact of demographic, clinical, or device-related factors and highlights the suitability of longitudinal monitoring using the RRVB approach.

Participant feedback provided valuable insights into the usability and effectiveness of the RRVB tool. High satisfaction levels in the feedback questionnaire indicated that the core objectives—to assess the feasibility and user experience of a novel voice-based tool for asthma monitoring—were met positively. Participants appreciated the convenience; accessibility; and potential for regular, in-home monitoring provided by the app. This aligned with the study’s aim to offer a noninvasive and user-friendly approach to asthma management, with many preferring the vocal approach over a peak flow meter, likely due to its ease of use, consistent with studies comparing peak flow to symptom monitoring [2].

Both raw and normalized RRVB scores showed limited utility in reflecting self-reported asthma control assessed via ACT despite their effectiveness in stratifying momentary respiratory states. This discrepancy may stem from the static averaging method used, which fails to capture dynamic fluctuations in asthma symptoms. In previous studies, manual counting of cough frequency overnight or over 24-hour periods was demonstrated to correlate with asthma control, but this required extended recording intervals and manual approaches [8,9]. Future research could explore more advanced statistical modeling to analyze RRVB time-series data, potentially providing better insights into asthma control by accounting for temporal patterns and symptom variability.

The difference in the measures used—momentary exacerbations versus ACT—may also contribute to this discrepancy rather than indicating a failure of alignment. RRVB effectively differentiated exacerbations from normal function as it captured immediate respiratory function based on PEF and self-reported well-being, which participants reported in real time. In contrast, the ACT relies on a 4-week recall period, reflecting broader symptom trends and overall control rather than immediate states. This recall period can introduce bias that is not reflected in RRVB scores [20]. Clinicians often rely on standard measures such as the ACT or informal self-reports of past symptoms, lacking access to historical data that were captured in real time. This is an important distinction and highlights the complementary nature of longitudinal RRVB and ACT scores. Used together, they may offer a more complete picture of asthma status than either alone.

In addition to the RRVB scores, tools such as this one can capture a wealth of other structured data—such as symptoms, severity, triggers, and rescue inhaler use—that could further enrich the clinician’s insight into patient condition and behavior over time. The symptom and medication use patterns observed in this study mirror well-established asthma phenotypes and reinforce the ecological validity of the RRVB signal in real-world conditions. These continuous, patient-reported data offer a more detailed and reliable historical record compared to patients’ verbal recollection of symptoms and behaviors during clinical visits. It also enables linkage between momentary physiological signals (such as the RRVB) and behavioral indicators of disease management (such as rescue medication use), supporting more context-aware monitoring. The RRVB could fill a critical gap among self-perceived wellness (which can be delayed or imprecise), environmental exposures (which are frequent but not always predictive), and behavior (which may lag behind symptom onset). Such longitudinal, multilayered data open up avenues for artificial intelligence–driven analysis to identify key patterns, enabling more personalized and proactive patient care.

Achieving and sustaining high engagement is essential for health apps but often challenging. In this study, a promising 58% (49/84) of participants had medium to high engagement. These participants used the app on average 53 times over 3 months, 94% (46/49) used the app at least once in month 3, and 78% (36/49) remained consistently engaged (≥8 uses). These metrics compare favorably to those of many app-based studies, which often report sharp early drop-off and limited sustained use [21,22]. For example, Pratap et al [21] found a median retention of just 5.5 days across large-scale studies, and Amagai et al [22] noted that only approximately one-third of reviewed studies maintained majority engagement for their full duration. Our definition of high engagement (≥8 uses per month and activity into month 3) aligns with top-quartile patterns identified in these reviews and highlights the feasibility of RRVB use in sustained, real-world contexts.

Subgroup analysis (Table S1 in Multimedia Appendix 1) indicated lower engagement for male participants and those aged <30 years. In addition, there was a trend of lower engagement for those with mild asthma and Black or African American participants. These factors are consistent with reported asthma medication adherence issues, indicating potentially common underlying causes [23-25]. However, the 58% (49/84) of participants who were consistently engaged compare favorably to 14% to 16% of patients with satisfactory medication adherence in a 15-location pharmacy database study [23], highlighting a potential opportunity for the RRVB tool to encourage more proactive asthma management behaviors.

Comorbidities associated with higher engagement included COPD, diabetes, nonseasonal allergies, reflux disease, sleep apnea, skin conditions, and obesity. In contrast, participants with hypertension, seasonal allergies, and sinus conditions had similar engagement rates to those without these conditions. These engagement differences may reflect the chronic nature of conditions such as COPD, diabetes, and sleep apnea and their direct impact on respiratory function, potentially increasing the perceived need for continuous asthma monitoring. In contrast, conditions such as hypertension and seasonal allergies, which are often more episodic or managed via stable long-term treatments, may not drive the same level of engagement. In addition, mental health conditions such as depression and anxiety, which were also linked to higher engagement, may amplify health-related concerns, prompting more frequent use of monitoring tools. While this study was not designed to definitively determine these drivers of engagement, the results suggest that individuals with more complex or chronic comorbidities may be more motivated to actively manage their asthma through digital health tools.

This study’s selective participant group limits the generalizability of the results. While diverse, the participants may not reflect the broader population of patients with asthma, and the longitudinal nature may bias outcomes toward consistently engaged participants. All participants were required to have access to a smartphone and complete app-based sessions, which may exclude individuals with limited digital access or technological literacy, including older adults or those from more digitally disconnected segments of the population.

Despite these limitations, the RRVB tool’s demonstrated generalizability in 2 previous studies [15] and again on this new cohort represents a level of robustness rarely shown in the validation of machine learning–based tools. In particular, the RRVB tool demonstrated strong engagement and performance among Hispanic participants (Table S1 in Multimedia Appendix 1), suggesting its potential utility in traditionally underserved communities. This finding compares favorably to reported equity challenges in digital intervention effectiveness across diverse populations [26] and suggests that the RRVB can serve a broad range of patients when designed and deployed with accessibility in mind.

This study used PEF in combination with symptoms to define asthma state (normal function, mild event, or exacerbation) rather than as an outcome variable for respiratory function. While PEF is a convenient and commonly used tool in asthma management, it is effort dependent and can vary based on the patient’s performance during testing. In addition, PEF is not the gold standard for assessing airflow limitation, which may affect the accuracy of the respiratory state classification in our analysis. As a result, variations in PEF performance could lead to misclassification of asthma states, potentially impacting the RRVB comparison to true respiratory function.

While our study advances the field in important ways, including its prospective design, longer study duration, and real-world ecological implementation, some limitations remain. Compared to previous studies with smaller cohorts or retrospective analyses [11], our work demonstrates improved scale and automation. However, subgroup analyses (eg, by age, sex, or engagement levels) were limited in statistical power. In addition, while we used predefined symptom and PEF-based labels to reduce bias in state classification, further validation in broader and more diverse populations is warranted. Finally, while the RRVB tool successfully identified exacerbations and respiratory states in real time, it did not directly trigger clinical or self-management interventions. Participants were instructed not to use the RRVB scores for asthma decision-making during the study, and no automated alerts or care pathways were linked to RRVB outputs. As such, the tool currently functions as a monitoring aid rather than a stand-alone intervention system. Future integration of the RRVB into broader asthma management platforms with tailored feedback, alerts, or clinical decision support could improve its impact on timely intervention.

Because the RRVB tool was intentionally tested as a stand-alone application to assess its signal fidelity in real-world conditions, engagement and perceived utility might have been limited. Without integration into a comprehensive asthma management system, participants lacked additional functionalities that could enhance effectiveness, such as personalized asthma education or medication tracking. This limited context likely affected use patterns and participants’ assessment of the tool’s utility. Indeed, feedback survey responses indicated a need for more comprehensive guidance on score interpretation and actionable insights such as advice on using rescue medication or contacting health care providers. The feedback suggests that integrating the RRVB within broader asthma management platforms could enhance user engagement and provide valuable guidance. Linking RRVB results to tailored asthma educational content, as explored in ongoing studies integrating the RRVB into the ASTHMAXcel app [27,28], was also suggested by participants.

Feedback was obtained from 55% (46/84) of the enrollment cohort, primarily from more engaged participants. This may skew satisfaction and usability ratings toward a more positive evaluation, potentially overestimating the tool’s effectiveness. That said, satisfaction levels varied minimally by engagement level (separate analysis; data not shown).

Existing digital asthma management platforms offer functionalities such as symptom tracking, medication management, peak flow and spirometry tracking, personalized asthma action plans, and access to educational resources [29]. Integrating advanced technologies such as the RRVB would represent a significant advancement in noninvasive, frequent monitoring of respiratory health. As digital platforms evolve, vocal biomarker technology is expected to offer a more proactive, personalized, and comprehensive approach to asthma management.

The RRVB’s ability to detect changes in respiratory status holds promise for improving asthma care by enabling a level of vigilance in disease management that is scalable, user-centric, and far beyond current capabilities. Frequent use of the RRVB could enhance patients’ perception of airflow limitation and medication adherence, as demonstrated with interventions using PEF prediction and feedback [30,31]. This suggests that the RRVB could help enhance patients’ awareness and management of their condition facilitated by the high engagement and positive reception demonstrated in this study. As a momentary signal that complements traditional tools such as the ACT, the RRVB could serve as a real-time input to future care models—flagging early changes in respiratory status that might prompt clinician review, self-management action, or digital triage. Practically, this could mean reduced frequency and severity of asthma attacks, better day-to-day symptom management, and potentially fewer urgent or emergency care visits and hospitalizations.

This study evaluated the real-world utility of an RRVB for detecting asthma exacerbations using smartphone-based voice recordings in participant home settings. Through a 3-month prospective cohort design with daily app-based data collection, RRVB scores were shown to effectively stratify respiratory states and predict exacerbations with statistically significant RRs. Engagement was sustained in most participants, and user feedback indicated strong usability despite some noted areas for improvement. These results support the RRVB as a promising tool for continuous asthma monitoring and lay the foundation for future integration into more comprehensive self-management systems, including a potential role in proactive health tracking or patient-clinician communication.