跳到正文
原文
Nature 健康护理主题·· 15 天前AI 评分55

评估用于检测胃轻瘫患者呕吐和干呕的音频监测系统:一项观察性队列研究

Evaluation of an audio-based monitoring system for detecting retching and vomiting in gastroparesis: an observational cohort study

AI 导读

研究团队评估了一种音频监测系统在胃轻瘫患者中的应用,结果显示该系统能够有效监测呕吐和干呕事件,且具有较高的患者依从性。

正文

Introduction

Gastroparesis (GP) is a chronic disorder characterized by delayed gastric emptying in the absence of mechanical obstruction1. Patients with GP commonly experience symptoms such as nausea, vomiting, and retching2, which can significantly impair quality of life3. However, these episodic symptoms vary in frequency and severity over time, posing challenges for accurate monitoring and effective management. Traditional methods for assessing vomiting and retching rely heavily on patient self-report via daily diaries4 or structured questionnaires with longer recall periods5. Although these tools provide valuable insights into patients’ subjective experiences, they are prone to recall and reporting biases6, presenting significant limitations for evaluating treatment effects in clinical trials and for monitoring clinical outcomes in both healthcare and home settings. Therefore, there is a critical need for objective, continuous, and reliable methods to capture vomiting-related symptom burden in patients with GP.

Emerging digital health technologies, particularly those leveraging passive sensing and artificial intelligence, present promising opportunities for objective, low-burden, and reliable symptom monitoring in ambulatory settings. Among these, audio-based monitoring systems have demonstrated feasibility in detecting respiratory and vocal biomarkers across various domains7. However, their application to gastrointestinal symptoms of retching and vomiting remains largely unexplored. Given the distinct acoustic signatures of retching and vomiting events, we hypothesized that audio analysis could provide a non-invasive, real-time solution for objectively monitoring vomiting patterns.

Here, we present an exploratory evaluation of an audio-based system designed to detect and quantify retching and vomiting (expulsion) events in a non-interventional pilot study involving participants with GP. This study aimed to differentiate auditory profiles of “retch” events (i.e., those without expelled material) from “expulsion” events (i.e., those involving the ejection of material). In addition, this study sought to evaluate the feasibility of audio-based symptom monitoring, with device compliance defined as the percentage of expected recording time successfully captured over the 14-day study period (calculated as recorded h divided by the total expected h [24 × 14]). The study also aimed to assess the reliability and validity of retch and expulsion event counts derived from audio recordings and to characterize the temporal patterns of these events in individuals with GP.

Results

Participant characteristics

This decentralized, observational study included 25 adult participants with nausea-vomiting gastroparesis. Baseline participant characteristics are summarized in Table 1. In line with the observational study design, baseline data focused on demographics and established symptom history. Throughout the study duration, there were no reported adverse events (AEs).

Table 1 Participants Demographics

Full size table

Device compliance

During the recruitment period, 28 GP participants were contacted. Of these, none declined participation due to privacy concerns around home monitoring requirements; however, n = 1 participant experienced symptoms that were too sparse and sporadic, n = 1 participant had no wireless internet access, and n = 1 participant had issues with the nondisclosure agreement.

Device compliance is defined as the percentage of expected recording time successfully captured over the 14-day study period. Compliance was higher, on average, for the iOS device (mean 20.66 ± 7.48 h/day; device compliance = 83.4%, N = 23) compared to the Watch (mean 17.24 ± 7.29 h/day; device compliance = 69.4%, N = 24). Participant engagement remained stable across the study cohort, with no significant longitudinal decline in daily recording hours for either platform (Supplementary Fig. 1). One participant was excluded from the Watch analyses and two were excluded from the iOS analyses due to technical setup issues; however, no participant experienced issues with both devices simultaneously (see individual participants’ device compliance in Supplementary Fig. 2).

Descriptive statistics: event counts

Table 2 lists all the endpoints collected and computed in this study. From the continuous audio data recorded remotely via the Watch and iOS devices, manual annotation by trained human raters showed that the Watch captured a total of 870 retch events and 275 expulsion events, while the iOS device recorded 796 retch events and 227 expulsion events. Using the web-based questionnaire to track symptoms via electronic patient-reported outcome (ePRO) counts, patients reported a total of 575 retch events and 247 expulsions.

Table 2 Summary of tabulated endpoints

Full size table

In addition to daily electronic patient-reported outcome (ePRO) event counts, participants also completed two 5-point rating scales: Patient Global Impression of Severity (PGI-S: None[0], Mild1, Moderate2, Severe3, and Very Severe4) and Patient Global Impression of Frequency (PGI-F: None[0], Rarely1, Sometimes2, Often3, and Very Often4). Figure 1 presents stacked bar charts visualizing daily item-level responses and longitudinal variability in the PGI-S and PGI-F scores for GP participants. For both scales, the distribution remained relatively consistent throughout the study period.

Fig. 1: Daily distribution of patient global impression responses across a 14-day tracking period.

Stacked bar plots illustrating the number of participants reporting each response category per day, where grey bars represent missing data, light blue bars represent none, orange bars represent rarely for frequency and mild for severity, yellow bars represent sometimes for frequency and moderate for severity, red bars represent often for frequency and severe for severity, and dark red bars represent very often for frequency and very severe for severity: a PGIF; b PGIS.

Daily and weekly event counts for retching (R), expulsion (E), and combined events (R + E), collected from multiple sources across all available days, are summarized in Table 3. A consistent pattern emerged across all timeframes, with the Watch capturing the highest mean number of events of each type, followed by the iOS device and then the ePRO. The audio-based devices demonstrated that retching events were typically 2–4 times more frequent than expulsion events, a pattern similarly observed in ePRO counts. Figure 2 illustrates the distribution of daily retch and expulsion counts by data source, highlighting the considerable variability both within and across patients over days.

Fig. 2: Daily retch and expulsion counts across different data sources.

Boxplots of daily retch and expulsion counts over the 14-day study period grouped by data source, where blue bars represent iOS data, cyan bars represent Watch data, and red bars represent ePRO data: a iOS; b Watch; c ePRO.

Table 3 Daily and weekly retch, expulsion and combined retch and expulsion events by data source for GP participants over entire study period

Full size table

Beyond raw counts, event bouts were defined as clusters of combined retching and expulsion events occurring in close temporal proximity to avoid splitting of physiologically continuous symptoms, as illustrated in Fig. 3. Specifically, if the time between the end of one event and the start of the next was 10 s or less, the events were grouped into a single bout. Event bout endpoints are only defined for data collected via audio devices, given that ePRO lacks a precise timestamp of events to allow for similar analyses.

Fig. 3: Demonstration of a retch and expulsion bout.

Representative example illustrating the temporal sequence of a single retch and expulsion event recorded during the monitoring period.

Bout events (combined R + E) collected via the Watch and iOS device are shown in Table 4. We found that bouts typically consist of 1–8 events per bout (10th–90th percentiles), and last 0.81–25.15 seconds (10th–90th percentiles). As with individual events, bout counts were consistently higher on the Watch compared to the iOS device. Across subjects, daily R + E bouts were 1.61 (SD 1.96; 10th–90th percentile 0–4) for Watch and 1.00 (1.25; 0–3) for iOS; the corresponding weekly R + E bouts were 11.28 (13.73; 0–27) and 7.02 (8.77; 0–18), respectively.

Table 4 Descriptive statistics for bout events (combined R + E) by data source for GP participants over entire study period

Full size table

Descriptive statistics: time of day effects on event counts

Figure 4 shows the analysis of the timestamp data of audio-derived events, presenting the median distribution and 10th–90th percentiles across 3 h intervals. A prominent increase in retch and expulsion events was observed during the morning h, peaking sharply between 09:00 AM and 11:59 AM. Additionally, retch and expulsion events persisted throughout the night across multiple intervals (especially in 03:00 AM–05:59 AM) at levels comparable to those observed during the daytime.

Fig. 4: Distribution of retch and expulsion events across time of day, presented in three-hour intervals.

Plots show the median distribution of events and error bars representing the 10th to 90th percentiles: a iOS; b Watch.

Human-annotated event counts vs. patient-reported event counts

We present the correlation between audio devices and ePRO with corresponding sample size and 95% CI in Table 5. The two audio devices demonstrated strong agreement with each other. Specifically, the correlation between the Watch and iOS devices was 0.86 (95% CI: 0.82–0.89) for retching, 0.91 (95% CI: 0.89–0.93) for expulsions, and 0.90 (95% CI: 0.87–0.92) for combined R + E events. In contrast, correlations between device-based and ePRO event counts were limited. For retching events, the correlation of ePRO counts with the iOS device and the Watch was 0.26 (95% CI: 0.14–0.37) and 0.40 (95% CI: 0.29–0.50), respectively. For expulsion events, both the iOS app and the Watch showed similar correlations with the ePRO (r = 0.68; 95% CI, 0.61–0.74 for both). For combined R + E events, the correlation of ePRO counts with the iOS device and Watch was 0.51 (95% CI: 0.42–0.60) and 0.60 (95% CI: 0.51–0.67), respectively.

Table 5 Repeated Measures Correlation (rmcorr) Estimates Between Audio Devices and ePRO

Full size table

When using bout-based combined event counts, correlations with ePRO improved. The correlation between ePRO and the iOS device was 0.58 (95% CI: 0.49–0.65), while the correlation between ePRO and the Watch was 0.60 (95% CI: 0.52–0.67).

Human-annotated audio and ePRO event counts vs. PGI-S and PGI-F

We hypothesized that daily expulsion events would show stronger correlations with the PGI scales than retching events, as the PGI items specifically reference vomiting. We also anticipated that ePRO daily event counts would show stronger correlations with PGI-F/PGI-S scores than device event counts would, given that both are based on patient self-report. Expulsion counts from ePRO showed the highest correlations with the patient global impression items (PGI-F: r = 0.65 and PGI-S: r = 0.65), followed by device-based counts (r ≈ 0.31–0.36). Correlations of retch counts with the patient global impression items were weaker across all sources (r ≈ 0.16–0.36). As expected, combined R + E raw counts showed moderate correlations with the PGI-F/PGI-S, with ePRO again outperforming both audio devices. Likewise, R + E bout counts from Watch and iOS device offered similar correlation strength to R + E raw counts. The corresponding sample size and 95% CI for all correlation coefficients are provided in Supplementary Table 1.

Bar plots illustrating R (daily), E (daily), and R + E (daily) counts against the corresponding PGI-F and PGI-S reported on the same day are presented in Fig. 5. For retching, ePRO event counts remained consistent across PGI-F levels, with underreporting of event counts observed for patients who rated themselves as having vomiting “Very Often” or “Often” on the PGI-F. In contrast, both the Watch and iOS device showed a relatively monotonic increase in retching counts with increasing PGI-F levels. For expulsion events, low counts across all sources were observed within the “None” PGI-F category. Expulsion counts from the Watch, iOS device and ePRO increased somewhat progressively with higher PGI-F levels, though variability was high due to day-to-day fluctuations. Since participants experienced more frequent retching than expulsion events, the combined R + E (daily) plot closely mirrored the trend observed for retching.

Fig. 5: Daily retching, expulsion, and combined retching and expulsion event counts across patient global impression levels.

Daily retching, expulsion, and combined retching and expulsion events per day across varying patient impression levels recorded via digital health devices, where red bars represent ePRO, blue bars represent iOS, cyan bars represent Watch, and error bars represent 95% confidence intervals: a Retching events per day across PGI-F levels; b Expulsion events per day across PGI-F levels; c Combined retching and expulsion events per day across PGI-F levels; d Retching events per day across PGI-S levels; e Expulsion events per day across PGI-S levels; f Combined retching and expulsion events per day across PGI-S levels.

Figure 6 presents bar plots of the R + E bout counts (daily) from Watch and iOS device against their reported PGI-F and PGI-S scores on the same day. Notably, the bout-based counts still showed a monotonic and clearer increase with rising PGI-F/PGI-S levels compared to the ePRO-based counts.

Fig. 6: Daily retch and expulsion bout counts across patient global impression levels.

Daily retch and expulsion bout counts across varying patient impression categories recorded via digital health devices, where cyan bars represent data from Watch and blue bars represent data from iOS, and error bars represent 95% confidence intervals: a PGI-F, b PGI-S.

The various device and patient-reported event counts failed to reliably distinguish between adjacent PGI-F groups (e.g., “Often” vs. “Very Often”) and PGI-S groups (e.g., “Severe” vs. “Very Severe”) based on pairwise comparisons. To address this problem, PGI-F responses were collapsed from five categories into three categories as follows: None, Rarely/Sometimes, and Often/Very Often. Similarly, PGI-S responses were grouped into None, Mild/Moderate, and Severe/Very Severe. As shown in Supplementary Table 2, ePRO counts had limited ability to differentiate between these combined categories, whereas the Watch and iOS devices showed clearer distinctions, with expulsion counts increasing consistently with increasing frequency and severity.

Outcome reliability and effect of averaging period

Supplementary Table 3 summarizes the intraclass correlation coefficients (ICCs) and their corresponding sample size and 95% CIs for daily retching, expulsion, and combined (R + E) counts across three day-pair comparisons: Day 1 vs. Day 8, Day 1 vs. Day 14, and Day 7 vs. Day 14. Reliability varied across devices and event types. For the Watch, ICCs for R, E and R + E counts ranged from 0.37 to 0.80, while the iOS device ranged from 0.22 to 0.67. ePRO counts showed more stable reliability, with ICCs ranging between 0.40 and 0.56 across different day-pair comparisons.

Weekly ICCs (Week 1 vs. Week 2) and their corresponding sample size and 95% CIs are summarized in Table 6. The Watch showed higher reliability for retching events (ICC = 0.53; 95% CI: 0.00–0.81) than for expulsion (ICC = 0.16; 95% CI: 0.00–0.44) and R + E events (ICC = 0.31; 95% CI: 0.00–0.67). The iOS device demonstrated stronger reliability for retching (ICC = 0.79; 95% CI: 0.30-0.95) and R + E events (ICC = 0.69; 95% CI: 0.12–0.92), compared to expulsion events (ICC = 0.36; 95% CI: 0.00–0.79). The ePRO had the most consistent performance, with values ranging from 0.76 to 0.90 across all event types: R (ICC = 0.90; 95% CI: 0.74–0.96), E (ICC = 0.76; 95% CI: 0.45–0.91), and R + E (ICC = 0.89; 95% CI: 0.75–0.96). These findings are visualized in Supplementary Fig. 2.

Table 6 Test-retest reliability of weekly retch, expulsion, and combined retch and expulsion events in GP participants

Full size table

Bout-based R + E counts did not consistently improve reliability. Daily ICCs ranged from 0.25 to 0.72 for the Watch and iOS, depending on the day pair as detailed in Supplementary Table 3 along with their respective sample size and 95% CIs. Weekly ICCs remained limited, with the Watch at 0.54 (95% CI: 0.00–0.84) and the iOS at 0.66 (95% CI: 0.02-0.92).

Discussion

This study provides foundational evidence supporting the use of audio-based devices for objectively monitoring retching and expulsion events in GP, a technology which may be applied to other disorders characterized by episodic vomiting. While these results establish proof-of-concept for audio-based monitoring within a home setting, they do not reflect the full complexities of naturalistic, 24-hour monitoring; thus, our findings should be interpreted as foundational evidence requiring further validation in more expansive, real-world environments. Our findings establish that audio-based monitoring is both feasible and effective, with objective data collected from devices complementing and building on the information gathered currently using PRO tools.

A key finding of this study was the high compliance and willingness of participants to be audio-recorded, indicating that audio-based symptom tracking is acceptable within this population. Notably, compliance was higher with the iOS device, likely due to its passive nature—once activated, it required minimal participant involvement. Although audio monitoring devices remain relatively uncommon in clinical studies, our observed compliance rates of 69.4% from Watch and 83.4% from iOS demonstrate strong feasibility and patient acceptability. These rates are highly comparable to recent literature; for example, studies deploying audio-enabled smartwatches have reported median daily wear times of 19.8 h (~82% of a 24 h period)8. It is especially notable that our participants achieved these robust compliance levels given that continuous at-home audio recording is inherently more demanding in terms of privacy. Furthermore, our results reinforce prior research demonstrating that individuals are willing to be recorded for extended periods, maintaining excellent compliance across 10 to 11 days of continuous ambulatory monitoring9. Our study tested two recording modalities (fixed at-home and mobile on-wrist) showing advantages and disadvantages of each. As noted above, recording times on the iOS device were longer, linked to the higher compliance. On the other hand, the Watch captured more events (despite shorter recording times), likely due to its on-wrist placement and greater mobility within home when worn. This suggests that while fixed systems may be logistically simpler, mobile wearable recorders may be better suited for tracking symptoms in real-world environments.

This study further highlights the complementary roles of ePRO and audio-based measures in clinical research. The correlations observed between ePRO and audio-based measures of vomiting and retching events were moderate and consistent with expectations for convergent validity, where the goal is to demonstrate a reasonable degree of association rather than equivalence. This is expected, as PROs reflect patients’ perception and recall of symptoms, whereas audio-based measures capture discrete physiological events. Similar moderate associations have been reported in related areas, such as objective cough monitoring, supporting the interpretation that subjective and objective measures capture complementary aspects of symptom experience10,11.

Additionally, the ePRO counts demonstrated stronger within-person correlations with patient global impression items (PGI-S and PGI-F), indicating greater sensitivity to perceived intra-individual changes over time. In contrast, audio-based measures showed stronger discriminative performance in distinguishing severity and frequency levels based on the global impression items, despite modest within-person correlations. In addition, audio devices are limited to capturing specific symptom events and cannot assess other important patient experiences, such as nausea. A multiple-endpoint strategy incorporating both measures may thus provide a more holistic, patient-centered understanding of clinical benefit, capturing distinct treatment effects aligned with regulatory expectations.

The data collected in this study also provided new insights into the symptoms of gastroparesis. For example, recordings from the Watch and iOS device suggest that retching may be a more common symptom of GP than previously recognized. This challenges traditional views of the disease and highlights the need for further research into the impact of frequent retching on quality of life. Analysis of time-of-day patterns, while limited due to imperfect knowledge of participant location and behavior, suggest the presence of frequent nighttime retching, which may contribute to sleep disruption—an area often overlooked in traditional symptom diaries. Morning symptoms were also higher than other times of the day, potentially related to morning movement or eating. The mid-day dip in symptoms data should, however, be interpreted with caution, as subjects may be more likely to be outside their home during this time, away from the iOS device and with the Watch recording paused. Future studies should confirm and further explore these time-of-day patterns, by potentially incorporating physical movement tracking and monitoring of mealtimes to determine links to GP symptoms.

There was substantial symptom variation at the daily level, which likely represents real day-to-day changes rather than measurement error and reflects the dynamic and episodic nature of GP symptoms. Accordingly, ICC estimates in this setting should be interpreted in light of this variability, as they may be attenuated by true within-person fluctuations. However, weekly aggregation of the same data (i.e., week-to-week test-retest reliability) marginally improved repeatability. Due to the length of the observation period, longer aggregation intervals (e.g., over 2 weeks or more) could not be assessed in this study but are expected to further stabilize the measures; future studies could test this hypothesis. Similar findings are seen in studies of chronic cough, where high day-to-day variability leads to a need for longer observation periods12,13. Less variability was observed in the daily ePRO data compared to audio-based metrics, perhaps reflecting an implicit averaging in participants’ recall of events.

This work explored the definition and utility of “retch and/or expulsion bouts”, inspired by similar investigations in the cough literature14. As described in “Methods”, we developed a data-driven strategy that allows for the identification of bouts that are both sufficiently frequent and of an appropriate length, seeking to capture disease burden while avoiding both excessive fragmentation and oversimplification of the event data. The use of bouts is conceptually supported by the observation that event counts were higher in raw audio than in diary, suggesting participants may have been reporting bouts rather than individual events. We found that bout-level audio data showed stronger correlations with ePRO data than raw count-level audio data, suggesting that participants may naturally perceive and report their symptoms in discrete bouts. Additionally, bouts allow computation of measures of bout severity (for example number of events per bout or bout length). However, aggregating individual events into bouts did not improve test-retest reliability.

Our work has important limitations that can be addressed in future studies. Although four raters participated in the annotation process, each recording was scored by only one individual, precluding the assessment of formal inter-rater reliability (IRR) metrics. This design decision was driven by the volume of the dataset, which totaled ~16,800 h of continuous audio (25 participants with 2 devices over 14 days). Given the significant cognitive burden and time intensity required to achieve high-fidelity manual scoring at this scale, the study design did not include overlapping evaluations of the same recordings. To ensure data integrity, we employed the four-step quality assurance framework described in the “Methods” section, which systematically cross-referenced all annotation summaries against contemporaneous clinical patient notes. While these measures provided robust internal verification, future work may explore targeted, nested IRR assessments on smaller, randomly sampled segments to further quantify inter-observer agreement.

To protect participant privacy, the study restricted audio recording to the home with visual reminders, allowed participants to manually pause the devices at their discretion, and utilized automated software to remove human speech before manual review. While these safeguards were highly acceptable to participants, they introduced inevitable data limitations. For instance, the restriction to home-based recording likely resulted in an undercounting of total events and created challenges in interpreting time-of-day findings. Our findings about time-of-day distribution of symptoms, while suggestive, are limited by the fact we do not know when subjects were outside the home and thus presumably not recorded. Additionally, our reliance on human scoring poses logistical challenges for scaling to larger populations. Moreover, placebo effects can be sizeable in self-reported gastrointestinal measures15. Objective, passively captured measures, such as audio-based symptom tracking may reduce placebo effects by reducing expectancy-driven responses inherent in subjective measures16. While this could not be evaluated in the non-interventional pilot study presented here, future interventional trials in GP could assess whether the use of objective monitoring of the symptoms could be valuable in settings where placebo effects are a concern.

The development of a wrist-worn, fully automated system could enable privacy-preserving monitoring both inside and outside the home and across larger populations. Transitioning from this home-based pilot to a 24-hour naturalistic assessment will require overcoming significant technical and logistical barriers, including the need for on-device, privacy-preserving AI processing to eliminate the limitations of human-annotated data collected solely in domestic environments. Although full automation is technically challenging given that retch and expulsion sounds constitute only a small fraction of a day’s recorded acoustic data, we anticipate that advancements in large audio datasets and AI will make such systems feasible. Additionally, modern smart watches capture physiological signals (e.g., acceleration, heart-rate data) which could enhance AI algorithms with additional inputs.

In conclusion, this exploratory study supports the feasibility of audio-based detection of retch and expulsion events in patients with gastroparesis and provides preliminary evidence of convergent validity. Reliability varied by endpoint and device, highlighting the need for further validation and automation before broader clinical-trial deployment.

Methods

Ethics approval and consent to participate

The study protocol was reviewed and approved by the Core Human Factors Institutional Review Board (Core Human Factors IRB; IORG0007854; IRB Registration #IRB00009432) via expedited review. All participants provided written informed consent prior to participation using the IRB-approved consent form. The study was conducted in accordance with applicable ethical guidelines for research involving human participants and the principles of the Declaration of Helsinki. Clinical trial number: not applicable.

Participants

A total of 25 participants with GP were recruited for this study through private clinical sites specializing in the treatment of GP patients that are not associated with any academic groups. Eligible adult patients were identified from clinical records by site investigators and were required to have a physician-established primary diagnosis of chronic vomiting associated with gastroparesis, and at screening, to have recalled vomiting at least one day per week for the 4 weeks prior to screening. Diagnoses were established by treating clinicians following standard clinical practice, as the study’s primary objective was to evaluate the real-world feasibility of audio-based symptom detection with wearable devices in a home setting. Prospective repeat gastric emptying scintigraphy (GES) was not mandated as a protocol-specific entry criterion. Exclusion criteria included pregnancy or current hospitalization at the time of consent, as well as significant developmental delay, intellectual disability, or other mental impairments. To maintain a real-world, non-interventional study environment, this protocol did not require restrictions on patient medications, supplements, diet, fluid intake, or daily activities, and such data were not systematically captured.

Study design and setting

A non-interventional, fully decentralized 14-day study was conducted to quantify the number of retching and expulsion events in participants with chronic, recurring, and episodic vomiting associated with GP. Upon enrollment, participants provided informed consent and were shipped a study kit that included two study watches (needed to swap for 14 days of continuous recording) and one iOS device (iPhone). Participants received training via teleconference upon receipt of the study kit and were given the opportunity to ask questions to ensure proper use of the devices. All data were collected remotely from participants’ home settings, allowing for real-world monitoring of symptoms while protecting participant privacy through home-based recording, participant-controlled pausing of recordings, and automated removal of speech from audio files before manual review.

Data collection

Audio data were recorded using two types of devices across 14 days. Participants were instructed to wear an audio recording Watch (Esonic, SR-600 Voice Activated Audio Recording Watch) throughout the 14-day period, with guidance to pause recording when they left home or had other privacy concerns. Additionally, participants were provided an iOS device with a customized audio recording application (Redenlab®). The iOS device was positioned in a location where participants typically experience symptoms, such as a bathroom, and was set to remain plugged in and recording continuously throughout the study period. Audio recordings were securely uploaded to a cloud-based system. For the iOS device, data was transferred in real time to a secure cloud location, while watch data was uploaded at the end of the study once the device was returned to the team. To optimize the manual annotation process and ensure privacy, a pre-processing step was performed on the collected raw data using Redenlab’s Voice Activity Detection module to remove silent segments and exclude speech. After pre-processing, all audio files were manually annotated by four trained human listeners skilled in evaluating audio data to identify and label retch and expulsion events (Redenlab Ltd., Melbourne, Australia). Raters recorded the start and stop times of all retch and expulsion events, as well as related sounds such as spits, coughs, throat clearing, and interference that occurred close in time to retching or expulsion. To ensure quality and consistency of annotations, a structured training and quality assurance four-step framework was implemented:

  1. 1.

    Raters received training using individual audio events sourced from publicly available repositories. This training familiarized raters with the range of signal characteristics encountered in the study and standardized their annotation approach.

  2. 2.

    All annotations were completed using a signal-based alignment and labelling tool, which provided objective temporal markers for audio events.

  3. 3.

    While individual recordings were typically annotated by a single rater, a consensus rating process was established to safeguard quality and handle ambiguous audio segments. We implemented bi-weekly rater calls where annotators could share challenging cases and gather input. Any case lacking initial certainty was reviewed until a consensus was reached between at least two trained raters.

  4. 4.

    To further strengthen reliability, the annotation summaries were cross-referenced with contemporaneous patient notes documenting daily events (described below), allowing for contextual validation of the audio labels. Specifically, participants completed a web-based questionnaire during each day of the study period.

The web-based questionnaire contains the following questions:

Electronic patient-reported outcome (ePRO) retch/expulsion counts

A gated question (“During the past 24 h, did you vomit or retch?” Yes/No) was first posed to participants. Participants who answered “yes” were then asked to report the number of times they vomited/retched separately, with vomiting defined as the expulsion of food or liquid, and retching defined as dry heaves without expulsion. Participants were instructed to record each instance of vomiting or retching based on the physical act itself, rather than the number of trips to the bathroom.

Patient global impression of severity (PGI-S)

The PGI-S is a single-item patient-reported outcome (PRO) assessing the severity of vomiting over the past 24 h. Participants rated severity on a 5-point scale as follows: None (0), Mild (1), Moderate (2), Severe (3), and Very Severe (4).

Patient global impression of frequency (PGI-F)

The PGI-F is a single-item PRO assessing the frequency of vomiting over the past 24 h. Responses were rated on a 5-point scale as follows: None (0), Rarely (1), Sometimes (2), Often (3), and Very Often (4). Note that the PGI-F was added as an amendment to the study protocol after the first five participants had completed data collection; as such, it is not available for all participants.

All patient-reported items (ePRO retch/expulsion counts, PGI-S, and PGI-F) were administered as part of a single, daily web-based session. If a participant did not complete the session, data for that day were considered missing for all three measures.

Event bout definition

To objectively define bouts of vomiting-related events (retching and/or expulsion), we developed a data-driven framework to identify the optimal maximum time gap between two consecutive events for them to be considered part of the same bout. This approach avoids arbitrary threshold selection and ensures that bout definitions reflect meaningful temporal clustering of symptoms observed in real-world data.

We evaluated candidate thresholds ranging from 1 to 20 seconds, each representing the maximum interval allowed between two events to be grouped into a single bout. For each threshold, two bout-level metrics were computed for every participant: the daily number of bouts and the mean bout duration. To ensure comparability, both metrics were z-score normalized using the training distribution.

Generalizability was assessed using a repeated cross-validation procedure. In each iteration, five participants were held out for validation, and the remaining participants were used for training. For each validation set, z-normalized bout counts and durations were calculated using training-derived means and standard deviations. An imbalance score was then computed as \({Score}=({z}_{{count}}-{z}_{{duration}})^2\), reflecting the extent to which a threshold produced too many short bouts (over-fragmentation) versus too few long bouts (over-merging).

This process was repeated 1000 times per candidate threshold with different random splits, and the average imbalance score was calculated, and the threshold with the lowest average score was selected as optimal.

Data analysis

Daily data were reviewed for completeness, duplicate entries, and outliers. No imputation was performed. To assess measurement stability, weekly event counts for retching (R), expulsion (E), and combined retching and expulsion (R + E) from the Watch and iOS devices were calculated by summing all events across all available recording hours and normalizing to a 7-day period. Events collected via ePRO were similarly summed over all 14 days and normalized to a 7-day period.

All statistical tests were two-tailed with a significance level of α = 0.05. The following analyses were performed for both count-based and bout-based endpoints:

Patient characteristics

Basic clinical and demographic characteristics of the participants were summarized, including age, gender, BMI, race, ethnicity, and duration of disease.

Device compliance

Participant compliance in capturing audio data was assessed separately for the Watch and iOS device. Compliance was calculated as the number of available recording h per day divided by 24 and averaged over the 14-day study. To standardize event rates, all count endpoints collected via audio devices were adjusted to a 24 h day using the following formula:

$${Adjusted}\,{count}/{day}\,=\,{round}(({count}\,/\,{recorded}\,{hours})\,\times \,24).$$

To ensure data quality and minimize bias from incomplete wear or technical issues, only days with more than 8 h of available recording were included in the analysis.

Descriptive statistics

Daily and weekly retch, expulsion, and combined event counts (by source) were summarized using means, standard deviations (SD), medians, and the 10th–90th percentiles. Similar descriptive statistics were calculated for combined R + E bouts.

Time-of-day effects

Temporal patterns of audio-derived daily events were evaluated by analyzing timestamp data for retch and expulsion events. A bootstrap approach was utilized to minimize the impact of outlier subjects. In each of 2000 iterations, 25 participants were randomly sampled with replacement. For every 3-hour time bin, the median and 10th–90th percentile range of the event proportions across all iterations were computed, providing an overview of the distribution of event timing throughout the day.

Association between audio-based endpoints and patient-reported outcomes

The agreement between daily event counts collected by audio devices (Watch and iOS device) and daily counts reported by patients (ePRO) was assessed using repeated measure correlations (rmcorr) developed by Bakdash and Marusich17. Further, within-subject associations between daily event counts (collected via ePRO, Watch, and iOS devices) and corresponding daily PGI-S/PGI-F ratings were assessed using repeated measure correlations. Differences in daily event counts collected via ePRO, Watch, and iOS devices across varying frequency and severity levels, as determined by the PGI-F and PGI-S scales, were analyzed using Friedman’s test, followed by Dunn’s tests for post hoc pairwise comparisons.

Outcome reliability and effect of averaging period

Test–retest reliability was assessed using intraclass correlation coefficients (ICCs), derived from a two-way mixed-effects ANOVA model for absolute agreement18 and computed using the iccCounts package in R19. Daily event counts from Day 1 (defined individually for each participant as the first day with available data) vs Day 8, Day 1 vs Day 14, and Day 7 vs Day 14 were compared in the primary analysis. Reliability was also evaluated by comparing total event counts from the first 7 days (week 1) with those from the final 7 days (week 2) of the study.

This study was conducted as a foundational, hypothesis-generating pilot investigation. Given the exploratory nature of these novel audio-based digital endpoints, we did not apply formal corrections for multiple comparisons, as our primary objective was to maximize sensitivity to potential biomarkers for future validation.

Data availability

Individual event–level datasets are not publicly available due to patient privacy and confidentiality requirements. Individual and de-identified summary-level data may be requested, and each request will be reviewed for compliance with privacy and data-use restrictions.

Code availability

Custom Python code used for feature extraction and R code for statistical analysis is available from the authors upon request.

References

  1. Camilleri, M. et al. Gastroparesis. Nat. Rev. Dis. Prim. 4, 41 (2018).

    Article  Google Scholar 

  2. Reddivari, A. K. R. & Mehta, P. Gastroparesis. In StatPearls [Internet] (StatPearls Publishing, 2022). Bookshelf ID: NBK551528. https://www.ncbi.nlm.nih.gov/books/NBK551528/.

  3. Parkman, H. P. et al. Factors that contribute to the impairment of quality of life in gastroparesis. Neurogastroenterol. Motil. 33, e14087 (2021).

  4. Revicki, D. A. et al. The content validity of the ANMS GCSI DD in patients with idiopathic or diabetic gastroparesis. J. Patient Rep. Outcomes 2, 61 (2018).

    Article  Google Scholar 

  5. Rentz, A. M. et al. Development and psychometric evaluation of the patient assessment of upper gastrointestinal symptom severity index (PAGI-SYM) in patients with upper gastrointestinal disorders. Qual. Life Res. 13, 1737–1749 (2004).

    Article  CAS  Google Scholar 

  6. Zini, M. L. L. & Banfi, G. A narrative literature review of bias in collecting patient-reported outcomes measures (PROMs). Int. J. Environ. Res. Public Health 18, 12445 (2021).

    Article  Google Scholar 

  7. Gong, Y., Yu, J. & Glass, J. VocalSound: A dataset for improving human vocal sounds recognition. In 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 151–155 (IEEE, 2022).

  8. Zhang, C. et al. Intelligent wearable devices with audio collection capabilities to assess chronic obstructive pulmonary disease severity. Digit. Health 11, 20552076251320730 (2025).

    Article  Google Scholar 

  9. Mehl, M. R. & Holleran, S. E. An empirical analysis of the obtrusiveness of and participants’ compliance with the electronically activated recorder (EAR). Eur. J. Psychol. Assess. 23, 248–257 (2007).

    Article  Google Scholar 

  10. Schelfhout, J. et al. Validation and meaningful change thresholds for an objective cough frequency measurement in chronic cough. Lung 200, 717–724 (2022).

    Article  Google Scholar 

  11. Zimmer, A. J. et al. Objective cough counting in clinical practice and public health: a scoping review. Lancet Digit. Health 7, e1–e10 (2025).

    Google Scholar 

  12. Chung, K. F. et al. Longitudinal cough frequency monitoring in persistent coughers: daily variability and predictability. Lung 202, 561–568 (2024).

    Article  Google Scholar 

  13. Morice, A. et al. Cough frequency has a high daily variation in patients with chronic cough. ERJ Open Res. 11, 00670–02024 (2025).

    Article  Google Scholar 

  14. Kelsall, A. et al. How to quantify coughing: correlations with quality of life in chronic cough. Eur. Respir. J. 32, 175–179 (2008).

    Article  CAS  Google Scholar 

  15. Elsenbruch, S. & Enck, P. Placebo effects and their determinants in gastrointestinal disorders. Nat. Rev. Gastroenterol. Hepatol. 12, 472–485 (2015).

    Article  Google Scholar 

  16. Abelson, M. B. & Dewey-Mattia, D. The placebo effect and other confounders. Rev. Ophthalmol. (2010).

  17. Bakdash, J. Z. & Marusich, L. R. Repeated measures correlation. Front. Psychol. 8, 456 (2017).

    Article  Google Scholar 

  18. Hays, R. D. & Revicki, D. A. Reliability and validity (including responsiveness). In Assessing Quality of Life in Clinical Trials: Methods and Practice (eds Fayers, P. M. & Hays, R. D.) 25–39 (Oxford University Press, 2nd edn, 2005).

  19. Carrasco, J. L. iccCounts: an R package to estimate the intraclass correlation coefficient for assessing agreement with count data. R. J. 14, 230–244 (2022).

    Article  Google Scholar 

Download references

Acknowledgements

This work was supported by Takeda Pharmaceutical Company Limited. The funder had a role in study design, data collection, analysis, interpretation, and the decision to publish. No specific grant number.

Author information

Authors and Affiliations

  1. Takeda Development Center Americas Inc., Cambridge, MA, USA

    Yishu Gong, Jasmin Imsirovic, Brian Tracey, Nicholas Stylopoulos, Yanwei Zhang, Muna Tahir & Jing Yuan

  2. Takeda Pharmaceuticals U.S.A. Inc., Cambridge, MA, USA

    Xuefang Wu

  3. University of Melbourne, Melbourne, VIC, Australia

    Lisa Graf, Thayabaran Kathiresan & Adam P. Vogel

  4. Redenlab Ltd, Melbourne, VIC, Australia

    Lisa Graf, Thayabaran Kathiresan & Adam P. Vogel

  5. Department of Medicine, Division of Gastroenterology, Hepatology, and Nutrition, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA

    David J. Levinthal

Authors

  1. Yishu Gong
  2. Jasmin Imsirovic
  3. Brian Tracey
  4. Nicholas Stylopoulos
  5. Yanwei Zhang
  6. Muna Tahir
  7. Jing Yuan
  8. Xuefang Wu
  9. Lisa Graf
  10. Thayabaran Kathiresan
  11. Adam P. Vogel
  12. David J. Levinthal

Contributions

Y.G. designed the methodology, performed data curation, formal analysis, and visualization, and wrote the original draft. J.I. conducted formal analysis, contributed to data acquisition and visualization, and reviewed and edited the methodology. B.T. contributed to conceptualization, formal analysis, and manuscript review and editing. N.S. provided conceptualization, clinical investigation, and resources, and reviewed and edited the manuscript. Y.Z. contributed to conceptualization, methodology, statistical analysis, supervision, and manuscript review and editing. M.T. provided supervision, methodology, statistical analysis, and manuscript review and editing. J.Y. contributed to methodology, statistical analysis, and manuscript review and editing. A.V. contributed to data acquisition, conceptualization, methodology, supervision, and manuscript review and editing. L.G. performed data acquisition, coordinated the project, and reviewed and edited the manuscript. T.K. contributed to data acquisition, methodology, and manuscript review and editing. X.W. contributed to conceptualization, project coordination, resources, and manuscript review and editing. D.L. provided conceptualization, clinical expertise, supervision, and manuscript review and editing. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Yishu Gong.

Ethics declarations

Competing interests

Yishu Gong, Jasmin Imsirovic, Brian Tracey, Nicholas Stylopoulos, Yanwei Zhang, Muna Tahir, Jing Yuan, and Xuefang Wu are employees of Takeda Pharmaceutical Company Limited. Adam P. Vogel, Lisa Graf, and Thayabaran Kathiresan are employees of Redenlab. David J. Levinthal has served as an advisor for Takeda. The authors declare no other competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Gong, Y., Imsirovic, J., Tracey, B. et al. Evaluation of an audio-based monitoring system for detecting retching and vomiting in gastroparesis: an observational cohort study. npj Gut Liver 3, 36 (2026). https://doi.org/10.1038/s44355-026-00086-8

Download citation

  • Received: 20 March 2026

  • Accepted: 16 September 2026

  • Published: 05 November 2026

  • Version of record: 05 November 2026

  • DOI: https://doi.org/10.1038/s44355-026-00086-8

来源:Nature 健康护理主题 · nature.com