Background

JMIR Form Res

formative

JMIR Formative Research

JMIR Form Res

2561-326X

JMIR Publications

Toronto, Canada

v10i1e87054

10.2196/87054

Original Paper

A Conversational Platform (Okaya) for Multimodal Digital Biomarkers of Fatigue, Cognition, and Mental Health: Feasibility Observational Study

Matthew

BS, MS1Sobolev

Michael

PhD2Menvielle

Gregory

AA, BA, MA1

Okaya

540 Mohawk Drive

Boulder

United StatesCornell Tech

New York

United States

Schwartz

Amy

Balcarras

Matthew

Rodriguez-Conde

Ivan

Correspondence to Matthew So, BS, MS, Okaya, 540 Mohawk Drive, Boulder, CO, 80303, United States, 1 9259841028; matthew@okaya.me

2026

142026

e87054

031120252702202627022026

2026

This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Formative Research, is properly cited. The complete bibliographic information, a link to the original publication on https://formative.jmir.org, as well as this copyright and license information must be included.

Background

Collection of multimodal data (video, audio, and text) can yield digital biomarkers relevant to mental health, fatigue, and cognition. However, the feasibility and signal characteristics in operational populations remain underexplored.

Objective

The objectives of this study were to (1) extract an evidence-based library of vision, speech, and language features; (2) assess the feasibility of a fully remote conversational platform (Okaya) for collecting analyzable multimodal data; and (3) conduct preliminary signal checks for depression, fatigue, and cognition.

Methods

Participants were recruited from the US Air Force and US Space Force. All participants completed the Okaya check-in, which included a voice conversation with a large language model. A total of 66 visual, acoustic, and text features were extracted from each interaction between the participant and the large language model. For validation purposes, the study also collected measures of depression (Patient Health Questionnaire–9), fatigue (Cancer Fatigue Scale), and cognition (trail making test). We evaluated the feasibility of the platform and correlation between the extracted features and the validated assessments.

Results

A total of 8 unique participants contributed with 62 sessions over a period from March 6, 2025, to August 6, 2025. The platform was deemed feasible as 6 of the 8 participants opted to complete more than one session, and the 3 participants who provided feedback reported high overall experience and usability. From the data perspective, preliminary correlations produced significant results for multiple potential digital biomarkers, including (1) pitch (P=.047), volume SD (P=.04), volume slope (P=.04), automated readability index complexity (P=.047), Flesch-Kincaid complexity (P=.04), and Gunning Fog complexity (P=.04) for depression; (2) pitch (P=.009), volume SD (P=.007), volume slope (P=.02), average F2 formant frequency (P=.03), Gunning Fog complexity (P=.049), and eyelid droop (P=.047) for fatigue; and (3) shimmer (P=.03) for cognition. We also observed how features varied over time among participants with multiple sessions.

Conclusions

The conversational and artificial intelligence–enabled platform was feasible among an operational sample as a method to collect multimodal data correlated with depression, fatigue, and cognition. These results align with those for previously discovered digital biomarkers of mental health, fatigue, and cognition and inform the development of personalized models for each user while detecting anomalies in a remote monitoring setting.

digital biomarkersmultimodal sensingcomputer visionspeech acousticsnatural language processingartificial intelligenceAI

Introduction

Mental health, fatigue, and cognitive functioning are critical to overall well-being and operational performance, yet their assessment often relies on intermittent self-report or clinician-administered instruments. Such methods, while clinically validated, can be burdensome, subjective, and limited in their ability to capture dynamic changes over time [1-9]. Mobile health technology provides an alternative solution due to the ability of apps and sensors to collect high-fidelity and high-frequency data pertaining to activity, behavior, symptoms, and cognition [10-12]. With increasing demands for continuous and remote monitoring—especially in occupational and high-performance settings—digital biomarkers have emerged as a promising avenue for objective and scalable assessment of psychological and cognitive states [13-16].

Recent advances in computer vision, speech acoustics, and natural language processing have enabled the extraction of behavioral and physiological signals from everyday digital interactions. Multimodal sensing that integrates video, audio, and language data provides an opportunity to quantify indicators of affect, fatigue, and cognitive load with greater ecological validity than traditional laboratory tests [11,17]. Prior research has demonstrated associations between acoustic features (eg, pitch variation [18-23], jitter [18,19,24-26], and spectral harmonicity [25-28]), facial expression dynamics (eg, eye gaze [29-32] and microexpressions [33]), and linguistic markers (eg, sentiment [28,34,35], complexity [28,36], and emotional valence [28,35,37-41]) with symptoms of depression, stress, and cognitive decline. However, most existing work has been conducted in tightly controlled environments, and less is known about the feasibility of collecting such data in naturalistic or operational contexts [42].

Conversational platforms, especially those powered by large language models, introduce new opportunities for naturalistic data capture [43-46]. By engaging users through dialogue, these systems can simultaneously elicit and record multimodal signals—speech, facial behavior, and language—while maintaining a familiar and low-burden interaction format [47-52]. Such interfaces align with the increasing integration of conversational artificial intelligence (AI) into health, wellness, and performance domains, offering a pathway toward longitudinal, user-centered monitoring of mental and cognitive health. Nevertheless, empirical evaluation of these systems’ feasibility and usability remains limited, particularly in real-world populations such as operational and occupational groups [53,54]. These groups can also be more susceptible to experiencing mental and cognitive health challenges due to their job functions [55,56].

The objective of this study was to evaluate the feasibility of a fully remote conversational platform, Okaya, for collecting analyzable multimodal data (video, audio, and text) in an operational population from the US Air Force (USAF) and Space Force. Specifically, we sought to determine whether participants would engage with the platform across multiple sessions and whether the system could reliably extract a comprehensive set of visual, acoustic, and linguistic features suitable for future digital biomarker modeling of depression, fatigue, and cognition. Finally, we present a case study for the design of a risk score based on a model of extracted digital biomarkers to illustrate the future potential of the platform.

MethodsStudy Population

This study was part of an existing Small Business Innovation Research Direct to Phase II contract with the USAF, and a pool of potential participants from the 71st Special Operations Squadron, provided by the USAF, was introduced to the Okaya platform as part of a pilot program. Participation was entirely optional. A total of 8 unique users participated in the study, contributing 62 sessions (median 2.5, range 1-38, IQR 1.5-7.5). Participants provided data during the period from March 6, 2025, to August 6, 2025.

Ethical Considerations

Ethics approval for this study was obtained from Sterling Institutional Review Board (approval 12300-GMenvielle), and all participants signed an informed consent statement that specified anonymized access to research records, potential risks, and voluntary participation. No data were collected except for email addresses, which were required to access the Okaya platform. No demographic or sample composition information was collected due to requirements from the USAF. Participants received no compensation for taking part.

Procedure and Data CollectionOverview

Completion of Okaya check-ins [57] was entirely voluntary. Data collection took place on the Okaya website, and users were allowed to select the device they preferred to access the website with, such as smartphones, tablets, or laptops. The browser used to access the website was also free to select by the user. Once users had successfully acknowledged the consent statement and registered for an Okaya account, they could complete a workflow as described in Figure 1. Participants were asked to complete 3 assessments: the Patient Health Questionnaire–9 (PHQ-9) [58], Cancer Fatigue Scale (CFS) [59], and trail making test (TMT) [60].

Figure 1.

Illustration of the Okaya platform and data collection. The 6 steps for a successful sample collection include log-in, device checks, Patient Health Questionnaire–9 (PHQ-9), Cancer Fatigue Scale (CFS), trail making test (TMT), and Okaya check-in.

Upon successful completion of all 3 assessments, users were presented with a text and voice prompt from a large language model (using OpenAI’s GPT-4o-mini model with a custom system prompt, referred to as “Sanora”), shown in Figure 2. Users were first required to meet the device and browser requirements, grant permission to access the microphone and camera, and meet an upload speed requirement of 2 Mb/s. Video was recorded continuously, and users could choose between webcam video, blurred video, a landmark representation of their face, an outline of their face, or a set of axes denoting the location of their face, designed to accommodate users with differing levels of comfort with being recorded. Although video was shown on the interface, raw video was not sent to our servers; instead, a set of facial landmarks was extracted via Google MediaPipe. These landmarks were automatically extracted using a proprietary machine learning model, and 478 landmarks were output, each with an x, y, and z coordinate, as well as 52 “blendshapes,” each representing various facial expressions. The check-in interface, as shown in Figure 2, included the text of Sanora’s prompt at the top; the representation of the user’s webcam video at the center; and a control button, either “Record,” “Stop,” or “Continue,” at the bottom of the screen.

Figure 2.

Screenshots of the Okaya check-in. Permission was obtained from the user depicted.

After passing device checks and clicking “I’m ready to check-in,” Sanora would generate a prompt for the user to respond to asking about their general mental well-being, emotions, or experiences they had had recently. Upon clicking “Record” to record a response to Sanora, the user’s microphone started recording, and an automatic transcription service from Amazon Transcribe started transcribing their speech until the user clicked “Stop” and stopped recording. Both the audio recording and transcribed text were sent back to the Okaya servers. Sanora then used the transcribed text to generate a response. This process continued for 4 turns (5 messages from Sanora and 4 responses from the participant). Upon conversation completion, the user clicked “Continue” to submit the session for analysis, after which a completion screen would indicate that the results were being processed. They would then be redirected to the research dashboard.

All back-end processing took place on Amazon Web Services infrastructure, with communication between the website (front end) and back end occurring via Representational State Transfer application programming interface calls. Data were also stored on Amazon Web Services infrastructure.

Feature Extraction

From the landmarks, audio, and transcribed text, we extracted 66 features based on the literature. A full list and descriptions of each implemented feature can be found in Multimedia Appendix 1, whereas summaries are shown in Tables 1-3.

Table 1.

Mapping features for depression.

Feature	Theoretical reasoning	References	Implemented features
Visual features
Flat affect	A predominant lack of facial expressiveness can be observed in some individuals with depression. This might manifest as a “blank” or “emotionless” appearance even in situations that would typically provoke an emotional response.	[33]	Affect_measure
Facial expression	“Sad” facial expressions, such as less frequent smiling, have been associated with the severity of depression.	[31,40,61-63]	Mouth_curvature and eyebrow_droop
Reduced eye contact	People with depression might avoid making eye contact, which can be a sign of feelings of worthlessness or guilt.	[29-32]	Gaze_down_dist, gaze_x_dist, and gaze_y_dist
Slow movements	Psychomotor retardation can manifest as slowed facial movements or reactions in those with depression.	[61]	Movement_speech_measure
Audio features
Monotone speech	Individuals with depression might exhibit a lack of variability in pitch, leading to speech that sounds monotonous.	[18-22,27]	Pitch and pitch_std
Reduced speech volume	Speaking more softly or with less projection than usual can be indicative of depression.	[18,22,64-67]	Vol, vol_std, and ealvi
Reduced vocal prosody	A decrease in the rhythmic and melodic aspect of speech can be indicative of depression.	[18-20,24,25,28,64,68]	Jitter, shimmer, timbre, and formant
Long pauses	Depression might result in more increased pauses between words or sentences, reflecting hesitancy, slowed thinking, or difficulty in organizing thoughts.	[18,21,68,69]	Audio_pauses
Transcript features
Long pauses	Depression might result in more increased pauses between words or sentences, reflecting hesitancy, slowed thinking, or difficulty in organizing thoughts.	[18,21,68,69]	Transcript_pauses
Slow speech rate	People with depression might speak more slowly, potentially reflecting slowed cognitive processing.	[22,23,68-71]	Words_per_s
Reduced responsiveness	A person with depression might be less verbally responsive in conversations, potentially taking longer to reply or offering shorter answers.	[21,22,69,70]	Response_latency and transcript_len
Decreased complexity and length of speech	Speech in individuals with depression might be less complex in terms of vocabulary and sentence structure, and individuals might be less talkative overall.	[28,34-37,39,40,67]	Complexity, emotion_keyword_prop, and sentiment_score

Table 2.

Mapping features for fatigue.

Feature	Theoretical reasoning	References	Implemented features
Visual features
Droopy eyelids	The muscles around the eyes may begin to sag due to tiredness, causing the eyelids to droop.	[72]	Eyelid_droop
Decreased blink rate	Fatigue can lead to a reduced rate of blinking.	[73-77]	Blinks_per_s and blink_len
Yawning	While yawning is a natural behavior, frequent yawning can be an overt sign of fatigue or drowsiness.	[76,78]	Yawns_per_s and yawn_len
Reduced expressiveness	Fatigue might cause an individual to have fewer facial movements or expressions.	[72]	Mouth_curvature
Audio features
Decreased volume	A tired individual might speak more softly or with less energy than when they are well rested.	[8,79]	Vol and vol_std
Decreased pitch variability	A fatigued voice might sound more monotonous, with less variation in pitch.	[23,25]	Pitch and pitch_std
Flatter voice profile	Decreased variation and width of the voice’s spectral profile can be indicative of fatigue.	[25,79]	Timbre and formant
Increased pauses	There might be more frequent and longer pauses between words or sentences, reflecting slowed cognitive processing or the need to gather thoughts.	[23,79,80]	Audio_pauses
Transcript features
Increased pauses	There might be more frequent and longer pauses between words or sentences, reflecting slowed cognitive processing or the need to gather thoughts.	[23,79,80]	Transcript_pauses
Slow speech rate	The overall rate of speech might decrease when a person is tired.	[23,80]	Words_per_s
Shortened responses	Fatigued individuals might offer shorter answers or engage less in conversation.	[23]	Transcript_len
Impaired memory recall	Fatigued individuals might struggle to remember certain words, names, or details, leading to more frequent use of filler words such as “um” or “uh.”	[80]	Hesitations_per_s

Table 3.

Mapping features for attention and cognition.

Feature	Theoretical reasoning	References	Implemented features
Visual features
Decreased blink rate	Blinking might increase when in periods of high attention demand during cognitive tasks.	[81]	Blinks_per_s
Eye movements	Eye movements outside the line of focus might be reduced when in a period of visual attention.	[82]	Eye_movements_per_s
Audio features
Volume	Speech volume can indicate differences in arousal and fatigue.	[26]	Vol, vol_std, vol_range, and volume_slope
Pitch and pitch variability	Pitch might be less varied during periods of low engagement or emotional expressivity.	[26,83]	Pitch and pitch_std
Voice timbre	Voice might be breathier or more strained when lacking attention.	[26]	Jitter, shimmer, and timbre
Transcript features
Reduced responsiveness	Response latency has been shown to enable the detection of the Stroop effect (delay in reaction time).	[84]	Response_latency

The webcam capture rate varied depending on a number of factors, including the user’s device and browser. As this capture rate was not controllable by the Okaya platform, the analysis had to handle landmark frames captured at nonregular times. Generally, the capture rate was 30 to 60 frames per second, but they were usually not captured at a consistent rate. All sessions were interpolated to a capture rate of 30 frames per second, where each time step was a linear interpolation between the 2 neighboring frames. Features were extracted in Python (Python Software Foundation) using the libraries SciPy and NumPy.

Similarly, audio capture rate varied per user, so all audio recordings were resampled to a sampling frequency of 22,050 Hz. Features were extracted in Python using the libraries Librosa, NumPy, SciPy, Parselmouth, and Statsmodels.

Text features were extracted in Python using the libraries NumPy, Textacy, and Natural Language Toolkit (Team NLTK). The exact implementation of each feature extractor is omitted as the details of the Okaya platform are proprietary.

Several temporal aggregations were used to calculate feature values: averaged per second, per frame, per response, or per detected event. The exact aggregation and unit used for each feature are listed in Multimedia Appendix 1.

ResultsEngagement and Usability Results

Three participants provided qualitative and survey feedback following completion of their Okaya sessions. Full survey responses can be found in Multimedia Appendix 2. Overall, preliminary user satisfaction with the conversational platform was given a rating of between 4 and 5 out of 5 on overall experience and usability among all participants who provided feedback (3/3). These participants described the platform as intuitive and easy to navigate and reported feeling adequately supported during setup and use. One participant highlighted the responsiveness of the support team and appreciated having direct access to human assistance when needed.

Motivations for engagement included curiosity about the technology and the opportunity for structured self-reflection during check-ins. One participant noted that the conversations prompted personal insight and self-awareness, describing the process as a valuable moment to “intentionally reflect” on life aspects discussed during the session. Common barriers to more frequent engagement included time constraints, insufficient or poorly timed reminders, and occasional concerns about privacy when speaking in shared environments. Notably, all participants who completed the survey expressed overall confidence in data privacy and platform security, each giving this aspect a rating of 5 out of 5.

Suggestions for improvement focused on enhancing convenience and personalization, including text-based reminders with direct log-in links, greater mobile accessibility, and shorter baseline tasks prior to the conversational check-in. Survey participants also expressed interest in receiving personalized summaries or progress visualizations of their responses. All the participants who provided feedback indicated interest in participating in future sessions using the Okaya platform, reinforcing the platform’s acceptability and potential for longitudinal engagement within operational populations.

Correlation AnalysisOverview

Pearson correlations and their associated significance are listed below. Although every participant was included in the dataset and most completed multiple sessions, we only took into account the earliest sample for each participant for the correlation analysis. The earliest sample was used to illustrate the use of an initial sample as a baseline, upon which personalized models can be built, illustrated in the case study below. We also observed low intraparticipant variability for each clinical measure, which hindered the application of more sophisticated statistical models at this stage. For each feature, outliers (defined as any measurement more than 2 SDs from the mean) were removed. The most significant features are plotted against the associated measures in Figure 3.

Figure 3.

Most significant features for each measure and their associated correlation. CFS: Cancer Fatigue Scale; PHQ-9: Patient Health Questionnaire–9; TMT: trail making test.

Correlations for Depression

Exploratory correlation analyses revealed several significant associations between multimodal features and depressive symptom severity as measured using the PHQ-9. Acoustic and linguistic indicators showed the strongest relationships. In this small sample specifically, pitch (r=0.71; P=.047) and volume slope (r=0.78; P=.04) were positively correlated with PHQ-9 scores. This suggests preliminary patterns and warrants further exploration into predictive modeling to validate that greater variability and upward trajectories in vocal intensity may correspond to higher self-reported depression levels. In contrast, volume SD was negatively correlated (r=−0.73; P=.04), encouraging a possible exploration into how participants with greater depressive symptoms may exhibit reduced dynamic range in speech.

Textual complexity metrics—including the automated readability index (r=−0.71; P=.047), Flesch-Kincaid complexity (r=−0.74; P=.04), and Gunning Fog index (r=−0.72; P=.04)—also demonstrated negative correlations with PHQ-9 scores; thus, the individuals in our preliminary dataset reporting higher depressive symptoms tended to produce shorter or linguistically simpler utterances during conversational interactions. Due to the low sample size, it is impossible to say whether there are any actual meaningful correlations; however, it does warrant further exploration with larger sample sizes and more resilient analyses. Due to our limited sample size, individual correlations are also not expected to be reliable or generalizable at this stage. Together, these relationships support the feasibility of deriving speech- and language-based digital biomarkers of depression from naturalistic conversational data.

Correlations for Fatigue

Analysis of correlations between multimodal features and fatigue as measured using the CFS identified several significant acoustic, visual, and linguistic markers. Among acoustic features, pitch (r=0.84; P=.009), volume slope (r=0.84; P=.02), and average F2 formant frequency (r=0.76; P=.03) were positively correlated with fatigue severity, whereas volume SD (r=−0.85; P=.007) showed a strong negative association. This warrants further investigation into how individuals reporting higher fatigue might exhibit flatter, less dynamically modulated vocal patterns, consistent with decreased speech energy and prosodic variation observed in fatigue-related speech studies.

Additionally, the linguistic Gunning Fog index was inversely correlated with CFS scores (r=−0.71; P=.049), indicating that higher fatigue levels were associated with simpler or less complex speech production. One visual feature, eyelid droop (r=0.71; P=.047), also demonstrated a significant positive association with fatigue, consistent with prior literature linking ocular and facial muscle changes to tiredness. Together, these findings highlight the sensitivity of both speech acoustics and facial metrics to self-reported fatigue, supporting their potential use as digital biomarkers for real-time fatigue monitoring. However, due to the limited sample size and analysis, individual associations should not be interpreted as predictive or applicable to general populations at this stage.

Correlations for Attention and Cognition

Exploratory analyses examining associations between multimodal features and cognitive performance as measured using the TMT identified a single significant relationship. The acoustic feature shimmer, which captures cycle-to-cycle variability in vocal amplitude, was positively correlated with TMT scores (r=0.74; P=.03). This association points to greater irregularity in voice amplitude when participants had longer task completion times. Further research into how lower cognitive efficiency might potentially relate to fatigue-related motor or attentional variability in speech production may be worthwhile. Although preliminary, this finding aligns with those of prior literature linking changes in vocal stability to fluctuations in attention and executive functioning. These results indicate that fine-grained acoustic features derived from conversational speech may offer a feasible, low-burden proxy for cognitive performance monitoring in remote or operational settings. Once again, although the exact associations found should not be treated as reliable or generalizable at this stage, this does suggest a potential future study with larger samples and analyses.

Descriptive Summary of Significant Associations

To complement the quantitative analyses, we conducted a descriptive review of the distribution of significant multimodal features across clinical domains. As shown in Table 4, depression-related features were primarily linguistic and acoustic in nature, with no significant visual markers identified. Fatigue exhibited the broadest feature coverage, encompassing 1 visual variable, 4 acoustic variables, and 1 linguistic variable. When examining fatigue subscales, distinct modality patterns emerged: the physical and affective fatigue components were each associated with multiple acoustic features, whereas the cognitive subscale was linked to both visual and acoustic indicators. In contrast, cognitive performance as measured using the TMT was primarily associated with a single acoustic feature, reflecting a narrower signal profile. It is important to note that feature set size also had an effect; for example, fatigue measures displayed a clear feature representation across audio features, but audio features also made up a larger proportion of the total features compared to visual and text features. There is no clear evidence of any specific feature modality showing broader significance or sensitivity.

Table 4.

Count of significant features for each subscale of the clinical measures and feature type out of the total number of features tested.

Clinical measure and subscale	Feature categories, n (%)
	Visual features (n=13)	Audio features (n=28)	Text features (n=25)
PHQ-9^a	0 (0.0)	3 (10.7)	3 (12.0)
CFS^b	1 (7.7)	4 (14.3)	1 (4.0)
Physical	1 (7.7)	4 (14.3)	3 (12.0)
Affective	0 (0.0)	5 (17.9)	0 (0.0)
Cognitive	1 (7.7)	5 (17.9)	0 (0.0)
TMT^c	0 (0.0)	1 (3.6)	0 (0.0)
Part A	0 (0.0)	0 (0.0)	1 (4.0)
Part B	0 (0.0)	1 (3.6)	0 (0.0)

^aPHQ-9: Patient Health Questionnaire–9.

^bCFS: Cancer Fatigue Scale.

^cTMT: trail making test.

Risk Score Case Study

To explore the feasibility of individualized monitoring, we conducted case studies on participants who completed multiple Okaya sessions. For each individual, a composite “risk score” was calculated by standardizing the significant multimodal features (z-scoring within participants) and summing the deviations from the mean, with directionality determined by the feature’s correlation with the clinical measure. This approach provided a demonstration of an interpretable, participant-specific index of deviation from baseline functioning across time; however, due to the low sample size and unstable correlations, it is purely used as a demonstration in this case, not a proof of actual meaningful capability.

An example use case can use a longitudinal visualization of these risk scores to reveal dynamic fluctuations in both depression- and fatigue-related indexes across sessions. For example, in participants with higher variability, increases in risk score can correspond to periods of increased self-reported symptom severity, whereas lower or stable values can align with more consistent well-being reports. Although exploratory, these case studies illustrate how multimodal conversational data can support within-person anomaly detection and temporal tracking of psychological and physiological states. This example provides a demonstration of how individualized, feature-based composite scores may offer a viable framework for early identification of changes in mood, fatigue, or cognitive function in remote monitoring contexts.

Data on depression and fatigue for the 2 participants with the most sessions are shown in Figures 4 and 5, respectively (TMT only exhibited 1 significant feature and, as such, a composite score was not meaningful).

Figure 4.

Risk score case study for depression.

Figure 5.

Risk score case study for fatigue.

DiscussionPrincipal Findings

We found several multimodal features that correlated with depression, fatigue, and cognition, some of which extended beyond patterns reported in prior literature. As expected, depression was characterized by reduced variability in vocal intensity and lower linguistic complexity, consistent with psychomotor slowing and affective flattening. Fatigue demonstrated a broader multimodal signature involving both acoustic and visual domains, including reduced volume variability, higher pitch, and greater eyelid droop, reflecting decreased energy and arousal. Notably, volume slope, a feature we did not find in the literature, was significant for both depression and fatigue, suggesting that individuals with depression or fatigue might have noticeable patterns in how speech volume progresses over time, especially during utterances provided on the Okaya platform. Additionally, complexity, a feature previously linked to depressive speech, was found to correlate with fatigue, indicating that depression and fatigue may share some similar effects on word choice and diversity [28,36]. Together, these findings highlight the potential for future study of established digital biomarkers in the Okaya platform and the emergence of cross-domain signals that may bridge affective, cognitive, and fatigue-related processes.

The findings of this feasibility study indicate that the Okaya conversational platform is both usable and acceptable for collecting multimodal data in an operational setting. Participants demonstrated relatively sustained engagement, with most completing multiple sessions, suggesting that the platform’s conversational format and remote accessibility supported ongoing participation. Usability feedback further underscored this positive experience—participants rated the system highly in terms of ease of use, clarity of instructions, and overall satisfaction. Qualitative feedback revealed that users found the interaction intuitive and meaningful, often describing the check-in as an opportunity for reflection rather than a burdensome task. Importantly, participants expressed high confidence in data privacy and security, which is critical for user trust in AI-enabled health technologies [85-87]. At the same time, feedback highlighted opportunities for improvement, including more flexible scheduling options, reminders, and greater mobile accessibility. Together, these findings suggest that Okaya’s design successfully balanced data collection rigor with user comfort and autonomy, supporting its feasibility for repeated deployment in diverse and high-performance populations. These results provide an encouraging foundation for the platform’s next phase of validation and refinement, emphasizing scalability, personalization, and long-term adherence.

The findings from the case study align closely with the long-term vision for the Okaya platform. The observed fluctuations in individualized risk scores and their correspondence with self-reported symptom changes demonstrate the platform’s potential for longitudinal, within-person monitoring—a central principle behind the Okaya Index and the simplified risk score we presented. Rather than relying on evidence from population-based studies, the Okaya platform emphasizes personalized baselines and interprets deviations relative to each individual’s unique behavioral and physiological profile. This individualized approach reflects the platform’s goal of enabling early identification of shifts in emotional, cognitive, or fatigue states. As the platform evolves, integration of these multimodal risk scores into the Okaya Index will enable continuous refinement of trajectories across time and populations, ultimately supporting adaptive, AI-driven insights that can inform both clinical decision-making and support for operational populations ranging from pilots to firefighters. This trajectory aligns with the goal of advancing proactive mental health and performance optimization in real-world settings.

From a clinical and research perspective, these findings illustrate the promise of conversational, multimodal sensing as a foundation for precision mental health and digital phenotyping. By capturing subtle changes in voice, facial behavior, and language over time, systems such as Okaya can complement traditional self-report and clinician-administered tools, providing a richer, continuous view of mental and cognitive functioning. Such individualized, data-driven assessments have potential applications across health care, occupational performance, and behavioral health monitoring, particularly in cases in which real-time insight and early intervention are critical. Importantly, the Okaya platform’s emphasis on nondiagnostic and privacy-preserving analytics aligns with emerging standards for responsible AI in health technology [13,88,89]. As digital biomarkers move closer to clinical integration, continued validation across diverse populations and conditions will be essential to ensure generalizability, interpretability, and equity [85,86,90]. Ultimately, this study of the Okaya platform represents a step toward an ethically grounded, adaptive ecosystem for mental health monitoring—one that transforms multimodal data into actionable insights while safeguarding user trust and autonomy.

This study has several limitations that should be considered when interpreting the findings. First, the sample size was small and limited due to our focus on operational populations, which are harder to recruit, constraining generalizability and the ability to model interindividual variability. Thus, the correlations we report are not meant to be statistically resilient or generalizable at this stage. Therefore, these findings preclude reliable statistical inference of robust estimates of the correlations in these data and, rather, suggest future avenues for discovery. Second, although validated instruments such as the PHQ-9, CFS, and TMT were used for comparison, these measures may not fully capture the temporal dynamics or subtle within-person fluctuations that multimodal digital biomarkers are designed to detect. Traditional self-report and performance-based assessments tend to be episodic and relatively stable, which may underestimate moment-to-moment variability in affect, fatigue, and cognition. Future studies should incorporate higher-resolution clinical measures—such as ecological momentary assessment or brief daily self-reports—that provide richer temporal data and enable stronger coupling between behavioral features and self-reported states [1-3,87,91]. Integrating these temporally aligned measures would allow for more precise modeling of intraindividual change and enhance the ecological validity of the derived digital biomarkers. Finally, while the Okaya platform demonstrated feasibility and acceptability in this study, further validation with larger, more diverse samples and extended longitudinal follow-up is necessary for predictive modeling and design of personalized remote monitoring systems.

Conclusions

The conversational and AI-enabled platform (Okaya) was feasible among an operational sample from the US Air and Space Force as a method for collecting multimodal data correlated with depression, fatigue, and cognition. Future work will examine larger samples with repeated measures to assess the test-retest reliability and predictive validity of multimodal digital biomarkers. Ultimately, the goal of the platform is the development of personalized models for each user while detecting anomalies in a remote monitoring setting.

The authors would like to thank Sara Ghandehari, MD, MCSO, FACHE, as well as Nick Uram, PsyD, for their invaluable contributions toward the development of the Okaya platform. The authors are also grateful to Will Jackson for his assistance in literature review. The authors declare the use of generative artificial intelligence (GenAI) in the research and writing process. According to the Generative Artificial Intelligence Delegation Taxonomy (2025), the following tasks were delegated to GenAI tools under full human supervision: literature search and systematization, code generation, code optimization, process automation, creation of algorithms for data analysis, data analysis, and visualization. The GenAI tool used was GPT-4o via the GitHub Copilot platform. Although the code design and functions were entirely defined by the authors, GitHub Copilot was used to generate autocomplete suggestions as the code was being written. The generated code was used for correlation calculations, risk score case study calculation, and generation of all shown figures. Responsibility for the final manuscript lies entirely with the authors. GenAI tools are not listed as authors and do not bear responsibility for the final outcomes.

Funding

This study was funded by SmartTec Inc to fulfill deliverables under a Direct to Phase II award from AFWERX (Department of the Air Force; research and innovation).

Data Availability

The datasets generated and analyzed during this study are not publicly available due to their use in the development of a proprietary platform as well as confidentiality requirements from the US Department of Defense. Public release of the data could enable reverse engineering of the platform or unauthorized commercial use. However, unidentified data may be available from the corresponding author on reasonable request.

Conceptualization: M So (lead), GM (equal), M Sobolev (equal)

Data curation: M So (lead), GM (equal)

Formal analysis: M So (lead), M Sobolev (equal)

Funding acquisition: GM

Investigation: M So (lead), M Sobolev (equal)

Methodology: M Sobolev (lead), M So (equal), GM (supporting)

Resources: M So (lead), GM (supporting)

Software: M So

Validation: M So (lead), M Sobolev (supporting)

Visualization: M So

Writing—original draft: M Sobolev (lead), M So (supporting)

Writing—review and editing: M So (lead), M Sobolev (equal), GM (supporting)

M So and GM are employed by and hold financial interest in SmartTec Inc, who develops the Okaya platform, and could benefit from research outcomes. M Sobolev receives consulting fees from SmartTec Inc.

Abbreviations

artificial intelligence

CFS

Cancer Fatigue Scale

PHQ-9

Patient Health Questionnaire–9

TMT

trail making test

USAF

US Air Force

References1

Sliwinski

Mogle

Hyun

Munoz

Smyth

Lipton

Reliability and validity of ambulatory cognitive assessments

Assessment2018012511430

10.1177/1073191116643164

27084835

Katz

Wang

Nester

T‐MoCA: a valid phone screen for cognitive impairment in diverse community samples

Alzheimers Dement (Amst)202102131e12144

10.1002/dad2.12144

33598528

Singh

Strong

Ecological momentary assessment of cognition in clinical and community samples: reliability and validity study

J Med Internet Res202306225e45028

10.2196/45028

37266996

McHugh

Behar

Readability of self-report measures of depression and anxiety

J Consult Clin Psychol20091277611001112

10.1037/a0017124

19968386

McHugh

Behar

Reply to “Further issues in determining the readability of self-report items: comment on McHugh and Behar (2009)”

J Consult Clin Psychol20121280611211122

10.1037/a0030455

23205507

Schinka

Further issues in determining the readability of self-report items: comment on McHugh and Behar (2009)

J Consult Clin Psychol2012805952955

10.1037/a0029928

22924333

Levis

Benedetti

Ioannidis

Patient Health Questionnaire-9 scores do not accurately estimate depression prevalence: individual participant data meta-analysis

J Clin Epidemiol202006122115128.e1

10.1016/j.jclinepi.2020.02.002

32105798

Kang

Guo

Discrepancies between self-rated depression and observed depression severity: the effects of personality and dysfunctional attitudes

Gen Hosp Psychiatry2021702530

10.1016/j.genhosppsych.2020.11.016

33689981

Inoue

Tanaka

Nakagawa

Utility and limitations of PHQ-9 in a clinic specializing in psychiatric care

BMC Psychiatry20120731273

10.1186/1471-244X-12-73

22759625

Sim

Mobile devices and health

N Engl J Med201909538110956968

10.1056/NEJMra1806949

31483966

Insel

Digital phenotyping: technology for a new science of behavior

JAMA20171033181312151216

10.1001/jama.2017.11295

28973224

Mohr

Zhang

Schueller

Personal sensing: understanding mental health using ubiquitous sensors and machine learning

Annu Rev Clin Psychol2017058132347

10.1146/annurev-clinpsy-032816-044949

28375728

Coravos

Khozin

Mandl

Developing and adopting safe and effective digital biomarkers to improve patient outcomes

NPJ Digit Med20192114

10.1038/s41746-019-0090-4

30868107

Dagum

Digital biomarkers of cognitive function

NPJ Digit Med2018110

10.1038/s41746-018-0018-4

31304295

Jacobson

Weingarden

Wilhelm

Digital biomarkers of mood disorders and symptom change

NPJ Digit Med201923

10.1038/s41746-019-0078-0

31304353

Sobolev

Gullapalli

Rahman

Advancing the science of digital biomarkers

DigiBiom ’21: Proceedings of the 2021 Workshop on Future of Digital Biomarkers

Jun 25, 2021

10.1145/3469266.3473711

Garcia-Ceja

Riegler

Nordgreen

Jakobsen

Oedegaard

Tørresen

Mental health monitoring with multimodal sensing and machine learning: a survey

Pervasive Mob Comput20181251126

10.1016/j.pmcj.2018.09.003

König

Tröger

Mallick

Detecting subtle signs of depression with automated speech analysis in a non-clinical sample

BMC Psychiatry20221227221830

10.1186/s12888-022-04475-0

36575442

Silva

Lopes

Galdino

Almeida

Voice acoustic parameters as predictors of depression

J Voice2024013817785

10.1016/j.jvoice.2021.06.018

34353686

Shin

Cho

Park

CHK

Detection of minor and major depression through voice as a biomarker using machine learning

J Clin Med202107810143046

10.3390/jcm10143046

34300212

Alpert

Pouget

Silva

Reflections of depression in acoustic measures of the patient’s speech

J Affect Disord2001096615969

10.1016/s0165-0327(00)00335-9

11532533

Cohen

Rodriguez

Opler

Evaluating speech latencies during structured psychiatric interviews as an automated objective measure of psychomotor slowing

Psychiatry Res202410340116104

10.1016/j.psychres.2024.116104

39137558

Vogel

Fletcher

Maruff

Acoustic analysis of the effects of sustained wakefulness on speech

J Acoust Soc Am201012128637473756

10.1121/1.3506349

21218906

Ozdas

Shiavi

Silverman

Wilkes

Investigation of vocal jitter and glottal flow spectrum as possible cues for depression and near-term suicidal risk

IEEE Trans Biomed Eng20040951915301540

10.1109/TBME.2004.827544

15376501

Gao

Yang

A rapid, non-invasive method for fatigue detection based on voice information

Front Cell Dev Biol202210994001

10.3389/fcell.2022.994001

36176279

Hamdan

Deeb

Sibai

Rameh

Rifai

Fayyad

Vocal characteristics in children with attention deficit hyperactivity disorder

J Voice200903232190194

10.1016/j.jvoice.2007.09.004

18082369

Zhao

Fan

Vocal acoustic features as potential biomarkers for identifying/diagnosing depression: a cross-sectional study

Front Psychiatry202213815678

10.3389/fpsyt.2022.815678

35573349

Arevian

Bone

Malandrakis

Clinical state tracking in serious mental illness through computational analysis of speech

PLoS One2020151e0225695

10.1371/journal.pone.0225695

31940347

Hinchliffe

Lancashire

Roberts

Eye-contact and depression: a preliminary report

Br J Psychiatry197011117540571572

10.1192/bjp.117.540.571

5480708

Suslow

Hoepfel

Kersting

Bodenschatz

Depressive symptoms and visual attention to others’ eyes in healthy individuals

BMC Psychiatry2024036241184

10.1186/s12888-024-05633-2

38448877

Fiquer

Moreno

Brunoni

Barros

Fernandes

Gorenstein

What is the nonverbal communication of depression? Assessing expressive differences between depressive patients and healthy volunteers during clinical interviews

J Affect Disord2018101238636644

10.1016/j.jad.2018.05.071

29957481

Waxer

Nonverbal cues for depression

J Abnorm Psychol197406833319322

10.1037/h0036706

4844922

Buyukdura

McClintock

Croarkin

Psychomotor retardation in depression: biological underpinnings, measurement, and treatment

Prog Neuropsychopharmacol Biol Psychiatry20110330352395409

10.1016/j.pnpbp.2010.10.019

21044654

Hadžić

Ohse

Alkostantini

Emotional dynamics in semi-clinical settings: speech emotion recognition in depression-related interviews

11th International Conference on Information and Communication Technologies for Ageing Well and e-Health

Apr 6-8, 2025

10.5220/0013415700003938

Rottenberg

Gross

Gotlib

Emotion context insensitivity in major depressive disorder

J Abnorm Psychol2005111144627639

10.1037/0021-843X.114.4.627

16351385

Trifu

Nemeș

Herta

Bodea-Hategan

Talaș

Coman

Linguistic markers for major depressive disorder: a cross-sectional study using an automated procedure

Front Psychol2024151355734

10.3389/fpsyg.2024.1355734

38510303

Hartnagel

Ebner-Priemer

Foo

Linguistic style as a digital marker for depression severity: an ambulatory assessment pilot study in patients with depressive disorder undergoing sleep deprivation therapy

Acta Psychiatr Scand2025031513348357

10.1111/acps.13726

38987940

Han

Feng

Correlation between word frequency and 17 items of Hamilton scale in major depressive disorder

Front Psychiatry202213902873

10.3389/fpsyt.2022.902873

35592381

Shinohara

Nakamura

Omiya

Depressive mood assessment method based on emotion level derived from voice: comparison of voice features of individuals with major depressive disorders and healthy controls

Int J Environ Res Public Health2021051918105435

10.3390/ijerph18105435

34069609

Gehricke

Shapiro

Reduced facial expression and social context in major depression: discrepancies between facial muscle activity and self-reported emotion

Psychiatry Res20000821952157167

10.1016/s0165-1781(00)00168-2

10963801

Cariola

Sheeber

Allen

Language use in depressed and non-depressed mothers and their adolescent offspring

J Affect Disord2024121366290299

10.1016/j.jad.2024.08.131

39187178

Kouba

Šmotek

Tichý

Kopřivová

Detection of air traffic controllers’ fatigue using voice analysis - an EEG validation study

Int J Ind Ergon20230595103442

10.1016/j.ergon.2023.103442

Torous

Topol

Assessing generative artificial intelligence for mental health

Lancet202540610504

10.1016/S0140-6736(25)01237-1

Cruciani

Fritsch

Cleland

Rey

Nugent

Lukowicz

Few-shot human activity recognition using lightweight language models

2025 International Conference on Activity and Behavior Computing

Apr 21-25, 2025

10.1109/ABC64332.2025.11118559

Bedi

Liu

Orr-Ewing

Testing and evaluation of health care applications of large language models: a systematic review

JAMA202501283334319328

10.1001/jama.2024.21700

39405325

Zhang

Lee

Kraut

Mohr

Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being

NPJ Digit Med2023121961236

10.1038/s41746-023-00979-5

38114588

Laranjo

Dunn

Tong

Conversational agents in healthcare: a systematic review

J Am Med Inform Assoc201809125912481258

10.1093/jamia/ocy072

30010941

Rizzo

Lucas

Gratch

Automatic behavior analysis during a clinical interview with a virtual human

Stud Health Technol Inform2016220316322

27046598

Cohen

Richter

Neumann

A multimodal dialog approach to mental state characterization in clinically depressed, anxious, and suicidal populations

Front Psychol2023141135469

10.3389/fpsyg.2023.1135469

37767217

Jiang

Seyedi

Griner

Multimodal mental health digital biomarker analysis from remote interviews using facial, vocal, linguistic, and cardiovascular patterns

IEEE J Biomed Health Inform20240328316801691

10.1109/JBHI.2024.3352075

38198249

Milintsevich

Sirts

Dias

Towards automatic text-based estimation of depression through symptom prediction

Brain Inform202302131014

10.1186/s40708-023-00185-9

36780049

Yang

Qian

Conversational agent interventions for mental health problems: systematic review and meta-analysis of randomized controlled trials

J Med Internet Res2023042825e43862

10.2196/43862

37115595

Ćosić

Popović

Wiederhold

Enhancing aviation safety through AI-driven mental health management for pilots and air traffic controllers

Cyberpsychol Behav Soc Netw202408278588598

10.1089/cyber.2023.0737

38916063

Wiederhold

From cyberspace to outer space: how technology will preserve mental health in the final frontier

Cyberpsychol Behav Soc Netw20250128113

10.1089/cyber.2024.0609

39823387

Gatti

Palumbo

Di Domenico

Mammarella

Affective health and countermeasures in long-duration space exploration

Heliyon2022051385e09414

10.1016/j.heliyon.2022.e09414

35607498

Yin

Liu

Fan

Long-term spaceflight composite stress induces depression and cognitive impairment in astronauts-insights from neuroplasticity

Transl Psychiatry2023118131342

10.1038/s41398-023-02638-5

37938258

Sobolev

Menvielle

Sanora: a conversational AI agent for multimodal digital biomarkers of mental health

TechRxiv2026-03-22Preprint posted online on Feb 10, 2026

https://www.techrxiv.org/users/1027431/articles/1387167-sanora-a-conversational-ai-agent-for-multimodal-digital-biomarkers-of-mental-health

Kroenke

Spitzer

Williams

The PHQ-9: validity of a brief depression severity measure

J Gen Intern Med200109169606613

10.1046/j.1525-1497.2001.016009606.x

11556941

Okuyama

Akechi

Kugaya

Development and validation of the Cancer Fatigue Scale

J Pain Symptom Manage200001191514

10.1016/S0885-3924(99)00138-4

10687321

Dahmen

Cook

Fellows

Schmitter-Edgecombe

An analysis of a digital variant of the Trail Making Test using machine learning techniques

Technol Health Care2017252251264

10.3233/THC-161274

27886019

Girard

Cohn

Mahoor

Mavadati

Hammal

Rosenwald

Nonverbal social withdrawal in depression: evidence from manual and automatic analysis

Image Vis Comput2014103210641647

10.1016/j.imavis.2013.12.007

25378765

Scherer

Stratou

Mahmoud

Automatic behavior descriptors for psychological disorder analysis

10th IEEE International Conference and Workshops on Automatic Face and Gesture Recognition

Apr 22-26, 2013

10.1109/FG.2013.6553789

Girard

Cohn

Mahoor

Mavadati

Rosenwald

Social risk and depression: evidence from manual and automatic facial expression analysis

International Conference on Automatic Face and Gesture Recognition 2013

Apr 22-26, 2013

10.1109/FG.2013.6553748

Menne

Dörr

Schräder

The voice of depression: speech features as biomarkers for major depressive disorder

BMC Psychiatry20241112241794

10.1186/s12888-024-06253-6

39533239

Wang

Liang

Zhang

Fast and accurate assessment of depression based on voice acoustic features: a cross-sectional and longitudinal study

Front Psychiatry2023141195276

10.3389/fpsyt.2023.1195276

37415683

Cummins

Sethu

Epps

Krajewski

Probabilistic acoustic volume analysis for speech affected by depression

Interspeech 2014

Sep 14-18, 2014

10.21437/Interspeech.2014-311

Shinohara

Toda

Nakamura

Evaluation of emotional arousal level and depression severity using voice-derived sound pressure change acceleration

Sci Rep2021063011113615

10.1038/s41598-021-92982-7

34193915

Mundt

Vogel

Feltner

Lenderking

Vocal acoustic biomarkers of depression severity and treatment response

Biol Psychiatry2012101727580587

10.1016/j.biopsych.2012.03.015

22541039

Mundt

Snyder

Cannizzaro

Chappie

Geralts

Voice acoustic measures of depression severity and treatment response collected via interactive voice response (IVR) technology

J Neurolinguistics2007012015064

10.1016/j.jneuroling.2006.04.001

21253440

Yamamoto

Takamiya

Sawada

Using speech recognition technology to investigate the association between timing-related speech features and depression severity

PLoS One2020159e0238726

10.1371/journal.pone.0238726

32915846

Cannizzaro

Harel

Reilly

Chappell

Snyder

Voice acoustical measurement of the severity of major depression

Brain Cogn2004105613035

10.1016/j.bandc.2004.05.003

15380873

Sundelin

Lekander

Kecklund

Van Someren

Olsson

Axelsson

Cues of fatigue: effects of sleep deprivation on facial appearance

Sleep201309136913551360

10.5665/sleep.2964

23997369

Zargari Marandi

Madeleine

Omland

Vuillerme

Samani

Eye movement characteristics reflected fatigue development in both young and elderly individuals

Sci Rep20180938113148

10.1038/s41598-018-31577-1

30177693

Stern

Boyer

Schroeder

Blink rate: a possible measure of fatigue

Hum Factors199406362285297

10.1177/001872089403600209

8070793

Yin

Liu

Hao

Yang

Feng

Evaluation of VDT-induced visual fatigue by automatic detection of blink features

Sensors (Basel)20220125223916

10.3390/s22030916

35161662

Hasan

Kashevnik

State-of-the-art analysis of modern drowsiness detection algorithms based on computer vision

29th Conference of Open Innovations Association (FRUCT)

May 12-14, 2021

10.23919/FRUCT52173.2021.9435480

Yamada

Kobayashi

Detecting mental fatigue from eye-tracking data gathered while watching video: evaluation in younger and older adults

Artif Intell Med201809913948

10.1016/j.artmed.2018.06.005

30026049

Guggisberg

Mathis

Schnider

Hess

Why do we yawn?

Neurosci Biobehav Rev20100734812671276

10.1016/j.neubiorev.2010.03.008

20382180

Krajewski

Trutschel

Golz

Sommer

Edwards

Estimating fatigue from predetermined speech samples transmitted by operator communication systems

Fifth International Driving Symposium on Human Factors in Driver Assessment, Training and Vehicle Design

Jun 22-25, 2009

10.17077/drivingassessment.1359

de Vasconcelos

Vieira

Kecklund

Yehia

Speech analysis for fatigue and sleepiness detection of a pilot

Aerosp Med Hum Perform2019041904415418

10.3357/AMHP.5134.2019

30922431

Magliacano

Fiorenza

Estraneo

Trojano

Eye blink rate increases as a function of cognitive load during an auditory oddball paradigm

Neurosci Lett20200925736135293

10.1016/j.neulet.2020.135293

32771601

Fischer

Breitmeyer

Mechanisms of visual attention revealed by saccadic eye movements

Neuropsychologia1987251A7383

10.1016/0028-3932(87)90044-3

3574652

Bloch

Aviram

Neeman

Methylphenidate mediated change in prosody is specific to the performance of a cognitive task in female adult ADHD patients

World J Biol Psychiatry2015168635639

10.3109/15622975.2015.1036115

25945954

Holmlund

Cohen

Cheng

Using automated speech processing for repeated measurements in a clinical setting of the behavioral variability in the stroop task

Brain Sci2023034133442

10.3390/brainsci13030442

36979252

Haltaufderheide

Ranisch

The ethics of ChatGPT in medicine and healthcare: a systematic review on Large Language Models (LLMs)

NPJ Digit Med202407871183

10.1038/s41746-024-01157-x

38977771

Tartaglia

Jaghab

Ismail

Assessing health technology literacy and attitudes of patients in an urban outpatient psychiatry clinic: cross-sectional survey study

JMIR Ment Health2024123011e63034

10.2196/63034

39753220

Sobolev

Vitale

Wen

The Digital Marshmallow Test (DMT) diagnostic and monitoring mobile health app for impulsive behavior: development and validation study

JMIR Mhealth Uhealth2021012291e25018

10.2196/25018

33480854

Radanliev

Santos

Brandon-Jones

Joinson

Ethics and responsible AI deployment

Front Artif Intell202471377011

10.3389/frai.2024.1377011

38601110

Trocin

Mikalef

Papamitsiou

Conboy

Responsible AI for digital health: a synthesis and a research agenda

Inf Syst Front20231225621392157

10.1007/s10796-021-10146-4

Chaturvedi

Angrisani

Troxel

American Life in Realtime: benchmark, publicly available person-generated health data for equity in precision health

PNAS Nexus202510410pgaf295

10.1093/pnasnexus/pgaf295

41064239

Wen

Sobolev

Vitale

mPulse mobile sensing model for passive detection of impulsive behavior: exploratory prediction study

JMIR Ment Health2021012781e25019

10.2196/25019

33502330

Multimedia Appendix 1

Feature list.

Multimedia Appendix 2

Survey responses.