Real-World Smartphone Data Predicts Mood After Ischemic Stroke and Transient Ischemic Attack Symptoms and May Constitute Digital Endpoints: A Proof-of-Concept Study

Abstract

Objective

To assess the feasibility of using smartphones to longitudinally collect objective behavior measures and establish the extent to which they can predict gold-standard depression severity in patients with ischemic stroke and transient ischemic attack (IS/TIA) symptoms.

Patients and Methods

Participants with IS/TIA symptoms were monitored in real-world settings using the Beiwe application for 8 or more weeks during March 1, 2024 to November 15, 2024. Depression symptoms were tracked via weekly Patient Health Questionnaire (PHQ)-8 surveys, monthly personnel-administered Montgomery–Åsberg Depression Rating Scale (MADRS) assessments, and weekly averages of smartphone sensor measures. Repeated measures correlation established associations between PHQ-8 scores and objective behavior measures. To investigate how closely smartphone data predicted MADRS scores, linear mixed models were used.

Results

Among enrolled participants (n=54), 35 completed the study (64.8%). PHQ-8 scores were associated with distance from home (r=0.173), time spent at home (r=−0.147) and PHQ-8 administration duration (r=0.151). Using demographic data and the most recent PHQ-8 scores, average root-mean-squared error for depression severity prediction across models was 1.64 with only PHQ-8 scores, 1.49 also including accelerometer and GPS data, and 1.36 also including PHQ-8 administration duration.

Conclusion

Smartphone sensors captured objective behavior measures in patients with IS/TIA. In predictive models, the accuracy of depression severity scores improved as measures from additional smartphone sensors were included. Future research should validate this decentralized, exploratory approach in a larger cohort. Our work is a step toward showing that real-world monitoring with active and passive data may triage patients with IS/TIA for efficient depression screening and provide digital mobility and response time endpoints.

Post-stroke depression (PSD) is an established risk factor for poor outcomes.¹ Although one-third of stroke survivors develop PSD, screening for PSD is not routine.² While PSD symptoms often improve in the first few months, most patients with PSD experience recurrent depression in subsequent years.³ Its diagnosis is further complicated by characteristic post-stroke fatigue and neurologic deficits that can complicate longitudinal monitoring and follow-up.⁴

The recent reclassification of stroke in the International Classification of Diseases, 11th edition, as a neurologic disorder, rather than a circulatory one, has sparked interest in studying populations with overlapping behavioral symptoms, including related diagnoses of transient ischemic attack (TIA); yet, the range of symptoms associated with individual PSD cases complicates the designation of a tool for PSD screening.4, 5, 6 Despite the introduction of several such tools, they are rarely used in practice.7, 8, 9 Historically, the identification of PSD has relied on observation by clinical staff with specialized training in stroke-related psychiatric conditions; however, questions from the Patient Health Questionnaire (PHQ) have emerged as reliable longitudinal measures of PSD, helping triage patients for diagnostic assessment.10, 11, 12, 13

Commonly used in clinical practice, self-reported surveys, like the PHQ, reflect patient self-perception, and objectively measured data from wearable devices may augment or supplement these surveys.¹⁴ Powered by digital sensors in activity trackers and other devices, novel data streams of passive measurements are actively being applied to exploratory biomarker research in neurology.¹⁵ Research evaluating measures from digital sensors to better phenotype and predict disease trajectories with neurologic conditions is on the rise, and opportunities exist to validate digital solutions in routine clinical practice.¹⁶ Considering that smartphone ownership is nearly ubiquitous, modern smartphones may be a useful wearable for remote monitoring, capturing high-quality data from built-in sensors.¹⁷ To date, a growing number of studies have demonstrated the feasibility of using wearable accelerometers to assess post-stroke function outside of clinical settings, but only a few have studied PSD, primarily focused on ischemic stroke (IS).17, 18, 19, 20

One such study by Bui et al²¹ required participants (n=212; 90% mild-moderate IS cases) to report activity and mood surveys via a smartphone application 5 times per day over a 2-week period. Of the 95% of participants who completed the study, most completed surveys at home (70%).²² Lau et al²³ combined real-world smartphone and wearable accelerometer monitoring to deliver 1 week of surveys, 8 times per day, to patients with stroke (n=40; 70% IS cases), finding that accelerometer-derived physical activity (PA) was associated with self-reported positive affect and, therefore, may be a behavioral proxy for mood.²¹ The results by Ashizawa et al²⁴ expand on the potential role of objectively measured PA as a proxy for mood. In a cohort of patients with minor IS (n=76), they applied a hybrid monitoring model, using a belt-worn accelerometer to measure in-hospital PA and delivered the Geriatric Depression Scale-15 survey to participants 3 months after discharge. They found direct and inverse relationships between depression severity and sedentary and light PA, respectively.²⁴ Błaszcz et al²⁵ also used a hybrid study design, instructing participants (n=21) to wear an accelerometer on their hips for 6 weeks and complete in-person Hospital Anxiety and Depression Scale assessments. In contrast to the results by Lau et al²³ and Ashizawa et al,²⁴ they found that although PA increased each week after hospital discharge, no change in depression was observed from baseline to study conclusion. Although these studies employed dissimilar methodologies, the results suggest that frequent assessment of mood in patients with stroke is feasible outside of clinical environments for short time periods, yet the potential for objectively measured behaviors, such as PA, to serve as proxies for mood is unclear.

Compared with traditional forms of remote stroke care, such as tele-stroke visits, little is known about novel wearable sensors, such as global positioning system (GPS), applied to stroke care.²⁶ More importantly, the translational potential of extracting useful information from these sensors, such as via the creation of digital study endpoints, remains largely unexplored. To date, no research on stroke or TIA has evaluated the potential of objectively measured data from real-world settings to augment clinical workflows via longitudinal monitoring, such as by predicting gold-standard depression severity scores.

The objectives of this pilot study were to (1) evaluate the feasibility of collecting objective behavior measures from smartphone accelerometer, GPS, and screen touch sensors and (2) assess the accuracy of those measures at predicting depression severity in patients with IS/TIA symptoms. Knowing that the longest comparable study design collected 6 weeks of wearable accelerometer data, we sought to extend the literature to assess the feasibility of collecting data over an 8-week study.²⁴ The scientific rationale underpinning this work is that self-reported mood is closely associated with functional ability in patients with stroke.²⁷

Patients and Methods

Study Design and Population

This study used a prospective cohort design to recruit adult participants (18 years and older) recently admitted to Mayo Clinic Hospital in Arizona with IS/TIA symptoms. Patients were screened for eligibility using electronic health records (EHRs) and recruited via email after discharge between March 1 and August 31, 2024. We recruited a cohort of participants with IS/TIA symptoms at hospital admission as well as those participating in sequelae monitoring with a previous IS/TIA diagnosis. Discharge diagnoses were confirmed by imaging results based on diagnostic criteria. Participants had to own a smartphone and were included if they had one of the following discharge diagnoses: TIA, IS, or other transient stroke-like symptoms. Those with a previous dementia diagnosis were excluded. All participants provided informed consent digitally and were registered remotely. Mayo Clinic Institutional Review Board approved the study procedure (22-009345).

Given that the impact of IS on functional status is, on average, less severe with shorter recovery times than that of hemorrhagic stroke, we focused recruitment on the former cohort, ensuring that mobility data would be generated for analysis.²⁸ Additionally, we included patients with TIA because TIA is often studied in combination with IS in the literature, due to its shared etiology, and does not permanently impair mobility.29, 30, 31, 32 Patients with suspected IS/TIA at hospital admission but whose discharge diagnosis was a non-stroke cerebrovascular syndrome were also included. Considering the lack of comparable studies and knowledge about how patients would receive such a study design, this study was launched as a proof-of-concept, offering a starting point for future predictive mood monitoring in cerebrovascular disease departments.

Data Collection

Demographic data were obtained from EHRs. Participants downloaded the HIPAA-compliant Beiwe application at baseline. Passive accelerometer, GPS, and screen touch sensor data were recorded along with remote self-entry of the PHQ-8 survey (score 0-24) using the application. Beiwe is an open-source smartphone platform for digital phenotyping compatible with both iOS and Android operating systems that includes Health Insurance Portability and Accountability Act–compliant data storage and the Forest library of data processing packages, generating measures from sensor streams (Table 1).³³ Study data were collected for 8 weeks, with the platform capturing up to 90 days of data to accommodate patient delays in scheduling Montgomery–Åsberg Depression Rating Scale (MADRS) assessments, as 7 days of passive data prior to a MADRS assessment were required for the predictive models (Supplemental Figure 1, available online at https://www.mcpdigitalhealth.org/).

Table 1.

Predictors Used in Linear Mixed Models

Baseline predictors	Active data predictors	Passive data predictors
Age	Antecedent PHQ-8 survey score	GPS (number of significant locations visited, significant location entropy, time spent at home, distance traveled, maximum diameter of travel, maximum travel distance from home, radius of gyration, average flight length, SD of flight length, average flight duration, SD of flight duration, total time spent in pause/stationary, average stationary/pause time, SD of stationary/pause time, and hours per day without GPS data)
Sex		Accelerometer (cadence, walking time, number of steps, and hours per day without accelerometer data)
Diagnostic category		Touch screen (time spent completing antecedent PHQ-8 survey, or PHQ-8 administration duration)
Previous depression diagnosis

The application sent a notification with the survey to participant phones at baseline and once per week (Supplemental Figure 2, available online at https://www.mcpdigitalhealth.org/). As continuous data streaming is not a standard approach for any wearable device study, due to battery drain, the sampling frequencies for smartphone sensors were determined based on previous studies using the application.¹⁷ Accelerometer sensors streamed data for 10 seconds on/10 seconds off, and GPS data were streamed for 1 minute on/10 minutes off.

Every 30 days, participants completed a virtual interview, during which personnel administered a MADRS assessment (score 0-60). Participants whose smartphone stopped streaming data or who were no longer responsive were classified as lost to follow-up. The application was uninstalled upon study completion.

Smartphone Sensor Data Processing

Accelerometer data were cleaned using the Oak algorithm, previously validated via 20 public datasets, in the open-source Forest Python library. The GPS data were processed using the Jasmine algorithm in the Forest library. This algorithm applies a validated quality control feature, filtering participants with suboptimal data collected, including those with location coordinates characterized by <50 m of horizontal accuracy.^33,34 Information about each GPS measure derived and its processing protocol is available in Supplemental Table 1 (available online at https://www.mcpdigitalhealth.org/). A summary table of data completeness by sensor for each participant is in Supplemental Table 2 (available online at https://www.mcpdigitalhealth.org/).

Statistical Analyses

First, we evaluated the feasibility of using smartphone sensor data to capture objective measures of behavior over time. Applying the approach by Wu et al,³⁵ we obtained correlations for repeated measures between weekly average smartphone sensor measures and PHQ-8 survey scores. Repeated measures correlation determines common within-subject associations using analysis of covariance.³⁵ All sensor measures were the average of 7 days prior to a survey.

Then, we assessed the potential of smartphone sensor measures to predict gold-standard depression severity scores using linear mixed models. Linear mixed models are uniquely able to use longitudinal study data with clusters of missing data, handling different numbers of repeated measures across participants and uneven time periods between repeated measures. Using a modified approach to the study focused on psychiatry patients by Pellegrini et al,³⁶ we developed 5 models with 2 variations—with and without demographic data—for predicting MADRS scores. Baseline characteristics along with study compliance metrics were obtained. Considering that the passive measures correlated with one another, principal component analysis (PCA) was used to obtain a principal component (PC) predictor (PC1) from weekly averages across measures. An antecedent survey score was defined as the most recent weekly PHQ-8 score submitted within 7 days before an MADRS assessment. Missingness variables for accelerometer and GPS sensors described earlier were included in PCA.

Before fitting the model, any instance with 1 or more missing predictor values was excluded. Leave-one-subject-out cross-validation was performed for each participant, with the model fitted with data from other participants. This required that all participants complete 2 or more MADRS assessments. Then, MADRS scores for the excluded participant were predicted with the model. Model 1 used only the antecedent survey score to predict MADRS scores; model 2 used only passive data (PC1 from weekly average); model 3 used both the survey score and PC1; model 4 applied the predictors from model 3 plus the open survey time (screen sensor) predictor; and model 5 used no predictors. Tests confirming linear mixed model assumptions were met are detailed in Supplemental Appendix 1 (available online at https://www.mcpdigitalhealth.org/).

To assess the fit of each model, the root-mean-squared error (RMSE) for each participant was calculated, taking the square of the error between the predicted MADRS and actual MADRS scores for each MADRS assessment, and averaged across participants. For a given model, lower RMSE values correspond to better model fit.

Sensitivity Analysis

We compared the antecedent smartphone survey scores with staff-administered MADRS scores, using Spearman's rank-order correlation coefficient across MADRS timepoints. In addition, we repeated the model development process using only accelerometer data to compare single and multisensor approaches.

Results

Participant Characteristics

Of the 54 participants enrolled (Figure), 44 downloaded the application (81.5%). Of the 38 participants who completed the initial MADRS assessment, 35 completed the study. Study participants were primarily men (59.1%; n=26), middle-aged (57 years), White (88.6%; n=39), and iPhone users (77.3%; n=34). Only 7.0% of participants reported Hispanic ethnicity (n=3). At baseline, most participants had 1 or more of the following diagnoses: heart disease (65.9%; n=27), diabetes (24.4%; n=10), anxiety (22.0%; n=9), and depression (26.8%; n=11). More than 35% had experienced multiple IS/TIAs. At the time of hospital admission or most recent sequelae monitoring visit, participants experienced a wide range of stroke-related symptoms, with one-sided weakness (20.5%; n=9), dizziness (15.9%; n=7), and numbness (15.9%; n=7) among the most common (Table 2).

Figure.

Consort diagram of study enrollment and retention. IS, ischemic stroke; MADRS, Montgomery–Åsberg Depression Rating Scale; TIA, transient ischemic attack.

Table 2.

Baseline Demographic Characteristics for Participants

Characteristic	Participants (n = 44)
Sex
Male	59.1 (26)
Female	40.9 (18)
Age (y)
Mean (SD)	57
Minimum, Q1, Q3, maximum	23, 49, 68, 83
Discharge diagnosis
TIA	22.7 (10)
Ischemic stroke	22.7 (10)
Other transient stroke-like symptoms	34.1 (15)
Previous ischemic stroke/TIA only	20.5 (9)
Race
White	88.6 (39)
Native American	2.3 (1)
Did not disclose	5.8 (3)
Asian	2.3 (1)
Hispanic ethnicity	7.0 (3)
Smartphone operating system
iOS	77.3 (34)
Android	22.7 (10)
Baseline PHQ-8, mean ± SD
TIA	5 ± 4
Ischemic stroke	3 ± 3
Other transient stroke-like symptoms	7 ± 6
Previous ischemic stroke/TIA only	7 ± 4
Heart disease	65.9 (27)
Diabetes	24.4 (10)
Multiple strokes	36.6 (15)
Depression	26.8 (11)
Anxiety	22.0 (9)
IS/TIA symptoms
Vision loss	9.1 (4)
Numbness	15.9 (7)
Headache/migraine	13.6 (6)
Tremor	2.3 (1)
Aphasia	13.6 (6)
Facial droop	11.4 (5)
Memory loss	13.6 (6)
Dizziness	15.9 (7)
One-sided weakness	20.5 (9)
Confusion	2.3 (1)
Loss of consciousness	2.3 (1)

Values are % (n) unless specified.

Survey completion rates over time were moderately high, with 73% of participants completing 8 or more weeks of PHQ-8 surveys (Supplemental Figure 3, available online at https://www.mcpdigitalhealth.org/). Individual PHQ-8 trajectories were highly variable (Supplemental Figure 4, available online at https://www.mcpdigitalhealth.org/). Two participants presented at the hospital for stroke symptoms during the study (Supplemental Table 3, available online at https://www.mcpdigitalhealth.org/). Additional information about post-stroke disabilities experienced by participants is available (Supplemental Appendix 2).

Feasibility of Capturing Behavior via Smartphone Sensors

Before analysis, 1 participant was excluded because of sampling below the quality check threshold for the Forest package. Using repeated measures correlation, we assessed relationships between smartphone sensor measures and survey scores (Table 3). PHQ-8 scores correlated positively with PHQ-8 administration duration (r=0.151) and distance from home (r=0.173) and negatively with time spent at home (r=−0.147). Relationships between individual PHQ-8 behaviors and passive measures are explained in detail in Supplemental Appendix 3 (available online at https://www.mcpdigitalhealth.org/).

Table 3.

Repeated Measures Correlations for PHQ-8 Survey Score and Smartphone Measures

Smartphone sensor measure	Correlation coefficient (r)	P	95% CI
PHQ-8 survey administration duration	0.151	0.004∗∗	(0.048-0.251)
Distance diameter	0.013	0.85	(−0.115 to 0.140)
Distance from home	0.173	0.008∗∗	(0.057-0.294)
Distance traveled	0.014	0.83	(−0.114 to 0.141)
Flight distance, average	0.008	0.90	(−0.119 to 0.135)
Flight distance, SD	0.021	0.75	(−0.106 to 0.148)
Flight duration, average	−0.066	0.31	(−0.192 to 0.062)
Flight duration, SD	−0.094	0.15	(−0.219 to 0.034)
Time spent at home	−0.147	0.024∗	(−0.269 to −0.020)
Gyration radius	0.006	0.92	(−0.121 to 0.134)
Significant location count	−0.055	0.40	(−0.181 to 0.073)
Significant location entropy	−0.070	0.28	(−0.196 to 0.058)
Pause time	−0.081	0.21	(−0.207 to 0.047)
Total flight time	−0.054	0.41	(−0.180 to 0.074)
Pause duration, average	−0.014	0.83	(−0.142 to 0.113)
Pause duration, SD	0.025	0.70	(−0.103 to 0.152)
Walking time	−0.026	0.65	(−0.137 to 0.085)
Steps	−0.028	0.62	(−0.139 to 0.083)
Cadence	−0.024	0.67	(−0.135 to 0.087)

Depression Severity Prediction Models

After excluding 15 participants with missing data for 1 or more measures, primarily due to GPS measures, PCA generated PC1, the passive data predictor in models (Supplemental Table 4, available online at https://www.mcpdigitalhealth.org/). Compared to the model variation without demographic variables, RMSE values were lower for the variation including demographic variables (variation 2), indicating better fit (Table 4). The lowest RMSE was 1.36 (model 4 with demographic data), which included accelerometer, GPS, screen sensor data, and the antecedent survey score. While variation 1 showed no clear trends, including passive smartphone data in variation 2’s model 1 (survey only) yielded a modest RMSE improvement of 0.15. Also, in variation 2, MADRS score prediction accuracy was improved by 0.13 with the addition of a measure from smartphone screen sensors: PHQ-8 administration duration.

Table 4.

RMSE Predicting MADRS Scores Using Models 1-5 With 2 Variations: No Demographic Characteristics and With Demographic Characteristics

Model	Variation 1 (no demographic characteristics)	Variation 2 (demographic characteristics)
Model 1 (PHQ-8 score)	1.70	1.64
Model 2 (Accelerometer + GPS)	1.70	1.56
Model 3 (Accelerometer + GPS + PHQ-8 score)	1.61	1.49
Model 4 (Accelerometer + GPS + PHQ-8 score + PHQ-8 survey administration duration)	1.60	1.36
Model 5 (only participant effect)	1.73	1.65

Sensitivity Analysis

In aggregate, antecedent PHQ-8 scores moderately correlated with personnel-administered MADRS scores (r=0.667), a finding supported by literature (Supplemental Figures 5 and 6a-c, available online at https://www.mcpdigitalhealth.org/). For the accelerometer-only MADRS prediction models, no trends were observed (Supplemental Table 5, available online at https://www.mcpdigitalhealth.org/).

Discussion

To the best of our knowledge, this is the first study to investigate the application of smartphone monitoring with GPS and accelerometer sensors in a population with IS/TIA.³⁷ It offers feasibility data for smartphone-based monitoring in patients with IS/TIA. We demonstrated that longitudinal measures derived from multiple smartphone sensors may capture objective measures of behavior and that those measures appear to predict depression severity in patients with IS/TIA. Some of those measures also correlate with self-reported mood, offering a first step for future researchers to further develop digital endpoints from smartphone measures. Though its results must be interpreted as preliminary, this type of monitoring program could one day use objective measures as proxies to detect clinically meaningful aberrations in mood in patients who otherwise would not receive in-person screening or complete surveys after hospital discharge, but such an analysis is beyond the scope of this study.

Compared to the model using only survey responses, the greatest improvement in model fit was found when measures from 3 smartphone sensors (accelerometer, GPS, and screen touch) were included, improving RMSE by 0.28. This preliminary finding highlights the potential for multimodal sensors combined with self-reported surveys to offer a more comprehensive observation of patients in future real-world studies. The improved prediction performance of passive data extends other proof-of-concept work, including the 1-week wearable accelerometer study on patients with mild stroke by Lau et al²³ and the 12-month Beiwe study with nurses by Yi et al,³⁸ which found that self-reported mood and daily activities correlate strongly with accelerometer measures.^39,40 In contrast to the study by Pellegrini et al,³⁶ which evaluated the performance of GPS and accelerometer data in predicting MADRS scores in a transdiagnostic population of psychiatric patients, we found that passive data did improve MADRS score prediction in an IS/TIA cohort; however, the range of PHQ-8 and MADRS scores in our cohort was smaller than theirs, as reflected in the lower RMSE values for our models.³⁶

A key strength of this study is that it shows the potential for smartphone sensors to collect passive and self-report data related to mood in an IS/TIA cohort for a minimum of 8 weeks. Also, we exclusively examined smartphone accelerometer data in our sensitivity analysis and found improved model fit when GPS and screen touch sensor data were included, creating an early-stage rationale for future researchers to use multiple sensors to capture behavior in patients with IS/TIA. Another strength of this study is the inclusion of participants with stroke-related disabilities, although TIA symptoms were resolved by study enrollment or during participation. Though previous work has demonstrated the feasibility of smartphone monitoring in memory loss, migraine, and tremor cohorts, our findings extend these results to patients with stroke-related disabilities and other transient neurologic symptoms, such as numbness, dizziness, and one-sided loss of function.41, 42, 43, 44

While this work should be considered as a starting point for future research, it is subject to numerous limitations. First, the pilot study design includes a small sample size, although comparable with other smartphone studies. Second, passive measures in this analysis were not processed to consider weekday versus weekend averages. Next, participants most likely did not carry their phones constantly. As such, differences in participant data could be partly driven by phone use patterns. Also, behaviors assessed via self-report are limited by the short set of PHQ-8 questions. Common in real-world studies, missing data due to missed surveys and sick days occurred, but interpretation of GPS data should be considered with extra caution, because of the longer off cycles required to prevent excessive battery drain.^17,45 Although the Forest package applies validated imputation methods to approximate sensor data points for off cycles, average measures may be biased when proportions of collected data are low (Supplemental Table 2).³⁶ Given that previous research found GPS and accelerometer data missingness levels predictive of symptoms in a cohort of patients with schizophrenia, researchers should explore similar hypotheses in IS/TIA cohorts.³² Also, GPS measures were the primary source of missing measures in the week before a MADRS assessment, potentially biasing the results by excluding patients with older smartphones and unreliable cellular data networks. In addition to enrolling larger cohorts, future research should explore how to design study protocols that reduce data missingness, perhaps by testing whether alternative GPS sensors, such as those in the Apple Watch, capture more complete data in some cohorts with IS/TIA.⁴⁶

Considering participant demographic characteristics, these results may not generalize to non-White patients. As most participants had an iOS smartphone, data may be biased in ways unnoticed. Beyond issues associated with wireless and cellular access in rural areas in Arizona, participant travel for vacation or work to different cities may also have introduced unforeseen variations in data. Although not a limitation, we chose to take the average of 7 days for each measure in our analyses, as is common in the growing set of studies in this space; however, other time periods may be more useful at predicting scores.³⁷ Lastly, it is important to consider that the improved model fit offered by the inclusion of passive measures, compared to self-reported scores alone, is small (RMSE decrease of 0.28) and may be more useful if applied to predicting scores during weeks of lapsed survey compliance rather than for screening purposes. A larger sample population is necessary to investigate the optimal role of remote monitoring with smartphones in this context and to validate the significant correlates of mood (distance from home, time spent at home, and PHQ-8 administration duration) as digital mobility and response time endpoints.

Conclusion

Our findings suggest that real-world data from smartphone sensors combined with self-reported data can (1) capture objective behavior measures and (2) approximate gold-standard depression screening scores in patients with IS/TIA. Repeated smartphone sensor measures, including distance from home, time spent at home, and PHQ-8 administration duration, correlated significantly with self-reported mood. Passive GPS, accelerometer, and screen sensor data slightly improved the prediction of gold-standard depression severity scores compared to models using fewer sensors or only patient-reported surveys. Our preliminary work lays a foundation for using mobility, spatial, and mobile device screen data to investigate environmental and behavioral factors influencing mood in patients with IS/TIA. Future research should involve larger cohorts to identify which measures are most meaningful for mood prediction, to compare the effectiveness of different combinations of wearable sensors and devices, and to establish clinically useful thresholds for screening patients with depression symptoms at different timepoints in recovery.

Potential Competing Interests

Given their role as Editorial Board Members, Drs Erickson and Demaerschalk had no involvement in the peer-review of this article and have no access to information regarding its peer-review.

Ethics Statement

The institutional review board at Mayo Clinic approved the study (22-009345). This study was conducted in compliance with the Declaration of Helsinki regarding ethics principles for medical research involving human subjects. Patients were informed of their right to object to the use of their data for the present study and gave their informed consent.