Personalized Smart Ring Models Forecast Weekly Panic Attack Risks in College Students

A longitudinal study published in Frontiers in Digital Health demonstrates that personalized machine learning models tracking continuous physiological and behavioral data from commercial smart rings can predict an individual’s risk of experiencing a panic attack during the upcoming week. Conducted by researchers at the University of Vermont and the Wake Forest University School of Medicine, the secondary analysis evaluated wearable sensor metrics alongside daily symptom surveys collected across an academic year from college students. Rather than relying on generic population-wide trends, the research team constructed participant-specific prediction algorithms designed to capture individualized physiological baseline shifts. The primary personalized models, which utilized logistic regression, achieved an overall precision-recall area under the curve lift of 1.17 over chance-level prediction across the complete cohort of thirty-nine modeled participants. In direct contrast, a generalized model pooling data across all participants performed at chance level with an area under the curve lift of 0.99, establishing that individual-level modeling significantly outperformed the generalized approach.

Panic attacks are clinically defined as unexpected, sudden surges of intense fear and discomfort accompanied by cognitive distress and acute autonomic symptoms such as tachycardia, chest pain, hyperventilation, dizziness, and sweating. While panic attacks are formally recognized as diagnostic criteria within panic disorder, which affects roughly 2.7 percent of adults in the United States each year, epidemiological data indicates that more than ten percent of the domestic population experiences a panic attack annually. The Diagnostic and Statistical Manual of Mental Disorders classifies panic attacks as a transdiagnostic specifier occurring across a wide spectrum of psychiatric presentations, including major depressive disorder and generalized anxiety. Historically, clinical predictive modeling for panic episodes has concentrated nearly exclusively on cohorts with formal panic disorder diagnoses, leaving the substantial majority of individuals who experience recurrent panic attacks without panic disorder unexamined. The investigators sought to address this diagnostic gap by investigating whether passive monitoring could forecast panic risk across diverse mental health profiles in real-world settings.

The underlying data was obtained from the Lived Experience Measured Using Rings Study, known as LEMURS, which enrolled full-time first-year and second-year undergraduate students aged eighteen to twenty-four at the University of Vermont. Eligible participants wore an Oura ring continuously across the academic year to capture physiological and behavioral metrics, including nocturnal heart rate, heart rate variability, respiration rate, body temperature fluctuations, sleep architecture stages, and daily physical activity intensities. To gather ground-truth outcome data while limiting respondent fatigue, the protocol deployed daily smartphone surveys within four separate twenty-one-day periods termed sprints, separated by six-week gaps. On these daily surveys, participants endorsed whether they experienced a panic attack based on their own conceptual understanding of the term. Diagnostic status was established through in-person Mini International Neuropsychiatric Interviews, which revealed that among participants with documented diagnostic data, over ninety percent met criteria for at least one psychiatric condition, encompassing panic disorder, anxiety disorder, and depression.

From an initial sample of one hundred ninety-seven participants reporting at least one panic attack during the daily survey windows, strict analytical filtering criteria were applied to ensure valid model training and evaluation. One hundred forty-eight individuals were excluded, predominantly because their chronological train or test data lacked representation of both classes, had fewer than two complete survey sprints, or contained insufficient total observations to compute precision-recall curves. The final analytical sample comprised thirty-nine college students who reported at least two panic attacks while wearing the sensor. Because panic attacks are episodic and rare, class balance between non-panic days and panic days exhibited marked variation across subjects, averaging a ratio of 2.12 with an observed range from 0.51 to 5.14. The cohort had an average age of 18.6 years, with women representing 91.9 percent of the sample where sex was recorded, and White individuals accounting for 83.8 percent of those reporting race and ethnicity.

To configure the forecasting task, the investigators extracted 148 statistical features from seven-day sliding windows of wearable measurements with a one-day step size, calculating the mean, standard deviation, and fifth and ninety-fifth percentiles for each behavioral and physiological metric. Outcome windows were defined as the seven-day period immediately following each feature window, labeled positive if at least one panic attack was recorded during that subsequent week. For every participant, features were standardized and selected using only the participant’s training sprints to prevent data leakage into the held-out final sprint test set. Model evaluation prioritized area under the precision-recall curve normalized as a lift metric against each individual’s baseline panic attack prevalence, establishing a uniform chance threshold of 1.0. This metric prevented distortions commonly caused by class imbalance and avoided the clinical arbitrariness of fixing hard classification thresholds for sensitivity and specificity.

In evaluating the primary logistic regression architecture across all thirty-nine students, the personalized framework produced a statistically significant pooled lift of 1.17 with a 95 percent confidence interval spanning 1.05 to 1.25, outperforming chance with an unadjusted p-value of 0.035. The pooled generalized logistic regression model produced a lift of 0.99 with a confidence interval between 0.90 and 1.08, failing to distinguish itself from chance at a p-value of 0.456. A direct comparison between the two modeling strategies demonstrated that the personalized pipeline significantly exceeded the generalized approach by a lift ratio of 1.18 at a p-value of 0.047. The researchers extended this comparison across three additional machine learning algorithms: support vector machines, extreme gradient boosting, and k-nearest neighbors. Although each algorithm showed personalized-to-generalized lift ratios greater than 1.0 across the full cohort, only the primary logistic regression formulation reached statistical significance.

Unsupervised one-dimensional k-means clustering applied to the test-set lift scores revealed marked heterogeneity in how well personalized models performed across individual participants. The clustering separated the thirty-nine participants into a well-fit subgroup of eight individuals and a poor-fit subgroup of thirty-one individuals. Among the well-fit cohort, personalized logistic regression models attained a pooled lift of 2.26 with a 95 percent confidence interval from 1.50 to 2.89, exceeding chance with a p-value of 0.001. Conversely, the poor-fit subgroup yielded a lift of 1.09 with a confidence interval of 0.98 to 1.17, which did not differ significantly from chance. The authors noted that because these groupings were formed directly from test-set metrics, the subgroup numbers represent descriptive observations of empirical variation rather than confirmatory evidence of inherent predictability differences. Descriptively, this twofold performance advantage of well-fit over poor-fit participants was replicated across all four machine learning algorithms, showing lift ratios ranging between 1.85 and 2.09.

The study also investigated whether self-reported psychological measures could match wearable sensors in predicting weekly panic episodes. When identical personalized models were trained using biweekly Depression Anxiety Stress Scales scores or daily tally counts of perceived life stressors, neither model achieved performance significantly above chance, yielding lifts of 1.08 and 0.90, respectively. To determine whether wearable metrics added incremental value over recent attack history, researchers implemented a persistence baseline using only the count and presence of panic attacks reported within the preceding seven-day feature window. While this baseline achieved a lift of 1.14 across the entire cohort and combining wearable features with persistence yielded a lift of 1.17, the authors explicitly emphasized that the difference between the combined and persistence-only models was not formally tested. Consequently, incremental predictive capacity beyond recent panic occurrence remains an open scientific question requiring formal evaluation.

Post-hoc interpretability analyses provided preliminary insight into the idiosyncratic nature of individual panic precursors. Computing the pairwise Kuncheva index across personalized models to quantify overlap among top predictive features produced a mean value of only 0.122, which was statistically indistinguishable from chance expectations under permutation testing. Lagged linear regressions conducted on representative individual subjects indicated that specific day-to-day behavioral shifts correlated with model-estimated risk fluctuations. For instance, in one modeled participant, increases in total nightly sleep relative to personal averages were associated with subsequent decreases in estimated panic likelihood, whereas in another participant, bedtime delays corresponded with heightened model-predicted risk. In an exploratory donor-recipient analysis across 248 pairings, five of the eight well-fit models significantly outperformed chance when evaluated against held-out data from at least one poor-fit participant, raising the hypothesis that poor individual model performance may partly stem from localized training constraints rather than a total absence of predictable physiological signal.

The investigators emphasized several important methodological and demographic limitations that contextualize their findings. Because the study sample was relatively small, geographically restricted to a single university, and predominantly composed of young, White female students, the findings cannot be generalized to broader demographic groups without external replication. In addition, panic attack occurrences relied on unadjudicated participant self-reporting without clinical verification, introducing subjective variability. Furthermore, while the models identified promising associations between wearable signals and weekly panic probabilities, false positives remained common, and the lagged analyses reflect internal mechanistic interpretations of fitted model parameters rather than clinical evidence that behavioral modifications will prevent panic attacks. The researchers concluded that while personalized passive sensing holds conceptual promise for informing future just-in-time adaptive interventions, extensive validation in larger, multi-device cohorts is essential before wearable prediction tools can be translated into clinical workflows.

Subject of Research: Personalized machine learning models for forecasting weekly panic attack risk using smart ring biometric data

Article Title: Development of personalized machine learning models for predicting panic attacks in college students

Article References: Kairamkonda, D. D., Cherian, J., Mascia, G., Hidalgo, J. E., Kim, J. N., Llorin, J., Stanton, K., Ha, J., Dodds, P. S., Price, M., Danforth, C. M., Sartini, J., McGinnis, E. W., & McGinnis, R. S. (2026). Development of personalized machine learning models for predicting panic attacks in college students. Frontiers in Digital Health, 8, Article 1978650. https://doi.org/10.3389/fdgth.2026.1978650

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1978650

Keywords: panic attacks, machine learning, digital phenotyping, wearable technology, mental health, predictive modeling, logistic regression, biometric monitoring, Development, personalized, machine, learning