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

AI Summary

Personalized logistic regression models significantly predicted next-week panic attack risk, outperforming generalized models and exceeding chance-level PR AUC (lift 1.17, p = .035).
Generalised pooled model did not differ from chance (PR AUC lift 0.99), highlighting advantage of individualised models.
Passively collected daily sleep, activity and physiological data from wearable rings hold signal for predicting panic attacks, enabling tailored lifestyle interventions.

TL;DR (AI Generated)

We assessed multiple prediction algorithms applied within both personalized and generalized frameworks, fitting unique models for each participant and pooling data, respectively.

Across all 39 participants, personalized models using logistic regression achieved a pooled PR AUC lift of 1.17 (95% CI: 1.05-1.25) over chance-level prediction (p =.035), whereas the generalized model did not differ from chance (PR AUC lift of 0.99; 95% CI: 0.90-1.08); in a direct comparison, personalized models outperformed the generalized model (lift ratio=1.18, p =.047).

Performance varied across participants: when personalized models were grouped by unsupervised clustering of test-set PR AUC lift, models clustered as well-fit (n = 8) showed a pooled PR AUC lift of 2.26 (95% CI: 1.50-2.89) and models clustered as poor-fit (n = 31) a lift of 1.09 (95% CI: 0.98-1.17).

Basic summary

Front Digit Health. 2026 Oct 30;8:1978650. doi: 10.3389/fdgth.2026.1978650. eCollection 2026.

ABSTRACT

Panic attacks (PAs) are unexpected episodes of intense fear and anxiety, accompanied by cognitive and physical symptoms that significantly impair day-to-day functioning. This study proposes a machine learning approach to predict PA risk in the upcoming week in individuals across mental health conditions, enabling timely and potentially preventive mental health intervention. We evaluate the approach in a longitudinal sample of behavioral and physiological data collected using Oura rings from 39 college students over six months. We assessed multiple prediction algorithms applied within both personalized and generalized frameworks, fitting unique models for each participant and pooling data, respectively. Across all 39 participants, personalized models using logistic regression achieved a pooled PR AUC lift of 1.17 (95% CI: 1.05-1.25) over chance-level prediction (p = .035), whereas the generalized model did not differ from chance (PR AUC lift of 0.99; 95% CI: 0.90-1.08); in a direct comparison, personalized models outperformed the generalized model (lift ratio=1.18, p = .047). Performance varied across participants: when personalized models were grouped by unsupervised clustering of test-set PR AUC lift, models clustered as well-fit (n = 8) showed a pooled PR AUC lift of 2.26 (95% CI: 1.50-2.89) and models clustered as poor-fit (n = 31) a lift of 1.09 (95% CI: 0.98-1.17). Because these groups were defined by test-set performance, they describe observed heterogeneity and require validation in independent data. Together, our findings point towards the potential value of personalized modeling approaches and suggest that passively collected daily measurements of sleep, activity, and physiology may carry signal for predicting PAs, which could ultimately be leveraged to identify individually tailored lifestyle interventions to reduce PA risk.

PMID:43031775 | PMC:PMC13759848 | DOI:10.3389/fdgth.2026.1978650

Document this CPD