Abstract
<sec> <title>BACKGROUND</title> <p>Stress-related disorders, including adjustment disorder and exhaustion disorder, are associated with disability, sickness absence, and heterogeneous treatment response. Internet-delivered interventions can improve stress-related symptoms, but tools for individualized prognosis are lacking.</p> </sec> <sec> <title>OBJECTIVE</title> <p>This study aimed to identify predictors of treatment response in patients with stress-related disorders and evaluate whether machine learning models could provide clinically useful individual-level prediction.</p> </sec> <sec> <title>METHODS</title> <p>We conducted a predictive modeling study using data from a randomized controlled trial of internet-delivered cognitive behavioral therapy versus general health promotion. Treatment arms were pooled because the parent trial found no between-group differences in efficacy. The outcome was responder status on the Perceived Stress Scale-10 at 12 weeks, defined using the reliable change index. Multivariable logistic regression examined prespecified predictors. Elastic net logistic regression, random forest, support vector machine, and AdaBoost models were trained using a 70/30 train-test split with 5-fold cross-validation and evaluated in an unseen hold-out test set. Clinical utility was prespecified as balanced accuracy of at least 67%.</p> </sec> <sec> <title>RESULTS</title> <p>Of 300 randomized participants, 282 had outcome data and 146 (51.8%) were responders. Higher baseline perceived stress (adjusted odds ratio [aOR] 2.03, 95% CI 1.50-2.74), higher quality of life (aOR 1.60, 95% CI 1.18-2.16), and higher educational attainment (aOR 1.57, 95% CI 1.14-2.15) predicted higher odds of response, whereas death of a close relative predicted lower odds (aOR 0.44, 95% CI 0.24-0.83). Elastic net logistic regression performed best in the per-protocol machine learning analysis (balanced accuracy 64.8%, 95% CI 54%-75%; area under the curve 0.74), but no per-protocol model met the prespecified clinical utility threshold. In a post hoc expanded-feature analysis, elastic net logistic regression reached balanced accuracy of 68.2% (95% CI 58%-78%; area under the curve 0.72).</p> </sec> <sec> <title>CONCLUSIONS</title> <p>Baseline clinical variables showed group-level prognostic value, but individual-level prediction remained modest. These findings support further development and external validation in larger samples before prediction models are used to guide routine care for stress-related disorders.</p> </sec> <sec> <title>CLINICALTRIAL</title> <p>ClinicalTrials.gov NCT04797273</p> </sec> <sec> <title>INTERNATIONAL REGISTERED REPORT</title> <p>RR2-10.2196/65790</p> </sec>