Прогнозная модель не оправдала отказ от онлайн-терапии для пациентов с депрессией

TOPLINE Patients with more severe depression at baseline experienced greater benefits from internet-based cognitive-behavioural therapy (iCBT), but the prediction model lacked precision to support withholding iCBT from those with low predicted benefits, a prognostic study found. METHODOLOGY - Researchers performed a prognostic study using data from five randomised clinical trials of three iCBT programmes conducted in Germany and Austria between 2012 and 2023 to evaluate a benefit prediction model for iCBT. - The analysis included 1589 adults (mean age, 41.48 years; 74.8% women) with depressive symptoms or disorders recruited from the samples: Krämer et al, Preiss et al, Moritz et al, and EVIDENT 1 plus EVIDENT 2 (the development sample). - Elastic net regression (ENR), ordinary least squares, and causal forest models were evaluated using 12 baseline predictors, including sex, age, relationship status, education, employment, antidepressant use, concurrent psychotherapy, diagnoses, baseline depression severity, and quality-of-life measures. - The primary outcome was a change in depression severity from baseline at 12 weeks assessed using the nine-item Patient Health Questionnaire (PHQ-9). If 12-week data were unavailable, assessment data at 8 or 10 weeks were used. - Performance was assessed for absolute outcome prediction and individual treatment benefit prediction, including calibration for benefits and decision accuracy analyses comparing model-guided selective treatment vs treat-all strategies. TAKEAWAY - ENR best predicted outcomes and benefits (coefficient of determination [R2] range, 0.19-0.50) using 12 predictors, and baseline depression severity was the strongest effect modifier; each unit increase in the baseline PHQ-9 score was associated with 0.16 points (95% CI, 0.06-0.25) of improvement under treatment. - In external validation, ENR showed consistently positive calibration slopes for benefit prediction across three trial samples (Moritz et al: R2, 0.32; slope, 0.64; Krämer et al: R2, 0.50; slope, 1.45; Preiss et al: R2, 0.39; slope, 1.04). - The difference in observed treatment effects between the predicted highest and lowest benefit quantiles from ENR was 1.34 points in the development sample, and CIs excluded zero in only one of three external validation samples (4.37 points). - Selectively treating patients by model recommendation performed worse than a treat-all strategy across samples (range, -2.01 to -0.25 points on the PHQ-9) as even patients in the lowest predicted benefit quartile improved substantially from iCBT (1.18-8.48 points). IN PRACTICE "In this prognostic study of individual treatment benefit from iCBT for depression, a sparse regularized model differentiated groups of patients by predicted iCBT benefit in a consistent manner across independent trials. This signal was insufficient to support model-guided treatment allocation," the authors wrote. "Our findings further support offering iCBT adjunctively even in more severe depression and where face-to-face treatment is unavailable or subject to long waiting times," they added. SOURCE The study was led by Cora Schefft, MD, Charité - University Medical Center Berlin, Berlin, Germany. It was published online on August 28, 2026, in JAMA Network Open. LIMITATIONS The study was limited by the analysis of only the participants with complete outcome data, varying and unrandomised guidance intensity across trials, and the exclusive reliance on samples from German-speaking countries. Additionally, participant race and ethnicity were not systematically recorded across trials, and one of the evaluated trials featured substantial attrition and an active sham control. DISCLOSURES No funding was received for this work. Several authors disclosed holding stocks and receiving grants, fellowships, or personal fees from the Berlin Institute of Health, Charité - University Medical Center Berlin, German Federal Ministry of Health, several European organisations, and pharmaceutical companies. One author reported serving as the vice chairman or past president of psychiatry research organisations; another author reported serving as an employee of GAIA AG, the developer of one of the digital therapeutic interventions described in this study. This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.