Abstract
BACKGROUND: Artificial intelligence (AI) models have been increasingly explored for predicting treatment response to cognitive behavioral therapy (CBT) in patients with anxiety disorders. Identifying potential responders in advance may help inform treatment planning and support clinical decision-making. Although a growing number of studies have applied AI techniques in this context, reported performance estimates vary across studies, and the overall predictive accuracy has not been comprehensively quantified. OBJECTIVE: This systematic review and meta-analysis aims to quantify the overall performance of AI models in predicting treatment response following CBT for anxiety disorders and to examine how data sources, algorithmic approaches, and diagnostic subtypes influence predictive performance. METHODS: A systematic literature search was conducted in PubMed, Embase, Web of Science, Cochrane Library, and PsycINFO up to August 2025. We included studies that validated AI models for predicting CBT treatment response (remission or response) in patients diagnosed with an anxiety disorder. The risk of bias was assessed using the PROBAST+AI (Prediction Model Risk of Bias Assessment Tool for Artificial Intelligence) tool. Predictive performance metrics, including sensitivity, specificity, accuracy, and area under the curve (AUC), were extracted and pooled. Pooled estimates for sensitivity, specificity, and diagnostic accuracy were derived using the Restricted Maximum Likelihood estimator, with CIs adjusted via the Hartung-Knapp-Sidik-Jonkman method. Prediction intervals were calculated and reported alongside these pooled estimates to illustrate the expected distribution of effects in real-world settings. RESULTS: Eleven studies were included in the meta-analysis. The pooled sensitivity of AI-based models for predicting treatment response was 0.73 (95% CI 0.58-0.85; I²=82.8%), and the pooled specificity was 0.75 (95% CI 0.59-0.89; I²=96.7%). The overall pooled accuracy was 0.74 (95% CI 0.62-0.84; I²=94.6%). The summary AUC was 0.81 (95% CI 0.78-0.85), indicating moderate discriminative performance. Subgroup analyses showed that models incorporating multimodal data achieved superior predictive performance, with a pooled sensitivity of 0.84 and an accuracy of 0.82. In addition, predictive performance was the highest in patients with social anxiety disorder compared with other anxiety disorder subtypes. CONCLUSIONS: This meta-analysis quantitatively synthesized AI performance in predicting CBT response for anxiety disorders, moving beyond narrative reviews to provide pooled evidence. In contrast to existing reviews that encompass broader diagnostic groups, our focused approach establishes a precise benchmark for this clinical domain, highlighting the current moderate overall performance. Furthermore, we extend beyond previous work by demonstrating the superior predictive utility of multimodal data, identifying social anxiety disorder as the most predictable subtype, and systematically evaluating the impact of data modalities and algorithm types. Future efforts should prioritize robustly validated multimodal models, laying essential groundwork for the potential development of AI-assisted tools to personalize treatment planning in anxiety disorders.