Speech-Derived Digital Markers of Treatment Response and Improvement Trajectories in Pediatric Anxiety

Anxiety disorders are highly prevalent in youth. Exposure-focused cognitive behavioral therapy (EF-CBT) is the first-line treatment, yet a substantial proportion of youth do not achieve full response. Clinicians currently lack scalable, objective markers to monitor treatment response as it unfolds during therapy, limiting the ability to make timely, data-informed treatment adjustments. This study examined whether youth speech features (acoustic and linguistic) from EF-CBT session audio recordings could predict anxiety improvement and model trajectories of response across treatment. The sample included 603 recorded sessions from 60 youth aged 7 to 17 years participating in a randomized clinical trial of 12-session EF-CBT. Improvement was defined as a ≥ 30% reduction in Pediatric Anxiety Rating Scale (PARS) scores between sessions, assessed by independent evaluators. Machine learning models were developed using data from pairs of sessions within individuals to predict whether improvement occurred between those sessions and were evaluated using internal cross-validation. Models using later-session features performed better than those using early-session features (AUC = 0.86 vs 0.82), indicating good discrimination between improvement and non-improvement. Acoustic features reflecting vocal variability and expressiveness (e.g., variation in pitch, loudness, and voice activity) were the most consistent predictors of improvement. A hybrid Empirical Bayes approach integrating speech-based model predictions with PARS-based priors produced the most accurate and clinically plausible trajectories compared to other approaches tested. These findings suggest that speech-derived markers may provide a scalable, session-level measure to support continuous monitoring of treatment response during EF-CBT for pediatric anxiety.