Abstract: DRKS00030915
Executive Summary
The PREACT study (Research Unit 5187) aims to bridge the "translational gap" in mental health by using machine learning to predict which patients will not respond to Cognitive Behavioral Therapy (CBT). By collecting a massive, multimodal dataset—ranging from brain scans and heart rate variability to smartphone usage patterns—researchers hope to move from "one-size-fits-all" therapy to precision psychotherapy tailored to the individual.
The Challenge of Treatment Non-Response
Cognitive Behavioral Therapy (CBT) is considered the gold standard for internalizing disorders such as depression, anxiety, and OCD. However, a significant challenge remains: fewer than 50% of patients achieve clinically meaningful improvement. Currently, clinicians lack reliable tools to identify non-responders before treatment begins. Most research relies on group averages, which provide little guidance for individual patient care. PREACT addresses this by focusing on "signatures" of treatment non-response (TNR) in a real-world outpatient setting.
Conceptual Framework: Emotion Regulation
The study identifies emotion (dys-)regulation as the core mechanism underlying internalizing disorders. By adopting a transdiagnostic approach, researchers aim to find universal predictors of therapy success across different diagnoses, linking specific brain circuits to behavioral symptoms.
Methodology: Deep Phenotyping and AI
PREACT employs a "deep phenotyping" approach, observing patients through multiple lenses simultaneously:
- Clinical & Psychological: Standardized interviews and questionnaires.
- Neuroimaging (fMRI): Measuring brain structure and activity during emotion regulation tasks.
- Psychophysiology: Tracking EEG and Heart Rate Variability (HRV).
- Digital Phenotyping: Using wearables and smartphone apps for Ecological Momentary Assessment (EMA) of mood and activity.
The AI strategy involves state-of-the-art machine learning, including Convolutional Neural Networks (CNNs) for neuroimaging, Transfer Learning from large datasets like the UK Biobank, and Explainable AI (XAI) to ensure model transparency.
Current Progress and Study Flow
As a study protocol, PREACT focuses on establishing a robust pipeline for predictive modeling. Preliminary simulations suggest that a target of 585 patients will provide sufficient data for high-accuracy training, with model performance typically plateauing around 300 samples.
As of late 2024, the study has successfully consented 434 patients, with 347 completing baseline assessments (T1). The ultimate goal is to achieve a single-patient prediction accuracy of at least 75%, representing a significant leap over current clinical judgment.
Limitations and Future Directions
The intensity of deep phenotyping presents a high participant burden, which may impact attrition rates. Furthermore, while the study is more naturalistic than traditional trials, it is limited to university outpatient populations. Future phases of PREACT aim to transition from predicting failure to preventing it by suggesting personalized augmentation strategies, such as neurofeedback or non-invasive brain stimulation, for predicted non-responders.
Summary updated