Quintuple Validation Inertial Framework: Multimodal Sensor Fusion and Deep Learning for Precise Detection of Body-Focused Repetitive Behaviors
Nafea M Alanazi, Muhammad Adnan · Preprints.org · 2025
Body-Focused Repetitive Behaviors (BFRBs), such as hair-pulling and skin-picking, affect millions worldwide, often leading to significant distress and impairment. Traditional self-report assessments suffer from bias and subjectivity, underscoring the need for objective, real-time monitoring tools. This study introduces the Quintuple Validation Inertial Framework (QVIF), a novel AI-driven approach leveraging multimodal wearable sensor data for precise BFRB detection. By fusing inertial measurement unit (IMU), time-of-flight, and thermopile sensor signals through Kalman filtering and deep fusion networks, QVIF extracts kinematic-enhanced metrics (KEMs) and employs gradient boosting ensembles (e.g., XGBoost) for classification. A hybrid CNN-LSTM architecture processes time-series data, with participant-stratified 5-fold cross-validation ensuring robustness against individual variability. Evaluated on the CMI-Detect Behavior dataset comprising 574,945 sensor readings from 80 participants, QVIF achieves a mean validation accuracy of 90.6\% ($\pm$0.8\%) and weighted F1-score of 0.903 ($\pm$0.006) for 18 gestures, outperforming single-modality baselines by 3.0\%. Phase prediction attains 87.39\% accuracy, highlighting superior temporal segmentation. These results demonstrate QVIF's potential for scalable, privacy-preserving mental health monitoring, paving the way for proactive interventions in clinical settings.