From Clinical Interview To Computational Signal: A Review Of DAIC-WOZ-Based Depression Detection Systems

Authors: Harsh, Dr. Pramod Kumar

Abstract: Major Depressive Disorder (MDD) is a disease that impacts more than 280 million people world-wide, and is woefully underdiagnosed because of subjectivity of clinical interview-based assessment measures. Distress Analysis Interview Corpus – Wizard of Oz (DAIC-WOZ) has become the key baseline dataset to assess automated, multimodal computational systems for depression screening. This review systemically discusses the machine learning and deep learning architectures tested on DAIC-WOZ ranging from unimodal facial, acoustic and textual approaches up to cutting-edge multimodal fusion frameworks that involve cross-modal attention and temporal graph neural networks. Important methodological weaknesses, subject leakage and shortcut learning through therapists prompts, are tested and quantified. Results show that some results of multimodal systems have a binary F1 score greater than 0.91, although many of the results are extremely overrated due to obsolete implementation of data division. Finally, the methodological recommendations and directions for future research to create clinically deployable, privacy-preserving and demographically generalizable depression screening systems are provided.

DOI: http://doi.org/10.5281/zenodo.21126640