The MADRS Pipeline: Supporting Depression Assessment in Clinical Trials

Clinical trials for depression rely heavily on structured interviews, such as those following the SIGMA guidelines, to assess symptom severity via the Montgomery–Åsberg Depression Rating Scale (MADRS). Because these assessments are subjective and dependent on clinician expertise, they often introduce variability that can compromise trial outcomes. This paper introduces the MADRS Pipeline, a computational framework designed to assist clinicians by automating the transcription, segmentation, and scoring of these interviews. The pipeline operates as a directed acyclic graph consisting of four stages: (1) audio-to-text transcription using Whisper; (2) transcript segmentation into ten symptom-specific items using GPT-4.1; (3) severity estimation using a fine-tuned Transformer encoder; and (4) quality assessment to flag potentially non-compliant clinical ratings.

The researchers evaluated the pipeline on a dataset of approximately 1,602 clinical interviews. The transcription component achieved high lexical and semantic alignment with manual references. The segmentation module demonstrated strong performance, with one-vs-all F1 scores exceeding 0.94 across all MADRS items. In the core assessment task, the pipeline achieved a Spearman correlation of 0.867 with expert human ratings. Furthermore, the quality assessment module successfully identified discrepancies between rater scores and expected clinical practice, providing a mechanism to monitor assessment reliability in real-world trial settings.

By providing an interpretable, automated tool for depression assessment, this pipeline offers a scalable way to reduce inter-rater variability in clinical trials. Unlike previous approaches that focus on binary detection from social media or non-clinical text, this system is specifically engineered for the structured, high-stakes environment of drug-development trials. It serves as a "computational clinician" that not only estimates severity but also flags potential errors in human scoring, thereby enhancing the consistency and reliability of trial data.