From Clinical Narratives to Predictive Signals: Data-Driven Modeling of Psychiatric Treatment Response
Main article
Abstract
Treatment selection in psychiatry remains a trial-and-error process because conventional predictors such as symptom rating scales, demographics, and basic comorbidity indicators carry only modest individual-level information about likely response. Clinical narratives — spoken interviews and the free-text notes that accumulate in electronic health records — encode signals about thought form, affect, functioning, and patient priorities that structured fields rarely capture. This paper develops a data-driven view of psychiatric treatment response in which language, processed by natural language processing and large language models, is treated as a first-class predictor alongside structured variables. We describe a workflow-native pipeline that turns interviews and progress notes into features, combines them with structured electronic health record variables, and feeds them to predictive and causal models for treatment-effect estimation and ranked decision support. We synthesise reported discriminative performance across predictor families, illustrate the operational gain that language adds at clinically relevant decision thresholds, examine fairness across subgroups, and quantify the calibration degradation that occurs when models are deployed without retraining. We argue that language-augmented modelling can move psychiatric decision support beyond the ceiling imposed by structured predictors alone, provided the pipeline is designed with explicit attention to confounding, distributional drift, fairness, privacy, and interpretability. The paper closes with concrete recommendations for the next phase of deployment-ready, language-aware psychiatric analytics.
