Monitoring Therapeutic Plans and Risk Signals from Clinical Narratives in Mental Health using Natural Language Processing

Clinical narratives in electronic health records (EHRs) are a key source of information in mental health care, but their unstructured nature limits systematic analysis. We propose an NLP framework to extract therapeutic plan information and identify indirect clinical risk signals from psychiatric narratives in Spanish. Using a retrospective dataset of outpatient consultations, we develop an interpretable approach combining rule-based extraction with a risk model based on TF-IDF and logistic regression, incorporating probability calibration and asymmetric thresholds to prioritize high-risk cases. We compare this approach against a fine-tuned transformer model (RoBERTuito) using patient-level cross-validation. These findings demonstrate that combining interpretability, robust evaluation, and domain-specific considerations enables the development of practical NLP tools for mental health applications using real-world clinical data.