R-GCRED: a retrieval-grounded and calibrated risk estimation for depression prediction and screening recommendation framework

Abstract

Early depression detection from textual narratives has advanced substantially with transformer-based language models and Large Language Models (LLMs). However, reliable screening-oriented decision support requires more than high predictive accuracy. Existing approaches often lack explicit symptom-level grounding, produce poorly calibrated confidence estimates, and provide limited safeguards against uncertainty and unsupported model-generated evidence. This paper introduces R-GCRED, a Retrieval-Grounded and Calibrated Risk Estimation framework for transparent, evidence-grounded, and uncertainty-aware depression risk estimation from text. R-GCRED employs ontology-guided retrieval to constrain LLM-based evidence extraction to clinically relevant symptom facets, producing structured and source-traceable representations that capture symptom identity, polarity, intensity, and supporting textual evidence. A dedicated evidence-verification mechanism further reduces the propagation of unsupported symptom assertions to downstream prediction. The verified evidence is subsequently processed by a dual-head architecture for joint depression-risk classification and severity estimation, followed by probability calibration, uncertainty-aware abstention, and cost-sensitive recommendation modeling. R-GCRED is evaluated on the DAIC-WOZ and CLEF eRisk benchmarks using participant-disjoint evaluation protocols and is compared against general-domain transformers, mental-health-adapted language models, evidence-based methods and LLM-based baselines. Across the evaluated benchmarks, R-GCRED consistently outperforms the compared approaches, indicating that its performance gains extend beyond those obtained against conventional transformer baselines. The framework achieves a Macro-F1 of 0.84, an AUC of 0.89, and an Expected Calibration Error of 0.038, while attaining an AUC of up to 0.90 under partial-observation early-detection settings. Results demonstrate that integrating retrieval-grounded symptom evidence, evidence verification, calibrated prediction, and uncertainty-aware decision modeling yields measurable improvements in predictive discrimination, confidence reliability and explanation traceability over the evaluated existing approaches. These findings establish the computational effectiveness of R-GCRED across the evaluated benchmark settings but do not constitute prospective clinical validation.