Measuring Depression Severity With Clinical Global Impression-Severity Scale Scores From Clinical Notes Using Large Language Models: Validation Study.

BACKGROUND

Real-world psychiatric care is marked by wide heterogeneity in clinical presentations and outcomes, underscoring the need for systematic approaches to outcome measurement. The Clinical Global Impression-Severity (CGI-S) scale is a brief, clinician-rated measure of overall illness severity that is widely used in psychiatric research, but rarely documented in routine care. Large language models (LLMs) may enable automated extraction of CGI-S scores from narrative clinical notes, thereby providing scalable outcome measures for real-world clinical care and research.

OBJECTIVE

The study aimed to evaluate whether LLMs can estimate CGI-S scores from psychiatric clinical notes for patients with major depressive disorder (MDD) and to compare performance across prompting strategies and model architectures.

METHODS

We extracted psychiatrist-authored notes from the Johns Hopkins electronic health record. Three board-certified psychiatrists independently rated 77 clinical notes using a validated depression-specific Clinical Global Impression (CGI) rubric. Weighted Cohen kappa coefficients were calculated to assess inter-rater reliability and model-human agreement. We evaluated GPT-4o under zero-shot and few-shot prompting conditions and Llama-4 under zero-shot prompting. Model performance was assessed by comparing LLM-generated scores to individual rater scores and consensus ratings. Exploratory analyses evaluated whether agreement varied by patient demographics, care setting, note length, or the percentage of copy-forwarded text within each note.

RESULTS

Interrater reliability among psychiatrists was high (κ=0.77-0.78). GPT-4o with zero-shot prompting demonstrated the highest agreement with average human ratings (κ=0.85, 95% CI 0.78-0.90), and few-shot prompting did not improve performance. In contrast, Llama-4 with zero-shot prompting demonstrated lower agreement with average human ratings (κ=0.70, 95% CI 0.55-0.80). Model agreement did not significantly differ across age, sex, race, treatment location, or the percentage of copy-forwarded text, but it was significantly lower for notes below the median note length than for notes at or above the median length (κ=0.72 vs 0.92; P=.003).

CONCLUSIONS

LLMs can estimate clinician-rated CGI-S scores from psychiatric clinical notes for patients with MDD at a level of agreement comparable to that of expert interrater reliability. Performance varied by model architecture, with GPT-4o outperforming an open-source alternative. If further validated, this approach may support scalable outcome measurement in research settings and inform future efforts to implement measurement-based care in real-world psychiatric practice.