Large Language Models for Distress Rating in Korean Psycho-Oncology Interviews: Exploratory Clinician-Benchmarked Evaluation Study

Abstract

Abstract Background Psychiatric distress is common among patients with cancer; yet, systematic interview-based screening remains difficult to scale in routine clinical care. Large language models (LLMs) have shown promise as scalable tools for mental health assessment, but most existing evidence is derived from clinician-authored records, translated text, or proxy data. The performance characteristics, error patterns, and explanatory behaviors of contemporary LLMs when applied to authentic, non-English psychiatric interviews remain insufficiently characterized. Objective This exploratory study evaluated how contemporary LLMs reproduced psycho-oncologists’ item-level symptom ratings from real-world Korean psycho-oncology interviews, focusing on concordance, directional bias, and clinician-adjudicated error characteristics. Methods Between April 2024 and May 2025, 101 adults receiving oncologic care in South Korea underwent semistructured interviews. Board-certified psycho-oncologists provided real-time, time-stamped ratings of 39 items. Generative pretrained transformer 4o (GPT-4o), Claude 3.5 Sonnet, and Gemini 2.5 Flash generated item scores and brief rationales using identical Korean zero-shot rubrics. Concordance between clinician ratings was evaluated using ordinal and binary screening metrics. A paired Wilcoxon test compared patient-level total symptom burden. Binary mismatches were clinically adjudicated as ambiguous or definite overestimation or underestimation with an 8-etiology taxonomy. Model-generated rationales were further meta-evaluated using GPT-5.4, Claude Sonnet 4.6, and Gemini Pro 3.1 across 4 dimensions: citation (use of quoted supporting statements), structure (logical organization of the rationale), mapping (consistency between the rationale and the assigned item rating), and expansion (degree of interpretive elaboration beyond the explicit transcript content). The associations between meta-evaluation results and absolute error were examined using cross-classified mixed-effects models. Results Out of 101 participants, 88 (87.1%) were predominantly female and had breast cancer as the primary cancer type (n=70, 69.3%). Across 3931 item-level ratings, all models showed good agreement with clinicians (intraclass correlation coefficient: 0.816‐0.872), with GPT-4o showing the highest agreement. Claude 3.5 and Gemini 2.5 yielded significantly higher patient-level symptom burden (adjusted P<.001 and adjusted P=.002, respectively), whereas GPT-4o did not (adjusted P=.15). Clinician adjudication attributed 97 of 248 (39.1%) GPT-4o mismatches, 88 of 301 (29.2%) Claude 3.5 mismatches, and 118 of 355 (33.2%) Gemini 2.5 mismatches to intrinsic ambiguity in patient speech. Among definite errors, misapplication of severity thresholds was the predominant mechanism across models. Exploratory mixed-effects models showed that higher expansion and lower mapping scores were associated with larger absolute errors. However, the association with expansion may reflect case difficulty or ambiguity rather than causation. Conclusions In this exploratory clinician-benchmarked evaluation of authentic interviews from a Korean psycho-oncology sample comprising predominantly women and patients with breast cancer, LLMs showed high aggregate concordance with psycho-oncologists’ item-level ratings while differing in their error profiles. These findings support further evaluation of LLM-based item-level symptom-rating approaches in psycho-oncology. Validation in larger, more diverse, and independent cohorts is needed.