Cross-Specialty Analysis of Japanese Perinatal EHR Narratives Using Large Language Models: Documentation Asymmetry and the Classification of Depressive States (Preprint)

<sec> <title>BACKGROUND</title> The perinatal period is associated with an elevated risk for depressive disorders, making timely identification a public health priority. Collaboration between obstetrics and psychiatry often remains suboptimal in clinical practice. A potential contributor to this gap is information asymmetry across specialties—qualitative and quantitative imbalances in how psychiatric symptoms are documented in electronic health records (EHRs). </sec> <sec> <title>OBJECTIVE</title> To quantify cross-specialty asymmetries in perinatal EHR documentation and to evaluate the performance of large language model (LLM)–based approaches for classifying depressive states from heterogeneous clinical narratives. </sec> <sec> <title>METHODS</title> We retrospectively collected Japanese-language EHR notes for 94 pregnant or postpartum patients with a history of psychiatric consultation from a single center. We compared the prevalence of symptom mentions by author specialty. Using stratified sampling, we then selected a cohort of 20 patients to benchmark multiple LLM-based inference pipelines for depressive-state classification. The best-performing pipeline combined LLM outputs with a rule-based layer encoding clinical knowledge for final adjudication. </sec> <sec> <title>RESULTS</title> Documentation practices differed markedly by clinical specialty. Mentions of “depression” were approximately 30-fold less frequent in notes authored by the obstetrics team compared with those by psychiatrists. Despite this documentation bias, the top-performing pipeline achieved an F1-score of 0.750 (95% CI 0.500–0.917) for classifying depressive states. </sec> <sec> <title>CONCLUSIONS</title> This study highlights the potential of LLMs to classify depressive states from heterogeneous EHRs despite significant documentation asymmetry. However, effective clinical deployment will require both technical advances (eg, bias-aware modeling) and organizational efforts to improve documentation quality and interspecialty information sharing. </sec>