Lessons from preventable ICU-related coronial deaths: A large-language-model-assisted national analysis of adult cases in Australia (1997 to 2025)

PURPOSE: Australian coroners independently examine the circumstances of reportable deaths and, where the evidence permits, may determine whether a death is explicitly or potentially preventable. Findings involving intensive care are dispersed across eight jurisdictions and rarely synthesised. We aimed to derive clinical lessons from publicly available Australian coronial findings on preventable ICU-related deaths. MATERIALS AND METHODS: We retrieved 8583 reports from coroners' court websites and processed them through a two-stage large-language-model pipeline (Gemini 3.1 Pro Preview), constrained by versioned prompts and clinician validation. Eligible cases were adult ICU-related deaths identified by the model as meeting a predefined coroner-language criterion for explicit or potential preventability. Each case was classified by clinical presentation, ICU interface, and cross-cutting failure mode. RESULTS: Of 1190 ICU-related cases, 563 met the preventability criterion (deaths 1997 to 2025; median age 52 years; 59.5% male). Leading presentations were trauma (19.7%) and perioperative or procedural (18.8%). The dominant failure mode was communication, handover and documentation (63.6%), followed by treatment, medication or device problems (50.6%) and diagnostic-process failure (49.4%). In a fully crossed two-reviewer audit of 130 cases, LLM-clinician agreement was substantial (presentation κ 0.64; mean failure-mode κ 0.63). CONCLUSIONS: Communication failures were the most frequently coded failure domain in this publicly available corpus, and perioperative deaths carried the heaviest multi-modal burden. Our findings support consideration for strengthening structured handover interfaces and postoperative ward-rescue protocols after major surgery. Integrating coronial findings into mortality review and a national ICU-specific synthesis are proposed, with the validated pipeline as a candidate enabling mechanism.