22 January 2025 in Expert Systems with Applications
1 November 2025 in IEEE Journal of Biomedical and Health Informatics
Clinical coding translates medical information from Electronic Health Records (EHRs) into structured codes such as ICD-10, which are essential for healthcare applications. Advances in deep learning and natural language processing have enabled automatic ICD coding models to achieve notable accuracy metrics on in-domain datasets when adequately trained. However, the scarcity of clinical medical texts and the variability across different datasets pose significant challenges, making it difficult for current state-of-the-art models to ensure robust generalization performance across diverse data distributions. Recent advances in Large Language Models (LLMs), such as GPT-4o, have shown great generalization capabilities across general domains and potential in medical information processing tasks. However, their performance in generating clinical codes remains suboptimal. In this study, we propose a novel ICD coding paradigm based on code verification to leverage the capabilities of LLMs. Instead of directly generating accurate codes from a vast code space, we simplify the task by verifying the code assignment from a given candidate set. Through extensive experiments, we demonstrate that LLMs function more effectively as code verifiers rather than code generators, with GPT-4o achieving the best performance on the CodiEsp dataset under zero-shot settings. Furthermore, our results indicate that LLM-based systems can perform on par with state-of-the-art clinical coding systems while offering superior generalizability across institutions, languages, and ICD versions.
9 July 2026 in European Heart Journal
Traditional Cox proportional hazards models show suboptimal performance for cardiovascular disease (CVD) risk prediction in Chinese populations. Transformer-based deep learning models have demonstrated promise in clinical risk prediction. In this study, sex-specific transformer-based models (China-AIHeart) for 10-year CVD risk prediction among Chinese adults were developed and validated. The derivation cohort included 156 790 participants [34.6% men; mean [SD] age, 56.7 [8.9] years) without CVD from the China Cardiometabolic Disease and Cancer Cohort. External validation was conducted in two independent Chinese cohorts (Xinjiang and CHARLS). Transformer-based time-to-event prediction models were developed, including a full model (22 predictors) and a simplified model (15 predictors). Performance was compared with Cox models using identical predictors and established risk scores (China-PAR, PREVENT-ASCVD, and SCORE2 Asia-Pacific equations). China-AIHeart demonstrated good discrimination (C-statistic [95% confidence interval, CI]: .767 [.754-.779] in men; .780 [.769-.791] in women), calibration (calibration χ2: 14.806 in men; 9.326 in women; Brier score: .104 in men; .077 in women), and net clinical benefit in predicting CVD risk. Predicted event rates closely matched observed risks across strata. Compared with Cox models with identical predictors, China-AIHeart showed improved discrimination (ΔC-statistic [95% CI]: .027 [.025-.028] in men; .031 [.029-.033] in women) and reclassification (net reclassification index [95% CI]: .478 [.466-.492] in men; .560 [.551-.572] in women), and outperformed China-PAR, PREVENT-ASCVD, and SCORE2 Asia-Pacific equations. External validation demonstrated robust performance, with C-statistics of .781/.825 (men/women) and .748/.820 for the full and simplified models in the Xinjiang cohort, and .740/.771 for the simplified model in the CHARLS cohort. The transformer-based China-AIHeart models predicted 10-year CVD risk and outperformed traditional Cox-based approaches, providing a practical tool for risk stratification in Chinese adults.
21 March 2026 in Pattern Recognition
21 May 2026 in Knowledge-Based Systems