Timely risk stratification of ICU patients with bloodstream infection (BSI) is essential for guiding clinical decisions. Prediction models developed in Western populations may not generalize to other regions because healthcare systems, case mix, and clinical practice differ. We evaluated the cross-national transportability of a parsimonious ICU mortality model and tested local logistic recalibration in a Chinese external cohort. We developed a 30-day mortality model using MIMIC-IV (n = 1,501) with Elastic Net regularization and multivariable logistic regression. The simplified model included five routinely available predictors: age, 24-hour Sequential Organ Failure Assessment (SOFA) score, Charlson Comorbidity Index (CCI), serum potassium, and fungal BSI. We externally validated the model in an independent cohort of 285 ICU patients with BSI from a tertiary teaching hospital in China. In internal validation, the simplified model achieved an area under the receiver operating characteristic curve (AUC) of 0.776, compared with 0.796 for the full model. In the Chinese cohort, discrimination remained acceptable (AUC 0.764), but calibration was poor, with systematic underestimation of 30-day mortality (calibration-in-the-large > 0). Local logistic recalibration improved calibration, bringing the calibration-in-the-large and slope closer to target values and reducing the Brier score to 0.182. The US-derived parsimonious model retained acceptable discrimination in this Chinese ICU cohort but showed important calibration drift. Local logistic recalibration may offer a practical, low-resource approach to support local adaptation. Because recalibration was estimated and evaluated in the same cohort, the improvement may be optimistic and requires validation in independent healthcare systems.
Indeterminate thyroid cytology remains a diagnostic challenge. We investigated whether ultrasound-derived clinical features and Raman spectra acquired from fine-needle aspiration samples could be integrated within an explainable multimodal machine learning framework. The analytical cohort comprised 203 patients from a single center who underwent thyroid surgery, including 64 follicular adenomas (FA) and 139 malignant lesions according to postoperative histology. The malignant class comprised 117 papillary thyroid carcinomas (PTC), 21 follicular-variant papillary thyroid carcinomas (FV-PTC), and one mixed FA+PTC lesion. The fixed primary architecture combined a clinical Logistic Regression classifier and a raw-Raman Extra Trees classifier through a Logistic Regression meta-learner. Internal validation used stratified fivefold cross-validation repeated 20 times, with all data-dependent preprocessing, stacking, and operating-threshold selection restricted to the corresponding training data. The stacked model achieved a mean repeated out-of-fold AUC of 0.856 (95
Adverse event (AE) reporting systems are central to hospital safety monitoring, yet severity classification remains dependent on human review and may be subject to variability among reviewers. Large language models (LLMs) may support more consistent classification of safety reports. To evaluate agreement between a large language model (LLM)-based classification system and adjudicated expert human classifications of adverse event (AE) severity, while characterizing variability in the human classification process. We analyzed 286 adverse event reports from the institutional safety reporting system of a tertiary academic medical center. Two quality and safety experts independently classified each report according to an A–I severity scale aligned with the National Coordinating Council for Medication Error Reporting and Prevention (NCC MERP) framework; disagreements were adjudicated by a third senior reviewer to establish the final human classification. The same reports were processed by a prespecified LLM-based workflow whose configuration was fixed before evaluation, without training, fine-tuning, calibration, or prompt modification on the study dataset; when the workflow generated clarification questions, the recorded responses were used to produce a second classification. Exact agreement with the final human classification was treated as the primary descriptive agreement measure; one-category differences and larger discordances were considered separately. Initial exact agreement between the two human reviewers was 65.4
Large Language Models have become impactful in medicine by facilitating the processing of difficult medical data and aiding diagnosis-making. Given general LLM’s weaknesses, specialized medical LLMs are constructed for clinical precision, contextual understanding and safety in risky diagnostic applications. This review summarizes recent trends in medical LLMs for diagnosis, reported during 2021–2025. A systematic literature review was performed following the PRISMA methodology. A search was designed and systematically carried out in the SCOPUS database using criteria of inclusion and exclusion (type of publication, language, period and country) and 75 peer-reviewed articles were chosen from a database of 1,911 records, after a multi-stage screening and eligibility evaluation. In summary, there have been significant advancements in the design and application of medical LLMs, with trends in hybrid architecture, multi-modal learning, prompt optimization techniques, diffusion-based imaging models, and multi-agent diagnostic system models. However, a number of persistent challenges that need to be overcome remain in the development and implementation of these models; namely, hallucination, bias, and poor probability calibration in LLMs, issues with explainability of LLMs, and ineffectiveness in a clinical environment. This review highlights the fast pace of development and limitations of domain-specific medical LLMs. While impressive improvements in diagnostic performance have been made, the future directions must address improvements in reliability, safety and clinical validation in order to be seamlessly and responsibly incorporated in the practice.
Echocardiographic assessment of diastolic dysfunction from small clinical cohorts is susceptible to overfitting, unstable feature attribution, and poorly calibrated confidence estimates, limiting the clinical translation of machine learning models. This study aimed to develop and evaluate CARDINAL, a physiology-guided framework combining a dataset-specific structural prior, directional graph regularization, prototype-aligned representation learning, and Monte Carlo dropout uncertainty estimation, and to assess whether it improves discrimination, calibration, and uncertainty-guided referral relative to established tabular learning models under a leakage-controlled protocol. We present CARDINAL, a physiology-guided framework that integrates four components: (i) a dataset-specific structural prior that encodes known physiological relations among clinical variables; (ii) directional graph refinement under sparsity and acyclicity penalties, serving as a regularizer and interpretability scaffold rather than as an identified causal model; (iii) prototype-aligned representation learning to enforce intra-class compactness under limited data; and (iv) Monte Carlo dropout-based uncertainty estimation ( T=30 , dropout rate 0.20) for calibrated confidence and selective referral of borderline cases. The study cohort comprised 118 cases classified by a composite reference index as normal ( n=37 ), mild ( n=42 ), or severe ( n=39 ). Evaluation followed a leakage-controlled protocol with a fixed stratified 70/15/15 split producing an 82-case training pool, an 18-case validation set, and an 18-case locked test set evaluated once after all modeling choices were finalized. On the locked test set, CARDINAL achieved a binary abnormal-versus-normal ROC-AUC of 0.969, AUPRC of 0.990, accuracy of 0.944, sensitivity of 0.923, specificity of 1.000, Brier score of 0.066, ECE of 0.095, and NLL of 0.223. Patient-level DeLong and McNemar tests were non-significant on this single split ( n=18 ), reflecting limited statistical power rather than equivalence. Across 20 matched repeated runs, CARDINAL showed consistent stability with Holm-adjusted Wilcoxon p = 3.05×10^-5 versus TabPFN-tuned. Three-class confusion analysis indicated reasonable grade-level discrimination, with misclassifications occurring predominantly between adjacent grades. Ablation analysis suggested that each framework component—directional graph regularization, prototype alignment, and uncertainty calibration—contributed to overall performance. Three public datasets (UCI Heart Disease, Heart Failure Clinical Records, and MIMIC-IV) served as tabular robustness benchmarks; CARDINAL showed consistently modest numerical advantages across all four datasets. These results suggest that physiological structure, prototype learning, and uncertainty quantification can be integrated within a leakage-controlled evaluation framework. Feature importance analysis indicated that E/e^' and LAVI—label-adjacent variables in the reference index—were the predominant contributors to model predictions, consistent with established cardiology guidelines. A complementary analysis excluding these label-adjacent variables preserved CARDINAL’s relative advantage (ROC-AUC 0.873 versus 0.846 for TabPFN-tuned). Uncertainty estimates may assist in identifying borderline cases for selective human review. These findings motivate prospective multi-center validation to evaluate clinical utility, transportability, and the impact of uncertainty-guided review on patient care.
