Background: Postoperative intensive care unit (ICU) admission affects 15% to 20% of surgical patients and represents a major source of morbidity and health care costs. Current anesthetic dosing relies on empirical guidelines rather than individualized risk assessment. We developed a counterfactual dose-response model to identify optimal fentanyl-propofol combinations. Objective: This study aimed to develop and evaluate a stratified, causal machine learning framework using electronic health record data to identify optimal fentanyl-propofol dose combinations and predict postoperative ICU admission risk, enabling precision anesthesia and individualized clinical decision support. Methods: We analyzed perioperative electronic health records of 67,134 surgical procedures from UC Irvine Medical Center (2017‐2022). A hierarchical learning framework was used to estimate causal effects while controlling for confounding variables. A total of 6 dose-sensitive subgroups were identified through stratified analysis. The primary end point was postoperative ICU admission. Results: High-risk combinations (fentanyl >5 mcg/kg with propofol <1 mg/kg) increased ICU admissions’ absolute risk difference by 36% (absolute risk increase; 95% CI 0.351-0.509; P <.001). A total of 6 patient subgroups demonstrated distinct dose-response patterns, with populations considered vulnerable (high glucose, elevated creatinine) showing elevated risk even at standard doses. The optimal dose range for decision-making was determined to be 1.25 to 4.25 mg/kg for propofol and 3.5 to 4.0 mcg/kg for fentanyl. Conclusions: Fentanyl-propofol combinations exhibit complex, nonlinear dose-response relationships with ICU admission risk. High-dose combinations markedly increase risk through synergistic effects, while specific patient subgroups require enhanced monitoring even at standard doses. These findings support the development of individualized dosing algorithms and risk assessment tools that could inform future decision support tools aimed at reducing postoperative ICU use, although their predictive performance and clinical impact would require external validation.

Background: The Tabular Prior-data Fitted Network (TabPFN) is a recently introduced Transformer-based foundation model designed specifically for structured tabular data. TabPFN enables task inference without the need for hyperparameter tuning or extensive data preprocessing. Despite its disruptive potential, the application of TabPFN in surgical data science remains unexplored. In this study, we evaluate the performance of TabPFN across six surgical classification tasks. Objective: To assess TabPFN’s performance against benchmark machine learning models in surgical classification tasks and identify optimal application scenarios based on sample size and outcome incidence characteristics. Methods: In this study, perioperative data from two independent medical centres were utilized, comprising a large-scale cohort (n=67,134) and a medium-scale cohort (n=6,888). Six clinically relevant classification tasks were developed. The performance of TabPFN was systematically compared to benchmark models including XGBoost, Random Forest, Support vector machine, Logistic regression, and Decision tree, using area under the receiver operating characteristic curve, accuracy, precision, recall, F1-score, and calibration. Results: In tasks with large-sample sizes (n > 3,000) and higher outcome incidence (>40%), TabPFN achieved the highest recall and F1-score among all models. For tasks with low outcome incidence ( 3,000), but its calibration performance declined noticeably in low outcome incidence (<20%). Conclusions: TabPFN represents a promising methodological approach for tabular modelling in surgical data science. It also shows considerable promise in tasks with sample sizes exceeding 10,000. However, it is not yet capable of fully replacing established benchmark models. The application scenarios of TabPFN should be selected based on key task-specific characteristics. Further targeted training is necessary in future.

An exploration into the application of large language models in diabetes care as a support tool for clinical decision-making is warranted due to the model’s ability to integrate clinical guidelines, synthesize large volumes of patient data, construct clinical notes, assist in clinical teaching, and provide a rationale for complex management decisions. Diabetes, in particular, presents a complex clinical scenario, as the formulation of management decisions must consider multiple factors, including, but not limited to, multiple targets and risk factors, monitoring data, medication safety, and the patient’s behavioral and psychosocial needs, burdens, and nutritional preferences. This narrative review describes the possibilities of clinician-facing diabetes decision-support systems utilizing large language models and proposes a patient safety architecture and evaluation framework for their clinical assessment. The emphasis is on the safe implementation of these systems, not in an autonomous care setting. Based on the current literature, the systems will most likely assist with preparing clinical visits, reviewing clinical medications, preparing patient summaries for continuous glucose monitoring, planning clinical procedures, and coordinating communication within the clinical team. Despite the above benefits, these systems remain limited, and the constraints of large language models may have serious consequences in the field of diabetes, considering the potential risks of treatment-related adverse events, including hypoglycemia, diabetic ketoacidosis, and other complications based on inequitable provision of care. The safest and most viable near-term solution for large language models in clinical practice is a controlled, evidence-based, and clinician-supervised model that supports the retrieval of clinical guidelines, indicates appropriate levels of uncertainty, identifies gaps in the data, and assumes professional accountability. The focus of future studies should extend beyond validation studies and accuracy benchmarks to evaluating the impact on clinical workflows, equity, patient care and safety, and to monitoring these systems once deployed. Large language models have the potential to support diabetes care; however, they must act to enhance the clinical voice and judgment rather than diminish it.

One limitation of Assisted Reproductive Technology (ART) is its relatively low pregnancy success rate per cycle. A key contributing factor to this issue is the lack of precise and efficient sperm selection methods. The emergence of multimodal artificial intelligence (AI), which integrates diverse data types—including imaging, genomics, and clinical parameters—represents a transformative advancement in reproductive medicine. This system enhances the accuracy and comprehensiveness of sperm screening through dynamic functional assessments, molecular feature recognition, and advanced data integration strategies, thereby enabling the identification of sperm with high genomic integrity. This review explores recent innovations in multimodal AI for sperm screening, highlighting its potential to overcome the inherent limitations of traditional static morphological assessments. Furthermore, this paper addresses significant challenges, including data heterogeneity, model interpretability, and barriers to clinical translation. By systematically integrating machine learning, deep learning, and explainable AI techniques, multimodal AI offers promising strategies to improve ART outcomes through precise and comprehensive sperm analysis.

Background/Objectives: Surgical site infection (SSI) remains a major complication after colorectal surgery. Comparative evidence for commonly used perioperative intravenous antibiotic regimens is limited, and risk prediction models require external evaluation before clinical use. We examined regimen–SSI associations and secondarily evaluated models using routine perioperative variables to distinguish patients with and without electronically recorded postoperative SSI. Methods: We analysed 732 adults from MIMIC-IV for regimen association and model development and 133 adults from ICU-based eICU-CRD for external model evaluation. SSI was identified using an electronic phenotype. Four intravenous antibiotic exposure groups were compared using multivariable logistic regression and exploratory interaction analyses. Predictors selected using LASSO screening followed by clinical refinement informed six models, which were assessed using discrimination, calibration, and classification metrics. Results: SSI was recorded in 146/732 patients (19.9%), with no evidence of an unadjusted association across observed exposure categories (p = 0.947). Logistic regression achieved an external ROC-AUC of 0.858 (95% CI 0.764–0.934; PR-AUC 0.560) and detected 17/21 SSI cases, but generated 27 false positives. Positive predictive value was 38.6%, with limited calibration (expected calibration error 0.236) and a Brier score of 0.197. Conclusions: These data provide no evidence of an overall association, nor evidence of equivalent prophylactic efficacy. Heterogeneous exposures and residual confounding preclude regimen-level causal conclusions. The small ICU-based external evaluation informs higher-acuity populations, not routine elective surgery. Uncertain SSI onset, limited calibration, and false-positive burden preclude clinical implementation. Prospective multicentre studies should evaluate risk-adapted surveillance, not model-directed antibiotic selection.

Background: Postoperative delirium is a serious complication in elderly ICU patients after gastric surgery, increasing mortality and healthcare burden. Early identification of high-risk patients may support timely prevention and optimise perioperative care. Objective: To develop and externally validate an interpretable machine learning model for predicting postoperative delirium in elderly ICU patients undergoing gastric surgery. Methods: This retrospective multicohort study used three perioperative and critical care databases. MIMIC-IV supported model development and internal validation, while eICU and MOVER supported external validation. Feature selection used LASSO regression, followed by comparison of six machine learning algorithms. Logistic regression was selected using the MIMIC-IV internal validation cohort based on discrimination, recall, and clinical interpretability; the locked model was then evaluated in eICU and MOVER. SHAP analysis quantified predictor contributions and enhanced model transparency. Results: The final population included four cohorts: MIMIC-IV training ( N = 4316), MIMIC-IV internal validation ( N = 1851), eICU external testing ( N = 1745), and MOVER external testing ( N = 1152). The internally selected logistic regression model achieved AUCs of 0.819 in eICU and 0.827 in MOVER, with recalls of 0.827 and 0.936, respectively. SHAP analysis identified chloride, sodium, bicarbonate, and anion gap as dominant predictors across cohorts. Conclusion: This study presents an interpretable postoperative delirium risk-prediction model in elderly ICU patients following gastric surgery. Although the model achieved high recall, marked class imbalance and low outcome prevalence were associated with low precision and a substantial false-positive burden. Prospective validation, threshold optimisation, and assessment of clinical utility are required before implementation.