BMC Medical Informatics and Decision Making

Achieving adequate enteral nutrition among mechanically ventilated patients is challenging, yet critical. We developed NutriSighT, a transformer model using learnable positional coding to predict which patients would achieve hypocaloric nutrition between days 3-7 of mechanical ventilation. Using retrospective data from two large ICU databases (3,284 patients from AmsterdamUMCdb - development set, and 6,456 from MIMIC-IV - external validation set), we included adult patients intubated for at least 72 hours. NutriSighT achieved AUROC of 0.81 (95% CI: 0.81 - 0.82) and an AUPRC of 0.70 (95% CI: 0.70 - 0.72) on internal test set. External validation with MIMIC-IV data yielded a AUROC of 0.76 (95% CI: 0.75 - 0.76) and an AUPRC of (95% CI: 0.69 - 0.70). At a threshold of 0.5, the model achieved a 75.16% sensitivity, 60.57% specificity, 58.30% positive predictive value, and 76.88% negative predictive value. This approach may help clinicians personalize nutritional therapy among critically ill patients, improving patient outcomes.

BackgroundIn a previous study, Kowadlo et al. [1] developed algorithms (POP - Patient OPtimizer) to predict outcomes for surgical patients at Austin Health. The POP algorithms predict postoperative complications, kidney failure, and hospital length-of-stay. The findings highlight the potential of risk prediction to improve health outcomes and justify further work to improve performance and generalisability. ObjectivesThe objectives are to: O_LIEstablish and validate a causal graph of elective surgery in a hospital setting C_LIO_LITest whether causal inference can be used in algorithm development to improve generalisation of predictive models to different patient cohorts C_LIO_LIImplement and test the concept of preventable risk, risk stratification that combines risk prediction with causal effect; to assist in decision making C_LI MethodIn order to achieve the objectives, we will: O_LIApply causal discovery methods to the INSPIRE dataset (Lim et al. [2]), to create a causal graph C_LIO_LIValidate the graph by combining clinical input with data analysis, to identify relevant confounding and collider variables C_LIO_LIMethodically control for confounders and colliders, while training and evaluating predictive models for length-of-stay, mortality, readmission or complications C_LIO_LIMeasure the generalisation of predictive models across patient populations, when controlling for identified confounders and colliders C_LIO_LIImplement Conditional Average Treatment Estimate (CATE) and combine it with risk prediction to calculate preventable risk C_LI

BackgroundIntracerebral hemorrhage (ICH) is a critical form of stroke resulting from bleeding within the brain, with a mortality rate of 40-50% within a few days and significant risk of long-term disability. Despite the high incidence of ICU readmissions among ICH patients, the specific factors contributing to these readmissions remain unclear. This study utilizes MIMIC-III and MIMIC-IV databases to develop machine learning models that predict ICU readmissions in ICH patients. MethodsData from 2,144 patients were extracted using ICD-9 and ICD-10 codes. Four machine learning models - AdaBoost, Random Forest, XGBoost, and LightGBM - were implemented. Recursive Feature Elimination with Cross-Validation reduced features from 50 to 18 key predictors. The RandomUnderSampler technique addressed class imbalance by reducing majority class samples to 60% of the minority class, while the Optuna framework with Tree-structured Parzen Estimator optimized model parameters. Performance was primarily evaluated using AUROC, which effectively handles class imbalance and provides threshold-independent assessment, complemented by accuracy, sensitivity, and specificity metrics. ResultsThe AdaBoost model achieved an AUROC of 0.877 (95% CI: 0.815-0.913) and accuracy of 0.810, improving from the previous best AUROC of 0.736. Sensitivity notably increased from 22.6% to 84.0%, demonstrating substantial improvement in identifying high-risk patients. This enhancement resulted from effective class imbalance handling and AdaBoosts adaptive weighting mechanism. Through comprehensive SHAP and ablation analyses, we identified oxygen saturation and cardiovascular disease as crucial predictive features for ICU readmission risk assessment, providing new insights into patient monitoring and care management. ConclusionsOur preprocessing methodology and model selection strategy significantly improved high-risk ICH patient identification. Through comprehensive improvements combining advanced hyperparameter optimization, balanced sampling techniques, and dual feature importance analysis, we achieved a 19.2% AUROC improvement while reducing feature dimensionality from 51 to 18. This integrated approach demonstrates the potential of machine learning in enhancing clinical decision-making. The framework provides a promising foundation for developing clinical decision support tools in ICU settings, improving resource allocation and enabling more personalized patient care interventions.

The potential of utilizing Japanese electronic medical record (EMR) data in global observational research is significant because of high EMR adoption and universal health insurance. However, a few studies have addressed the conversion of Japanese EMR data to the Observational Medical Outcomes Partnership Common Data Model (OMOP CDM) standard, which regulates EMR data for global observational research. In this study, we investigated the feasibility of converting Japanese oncology EMR data to the OMOP CDM and applying the Observational Health Data Sciences and Informatics (OHDSI) tools for analysis. We focused on data from the National Cancer Center Hospital East, encompassing 8,447 patients with breast cancer between January 2015 and November 2023. The main objectives included vocabulary standardization and data structure standardization. The anonymized dataset included clinical information such as patient demographics, diagnoses, treatments, and laboratory results. A total of 3,697 unique disease names, 987 specimen test result terms, and 1,144 drug terms were successfully mapped to OMOP CDM standards, with IC-10 terms showing the highest success rate for disease names. A total of 90% of clinical terms were successfully mapped to OMOP CDM standards, with 80% of source data fully integrated. However, only 32 surgical terms were identified. The feasibility of converting EMR data to OMOP CDM was evaluated by mapping source terms, comparing local raw datasets, and conducting a comprehensive quality assessment using a Data Quality Dashboard. A total of 1,991 validation checks were performed to evaluate the validity of data, suitability, and completeness. The results revealed 24 checks flagged as FAIL or ERROR, with the most frequent issues in the measurement table (10 errors). Despite these issues, the conversion process demonstrated high feasibility. Overall, this study positions Japan as a key player in international observational oncology research, enhancing the global understanding of treatment effectiveness and patient outcomes in real-world settings.

Predictive accuracy of surgical case duration plays a critical role in reducing cost of operation room (OR) utilization. The most common approaches used by hospitals rely on historic averages based on a specific surgeon or a specific procedure type obtained from the electronic medical record (EMR) scheduling systems. However, low predictive accuracy of EMR leads to negative impacts on patients and hospitals, such as rescheduling of surgeries and cancellation. In this study, we aim to improve prediction of operation case duration with advanced machine learning (ML) algorithms. We obtained a large data set containing 170,748 operation cases (from Jan 2017 to Dec 2019) from a hospital. The data covered a broad variety of details on patients, operations, specialties and surgical teams. Meanwhile, a more recent data with 8,672 cases (from Mar to Apr 2020) was also available to be used for external evaluation. We computed historic averages from EMR for surgeon- or procedure-specific and they were used as baseline models for comparison. Subsequently, we developed our models using linear regression, random forest and extreme gradient boosting (XGB) algorithms. All models were evaluated with R-squre (R2), mean absolute error (MAE), and percentage overage (case duration > prediction + 10 % & 15 mins), underage (case duration < prediction - 10 % & 15 mins) and within (otherwise). The XGB model was superior to the other models by having higher R2 (85 %) and percentage within (48 %) as well as lower MAE (30.2 mins). The total prediction errors computed for all the models showed that the XGB model had the lowest inaccurate percent (23.7 %). As a whole, this study applied ML techniques in the field of OR scheduling to reduce medical and financial burden for healthcare management. It revealed the importance of operation and surgeon factors in operation case duration prediction. This study also demonstrated the importance of performing an external evaluation to better validate performance of ML models.

BackgroundThere is a growing demand for advanced methods to improve the understanding and prediction of illnesses. This study focuses on Sepsis, a critical response to infection, aiming to enhance early detection and mortality prediction for Sepsis-3 patients to improve hospital resource allocation. MethodsIn this study, we developed a Machine Learning (ML) framework to predict the 30-day mortality rate of ICU patients with Sepsis-3 using the MIMIC-III database. Advanced big data extraction tools like Snowflake were used to identify eligible patients. Decision tree models and Entropy Analyses helped refine feature selection, resulting in 30 relevant features curated with clinical experts. We employed the Light Gradient Boosting Machine (LightGBM) model for its efficiency and predictive power. ResultsThe study comprised a cohort of 9118 Sepsis-3 patients. Our preprocessing techniques significantly improved both the AUC and accuracy metrics. The LightGBM model achieved an impressive AUC of 0.983 (95% CI: [0.980-0.990]), an accuracy of 0.966, and an F1-score of 0.910. Notably, LightGBM showed a substantial 6% improvement over our best baseline model and a 14% enhancement over the best existing literature. These advancements are attributed to (I) the inclusion of the novel and pivotal feature Hospital Length of Stay (HOSP_LOS), absent in previous studies, and (II) LightGBMs gradient boosting architecture, enabling robust predictions with high-dimensional data while maintaining computational efficiency, as demonstrated by its learning curve. ConclusionsOur preprocessing methodology reduced the number of relevant features and identified a crucial feature overlooked in previous studies. The proposed model demonstrated high predictive power and generalization capability, highlighting the potential of ML in ICU settings. This model can streamline ICU resource allocation and provide tailored interventions for Sepsis-3 patients.

In recent years, most hospitals have implemented Electronic Health Records to manage and integrate a wide range of medical information, including diagnostics, medication admission and laboratory test results. Certain laboratory variables may serve as indicators of a patients clinical deterioration, making laboratory data a valuable tool for identifying high-risk patients. This work introduces a framework for predicting imminent health outcomes (IHO) of multimorbidity patients using laboratory test data. Our cohort includes 322,316 multimorbidity patients that performed laboratory tests in a large teaching hospital between January 2007 and August 2021. Two Imminent Health Outcomes predictive tools were developed. The first considers all patients in the dataset. The second was developed using a subset of patients with Heart Failure (HF) as the main comorbidity (5% of the entire dataset), considering that HF is a highly prevalent syndrome in multimorbidity patients. This predictive model achieved a reasonable predictive performance (AUROC = 0.718, 95% CI 0.708-0.756, and AUPRC = 0.663, 95% CI 0.630-0.701). C-reactive protein and NT-proBNP are the lab tests that most positively contribute to the prediction of IHO. The IHO predictive tool has the potential to help the medical team identify patients at high-risk of an imminent adverse event, highlighting the laboratory variables that are most contributing to the deterioration of the patient.

Medical practitioners record the condition status of a patient through qualitative and quantitative observations. The measurement of vital signs and molecular parameters in the clinics gives a complementary description of abnormal phenotypes associated with the progression of a disease. The Clinical Measurement Ontology (CMO) is used to standardize annotations of these measurable traits. However, researchers have no way to describe how these quantitative traits relate to phenotype concepts in a machine-readable manner. Using the WHO clinical case report form standard for the COVID-19 pandemic, we modeled quantitative traits and developed OWL axioms to formally relate clinical measurement terms with anatomical, biomolecular entities and phenotypes annotated with the Uber-anatomy ontology (Uberon), Chemical Entities of Biological Interest (ChEBI) and the Phenotype and Trait Ontology (PATO) biomedical ontologies. The formal description of these relations allows interoperability between clinical and biological descriptions, and facilitates automated reasoning for analysis of patterns over quantitative and qualitative biomedical observations.

ObjectivesPatient non-attendance at outpatient appointments is a major concern for healthcare providers. Non-attendances increase waiting lists, reduce access to care and may be detrimental not for the patient who did not attend. We aim to produce a model which can accurately predict which appointments will be attended. SettingA teaching hospital in London, UK combining secondary and tertiary care. ParticipantsA set of 9.6 million outpatient appointments between April 2015 and September 2019 including all ages and specialities. Primary and secondary outcome measuresArea under the receiver operating characteristic curve (AU-ROC) for prediction of outpatient appointment non-attendances. ResultsThe model uses 27 predictors to achieve an AUROC score of 0.768 (95% CI: 0.767-0.769) and accuracy of 89.2% (95% CI: 89.16%-89.24%) on test data. We find that the waiting period between booking and the appointment, the patients past attendance behaviour, and the levels of deprivation in their local area are important factors in predicting future attendance. ConclusionOur model successfully predicts patient attendance at outpatient appointments. Its performance on both patients who did not appear in the training data and appointments from a different time period which covers the Covid-19 pandemic indicate it generalized well across both face to face and virtual appointments and could be used to target resources and intervention towards those patients who are likely to miss an appointment. Moreover, it highlights the impact of deprivation on patient access to healthcare Strengths and Limitation of this StudyO_LIWe make use of a large dataset which enables us to use complex machine learning algorithms. C_LIO_LIWe validate the model on two large, distinct datasets giving high confidence in our model performance. C_LIO_LIAn unknown amount of patient data is missing due to a nearby hospital which shares patients with the study setting. C_LI

Patient length of stay (LoS) is a critical metric for evaluating the efficacy of hospital management. The primary objectives encompass to improve efficiency and reduce costs while enhancing patient outcomes and hospital capacity within the patient journey. By seamlessly merging data-driven techniques with simulation methodologies, the study proposes an all-encompassing framework for the optimization of patient flow. Using a comprehensive dataset of 2.3 million de-identified patient records, we analyzed demographics, diagnoses, treatments, services, costs, and charges with machine learning models (Decision Tree, Logistic Regression, Random Forest, Adaboost, LightGBM) and Python tools (Spark, AWS clusters, dimensionality reduction). Our model predicts patient length of stay (LoS) upon admission using supervised learning algorithms. This hybrid approach enables the identification of key factors influencing LoS, offering a robust framework for hospitals to streamline patient flow and resource utilization. The research focuses on patient flow corroborates the efficacy of the approach, illustrating decreased patient length of stay within a real healthcare environment. The findings underscore the potential of hybrid data-driven models in transforming hospital management practices. This innovative methodology provides generally flexible decision-making, training, and patient flow enhancement; such a system could have huge implications for healthcare administration and overall satisfaction with healthcare.

ObjectiveTo evaluate the feasibility of continuous paO2 prediction in an intraoperative setting among neurosurgical patients with modern machine learning methods. Materials and MethodsData were extracted from routine clinical care of lung-healthy, neurosurgical patients. We used recursive feature elimination to identify relevant features for the prediction of paO2. Five machine learning algorithms (gradient boosting regressor, k-nearest neighbors regressor, random forest regressor, support vector regression, multi-layer perceptron regressor) and a multivariate linear regression were then tuned and fitted to the selected features. A performance matrix consisting of Spearmans {rho}, mean absolute percentage error (MAPE) and root mean squared error (RMSE) was finally computed based on the test set and used to compare and rank each algorithm. ResultsWe analyzed 4,180 patients with 12,497 observations. A total of 20 features were selected from analysis of the training dataset comprising 836 patients with 9,992 observations. The best algorithm, random forest, was able to predict paO2 values with {rho}=0.90, MAPE=10.4%, and RMSE=30.9mmHg, closely followed by gradient boosting and multi-layer perceptron. Support vector regression, k-nearest neighbors regressor and the linear regression did not achieve the performance metrics. DiscussionWe successfully applied and compared several machine learning algorithms to estimate continuous paO2 values in neurosurgical patients. The random forest regressor performed best over all three categories of the performance matrix. ConclusionPaO2 can be predicted by perioperative routine data in neurosurgical patients.

BackgroundElectronic Health Records face a fundamental challenge: the semantic gap between relational data storage and clinical reasoning patterns. Traditional databases struggle with complex healthcare queries requiring multiple joins and temporal analysis, creating performance bottlenecks that limit real-time clinical applications. MethodsWe developed a Neo4j-based framework integrating MIMIC-IV clinical data (1,504 patients, 4,967 admissions) with SNOMED CT medical ontology through ICD-10-CM mappings. The implementation created a unified graph comprising 625,708 nodes and 2,189,093 relationships, with systematic preservation of temporal and semantic connections. ResultsPerformance analysis demonstrated substantial improvements over PostgreSQL across five query types, with Neo4j showing 5.4x to 48.4x faster execution times. The framework successfully enabled three clinical applications: ventilator-associated pneumonia temporal analysis (revealing 47.79% pneumonia rates among ventilated ICU stays), hypertension semantic network mapping through multi-level SNOMED-CT relationships, and Medicare Part D quality measure monitoring. Notably, the system identified that 96.7% of eligible diabetic patients lacked statin prescriptions, demonstrating practical utility for healthcare quality improvement initiatives. ConclusionThis graph-based approach provides a robust foundation for next-generation clinical decision support systems by bridging the gap between fragmented clinical data and integrated patient-centric analysis. The frameworks demonstrated performance advantages and practical applications in quality measure monitoring establish its potential for addressing real-world healthcare challenges while supporting the transition toward more effective, evidence-based patient care.

BackgroundThe digitisation of healthcare records has generated vast amounts of unstructured data, presenting opportunities for improvements in disease diagnosis when clinical coding falls short, such as in the recording of patient symptoms. This study presents an approach using natural language processing to extract clinical concepts from free-text which are used to automatically form diagnostic criteria for lung cancer from unstructured secondary-care data. MethodsPatients aged 40 and above who underwent a chest x-ray (CXR) between 2016-2022 were included. ICD-10 and unstructured data were pulled from their electronic health records (EHRs) over the preceding 12 months to the CXR. The unstructured data were processed using named entity recognition to extract symptoms, which were mapped to SNOMED-CT codes. Subsumption of features up the SNOMED-CT hierarchy was used to mitigate against sparse features and a frequency-based criteria, combined with univariate logarithmic probabilities, was applied to select candidate features to take forward to the model development phase. A genetic algorithm was employed to identify the most discriminating features to form the diagnostic criteria. Results75002 patients were included, with 1012 lung cancer diagnoses made within 12 months of the CXR. The best-performing model achieved an AUROC of 0.72. Results showed that an existing disorder of the lung, such as pneumonia, and a cough increased the probability of a lung cancer diagnosis. Anomalies of great vessel, disorder of the retroperitoneal compartment and context-dependent findings, such as pain, statistically reduced the risk of lung cancer, making other diagnoses more likely. The performance of the developed model was compared to the existing cancer risk scores, demonstrating superior performance. ConclusionsThe proposed methods demonstrated success in leveraging unstructured secondary-care data to derive diagnostic criteria for lung cancer, outperforming existing risk tools. These advancements show potential for enhancing patient care and results. However, it is essential to tackle specific limitations by integrating primary care data to ensure a more thorough and unbiased development of diagnostic criteria. Moreover, the study highlights the importance of contextualising SNOMED-CT concepts into meaningful terminology that resonates with clinicians, facilitating a clearer and more tangible understanding of the criteria applied.

ObjectiveThe aim of this study was to investigate predictive capabilities of historical records of patients maintained at hospitals towards predicting an impending adverse outcomes such as, mortality, readmission, and prolonged length of stay (PLOS). MethodsLeveraging a de-identified dataset from a tertiary care university hospital, we developed a eXplainable Artificial Intelligence (XAI) framework combining tree-based and traditional ML models with interpretations, and statistical analysis of predictors of mortality, readmission, and PLOS. ResultsOur framework demonstrated exceptional predictive performance with notable Area Under the Receiver Operating Characteristic (AUROC) of 0.9625 and Area Under the Precision-Recall Curve (AUPRC) of 0.8575 for 30-day mortality at discharge and an AUROC of 0.9545 and AUPRC of 0.8419 at admission. For the readmission and PLOS risk the highest AUROC achieved were 0.8198 and 0.9797 repectively. The tree-based machine learning (ML) models consistently outperformed the traditional ML models in all the four prediction tasks. The key predictors were age, derived temporal features, routine laboratory tests, and diagnostic and procedural codes. ConclusionThe study underscores the potential of leveraging medical history for enhanced predictive analytics in hospitals. We present a accurate and intuitive framework for early warning models that can be easily implemented in the current and developing digital health platforms to accurately predict adverse outcomes.

Through the utilisation of algorithms and data-driven analytics, predictive technology can be leveraged to provide clinicians with invaluable insight into their patients conditions, allowing for more accurate, informed, and timely decision-making in the fast-paced clinical environment. This study aims to provide a preliminary proof of concept with a double-validated real-world dataset from the UK. The YouDiagnose predictive model, developed using retrospective data from over 41,257 patients data, was assessed by testing it with a double-validated real-world dataset from the UK and the machine predictions have been compared here with the final diagnosis of the diseases as the gold standard. Out of the total of 433 cases, 60 cases had a mismatch in their prediction, all of them being cancer overdiagnosis, resulting in a lower specificity rate of 84.3%. The combined prediction accuracy at the first prediction level was 86% (n=373) while prediction 1-3 combined was successful in predicting diseases in 93% of the cases when evaluated against the gold standard. The model accurately predicted all 52 cases of cancer, indicating a 100% sensitivity rate. The study shows that the tool can be used in the frontline to accurately screen patients with a high level of confidence in the inclusion of cancer patients. This tools high sensitivity means that there is little chance of missing any cancer cases.

BackgroundAccurate prediction of postoperative outcomes, particularly 30-day in-hospital mortality, is crucial for improving surgical planning, patient counseling, and resource allocation. This study aimed to develop and validate a logistic regression model to predict 30-day in-hospital mortality using comprehensive perioperative data from the INSPIRE dataset. MethodsWe conducted a retrospective analysis of the INSPIRE dataset, comprising approximately 130,000 surgical cases from Seoul National University Hospital between 2011 and 2020. The primary objective was to develop a logistic regression model using preoperative and intraoperative variables. Key predictors included demographic information, clinical variables, laboratory values, and the emergency status of the operation. Missing data were addressed through multiple imputation, and feature selection was performed using univariate analysis and clinical judgment. The model was validated using cross-validation and assessed for performance using ROC AUC and precision-recall AUC metrics. ResultsThe logistic regression model demonstrated high predictive accuracy, with an ROC AUC of 0.978 and a precision-recall AUC of 0.958. Significant predictors of 30-day in-hospital mortality included emergency status of the operation (OR: 1.56), preoperative prothrombin time (PT/INR) (OR: 1.53), potassium levels (OR: 1.49), body mass index (BMI) (OR: 1.37), serum sodium (OR: 1.11), creatinine levels (OR: 1.04), and albumin levels (OR: 0.85). ConclusionThis study successfully developed and validated a logistic regression model to predict 30-day in-hospital mortality using comprehensive perioperative data. The models high predictive accuracy and reliance on routinely collected clinical and laboratory data enhance its feasibility for integration into existing clinical workflows, providing real-time risk assessments to healthcare providers. Future research should focus on external validation in diverse clinical settings and prospective studies to assess the practical impact of this predictive model.

ObjectiveVancomycin is a widely used antibiotic that requires therapeutic drug monitoring (TDM) for optimized individual dosage. The deep learning-based model PKRNN-1CM has shown the advantage of leveraging time series electronic health record (EHR) data for individualized estimation of vancomycin pharmacokinetic (PK) parameters. While one-compartment (1CM) PK models are commonly used because of their simplicity and previous trough-based clinical practices for dose adjustment, the pre-deep learning literature suggests the superiority of two-compartment models (2CM). Motivated by this, we introduce a novel deep-learning-based approach, PKRNN-2CM, for vancomycin TDM. MethodsPKRNN-2CM combines RNN-driven PK parameter estimation with a 2CM PK model to predict vancomycin concentration trajectories. Training on both simulated data and real-world EHR data allows for a comprehensive evaluation of its performance. ResultsExperiments based on simulated data highlight PKRNN-2CMs superiority over the simpler 1CM model PKRNN-1CM (PKRNN-2CM RMSE=1.30, PKRNN-1CM RMSE=2.50). Application to real data showcases significant improvement over PKRNN-1CM (PKRNN-2CM RMSE=5.62, PKRNN-1CM RMSE=5.84, two-sample unpaired t-test p-value=0.01), with potential further gains expected with non-trough level measurements. ConclusionPKRNN-2CM is an important improvement in vancomycin TDM, demonstrating enhanced accuracy and performance compared to the PKRNN-1CM model. This deep learning model holds potential for future individualized vancomycin TDM optimization and broader application in diverse clinical scenarios.

BackgroundSepsis-Associated Acute Kidney Injury (SA-AKI) leads to high mortality in intensive care. This study develops machine learning models using the Medical Information Mart for Intensive Care IV (MIMIC-IV) database to predict Intensive Care Unit (ICU) mortality in SA-AKI patients. External validation is conducted using the eICU Collaborative Research Database. MethodsFor 9,474 identified SA-AKI patients in MIMIC-IV, key features like lab results, vital signs, and comorbidities were selected using Variance Inflation Factor (VIF), Recursive Feature Elimination (RFE), and expert input, narrowing to 24 predictive variables. An Extreme Gradient Boosting (XGBoost) model was built for in-hospital mortality prediction, with hyperparameters optimized using GridSearch. Model interpretability was enhanced with SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). External validation was conducted using the eICU database. ResultsThe proposed XGBoost model achieved an internal Area Under the Receiver Operating Characteristic curve (AUROC) of 0.878 (95% Confidence Interval: 0.859-0.897). SHAP identified Sequential Organ Failure Assessment (SOFA), serum lactate, and respiratory rate as key mortality predictors. LIME highlighted serum lactate, Acute Physiology and Chronic Health Evaluation II (APACHE II) score, total urine output, and serum calcium as critical features. ConclusionsThe integration of advanced techniques with the XGBoost algorithm yielded a highly accurate and interpretable model for predicting SA-AKI mortality across diverse populations. It supports early identification of high-risk patients, enhancing clinical decision-making in intensive care. Future work needs to focus on enhancing adaptability, versatility, and real-world applications. Graphical Abstract HighlightsO_LIThe study implemented a robust machine learning pipeline for predicting ICU mortality in sepsis-associated acute kidney injury (SA-AKI) patients. This pipeline included advanced data preprocessing techniques, stratified imputation for handling missing values, and a three-stage feature selection strategy using Variance Inflation Factor (VIF), Recursive Feature Elimination (RFE), and expert clinical input. The optimized feature set was then used to train an XGBoost model with hyperparameter tuning via GridSearchCV, achieving high predictive accuracy with an AUROC of 0.878 (95% CI: 0.859-0.897) and enhanced clinical applicability. The interpretability analysis using SHAP and LIME identified critical features such as SOFA score, serum lactate, and respiratory rate as key mortality predictors. C_LIO_LIThe model was externally validated using the eICU Collaborative Research Database, confirming its generalizability and robustness across diverse patient populations with an AUROC of 0.720 (95% CI: 0.708-0.733). This transparent, data-driven approach supports early identification of high-risk patients, optimizing clinical decision-making and resource allocation in intensive care settings. C_LI

ObjectivePediatric heart transplantation is challenged by limited donor organ availability, prolonged waitlist times, and elevated risks of late acute rejection (LAR) and hospitalization. Current predictive models for post-transplant outcomes lack high accuracy due to reliance on registry data without integrating dynamic clinical and social factors. This study aimed to improve predictive performance and model interpretability by incorporating electronic health records (EHR), social determinants of health (SDoH), and United Network for Organ Sharing (UNOS) data. Materials and MethodsWe used EHR and UNOS data from 111 pediatric heart transplant patients (ages 0-18) at the University of Florida Health Shands Childrens Hospital to build predictive models for organ rejection at 1-, 3-, and 5-year intervals post-transplant. UNOS data includes pre- and post-transplant health and medical records, encompassing procedures, clinical evaluations, and post-transplant follow-up information, EHR data included evolving clinical parameters (e.g., comorbidities, medication adherence, and laboratory results), while SDoH encompassed socioeconomic status, living conditions, and healthcare access. Feature importance was assessed using Shapley Variable Importance Cloud (ShapleyVIC), which integrates Shapley Additive Explanations (SHAP) to provide robust, interpretable insights across nearly optimal models. ResultsModels integrating EHR, SDoH, and UNOS data outperformed those using UNOS data alone, with AUROC of 0.743 (0.607-0.879), 0.798 (0.725-0.871), and 0.760 (0.692-0.828). Key predictors of rejection included severe pre-transplant conditions (e.g., life support, prolonged waitlist times), elevated bilirubin and creatinine levels, and social factors (e.g., transportation barriers, BMI, insurance type). DiscussionFindings reveal the importance of integrating clinical and social data to address multisystem dysfunction, disparities in healthcare access, and adherence challenges. ShapleyVIC enhanced model interpretability, providing actionable insights for improving post-transplant care. ConclusionHolistic, data-driven approaches that combine EHR, SDoH, and registry data significantly enhance predictive accuracy and interpretability, supporting improved long-term outcomes for pediatric heart transplant patients.

Acute Myeloid Leukemia (AML) is a complex disease requiring accurate risk stratification for effective treatment planning. This study introduces an innovative ensemble machine learning model integrated with the European LeukemiaNet (ELN) 2022 recommendations to enhance AML risk stratification. The model demonstrated superior performance by utilizing a comprehensive dataset of 1,213 patients from National Taiwan University Hospital (NTUH) and an external cohort of 2,113 patients from UK-NCRI trials. On the external cohort, it improved a concordance index (c-index) from 0.61 to 0.64 and effectively distinguished three different risk levels with median hazard ratios ranging from 18% to 50% improved. Key insights were gained from the discovered significant features influencing risk prediction, including age, genetic mutations, and hematological parameters. Notably, the model identified specific cytogenetic and molecular alterations like TP53, IDH2, SRSF2, STAG2, KIT, TET2, and karyotype (-5, -7, -15, inv(16)), alongside age and platelet counts. Additionally, the study explored variations in the effectiveness of hematopoietic stem cell transplantation (HSCT) across different risk levels, offering new perspectives on treatment effects. In summary, this study develops an ensemble model based on the NTUH cohort to deliver improved performance in AML risk stratification, showcasing the potential of integrating machine learning techniques with medical guidelines to enhance patient care and personalized medicine.