Development and validation of a bedside-available machine learning model to predict discrepancies between SaO₂ and SpO₂: Exploring factors related to the discrepancies.

In critically ill patients, a discrepancy frequently exists between percutaneous oxygen saturation (SpO₂) and arterial blood oxygen saturation (SaO₂), which can lead to potential hypoxemia being overlooked. The aim of this study was to explore the factors related to the discrepancy and to develop an easy-to-use prediction model that uses readily available bedside information to predict the discrepancy and suggest the need for arterial blood gas measurement. This is a prognostic study that used eICU Collaborative Research Database from 2014 to 2015 for model development and MIMIC-IV data from 2008 to 2019 for model validation. To predict the outcome of SpO₂ exceeding SaO₂ by 3% or more, non-invasive, readily available bedside information (patient demographics, vital signs, vasopressor use, ventilator use) was used to develop prediction models with three machine learning methods (decision tree, logistic regression, XGBoost). To make the model accessible, the model was deployed as a web-based application. Additionally, the contribution of each variable was explored using partial dependence plots and SHAP values. From 4,781 admission records in eICU data, a total of 19,804 paired SpO₂ and SaO₂ measurements were used. Among three machine learning models, the XGBoost model demonstrated the best predictive performance with an AUROC of 0.73 and a calibration slope of 0.90. In the validation cohort of MIMIC-IV paired dataset, the performance was AUROC of 0.56. An exploratory model-updating step followed by temporal validation raised performance to AUROC of 0.70 with a calibration slope of 0.85. In both datasets, worse vital signs were associated with the discrepancy (e.g., low blood pressure, low temperature) between SpO₂ and SaO₂. Using non-invasive bedside data, a machine learning model was developed to predict SpO₂–SaO₂ discrepancy and identified vital signs as key contributors. These findings underscore the awareness for hidden hypoxemia and provide the basis of further study to accurately evaluate the actual SaO₂.

Background: Traumatic brain injury (TBI) induced coagulopathy (TIC) patients, is a disease with poorer prognosis and increased mortality rate. Objectives: Our study aimed to identify predictors as well as develop machine learning (ML) models to predict the risk of coagulopathy in this population. Methods: ML models were developed and validated based on two public databases named Medical Information Mart for Intensive Care (MIMIC)-IV and the eICU Collaborative Research Database (eICU-CRD). Candidate predictors, including demographics, personal history, comorbidities, vital signs, laboratory findings, injury type, therapy strategy and scoring system were included. Models were compared on area under the curve (AUC), accuracy, sensitivity, specificity, positive and negative predictive values, and decision curve analysis (DCA) curve. Results: Of 999 patients in MIMIC-IV included in the final cohort, a total of 493 (49.35%) patients developed coagulopathy following TBI. Recursive feature elimination (RFE) selected 15 variables, including international normalized ratio (INR), prothrombin time (PT), sepsis related organ failure assessment (SOFA), activated partial thromboplastin time (APTT), platelet (PLT), hematocrit (HCT), red blood cell (RBC), hemoglobin (HGB), blood urea nitrogen (BUN), red blood cell volume distribution width (RDW), creatinine (CRE), congestive heart failure, myocardial infarction, sodium, and blood transfusion. The external validation in eICU-CRD demonstrated that adapting boosting (Ada) model had the highest AUC of 0.924 (95% CI: 0.902–0.943). Furthermore, in the decision curve analysis, the Ada model had a higher net benefit (ie, the correct classification of coagulopathy considering a trade-off between false- negatives and false-positives)—over other models across a range of threshold probability values. Conclusions: The ML models, as indicated by our study, can be used to predict the incidence of TBI-induced coagulopathy (TBI-IC) in the ICU.

Background: Traumatic brain injury-induced coagulopathy (TBI-IC), is a disease with poor prognosis and increased mortality rate.Objectives: Our study aimed to identify predictors as well as develop machine learning (ML) models to predict the risk of coagulopathy in this population.Methods: ML models were developed and validated based on two public databases named Medical Information Mart for Intensive Care (MIMIC)-IV and the eICU Collaborative Research Database (eICU-CRD). Candidate predictors, including demographics, family history, comorbidities, vital signs, laboratory findings, injury type, therapy strategy and scoring system were included. Models were compared on area under the curve (AUC), accuracy, sensitivity, specificity, positive and negative predictive values, and decision curve analysis (DCA) curve.Results: Of 999 patients in MIMIC-IV included in the final cohort, a total of 493 (49.35%) patients developed coagulopathy following TBI. Recursive feature elimination (RFE) selected 15 variables, including international normalized ratio (INR), prothrombin time (PT), sepsis related organ failure assessment (SOFA), activated partial thromboplastin time (APTT), platelet (PLT), hematocrit (HCT), red blood cell (RBC), hemoglobin (HGB), blood urea nitrogen (BUN), red blood cell volume distribution width (RDW), creatinine (CRE), congestive heart failure, myocardial infarction, sodium, and blood transfusion. The external validation in eICU-CRD demonstrated that adapting boosting (Ada) model had the highest AUC of 0.924 (95% CI: 0.902–0.943). Furthermore, in the DCA curve, the Ada model and the extreme Gradient Boosting (XGB) model had relatively higher net benefits (ie, the correct classification of coagulopathy considering a trade-off between false- negatives and false-positives)—over other models across a range of threshold probability values.Conclusions: The ML models, as indicated by our study, can be used to predict the incidence of TBI-IC in the intensive care unit (ICU).

As more machine learning and deep learning models are applied in studying the quantitative relationship between the climate and terrestrial vegetation growth, the uncertainty of these advanced models requires clarification. Partial dependence plots (PDPs) are one of the most widely used methods to estimate the marginal effect of independent variables on the predicted outcome of a machine learning model, and it is regarded as the main basis for conclusions in relevant research. As more controversies regarding the reliability of the results of the PDPs emerge, the uncertainty of the PDPs remains unclear. In this paper, we experiment with real, remote sensing data to systematically analyze the uncertainty of partial dependence relationships between four climate variables (temperature, rainfall, radiation, and windspeed) and vegetation growth, with one conventional linear model and six machine learning models. We tested the uncertainty of the PDP curves across different machine learning models from three aspects: variation, whole linear trends, and the trait of change points. Results show that the PDP of the dominant climate factor (mean air temperature) and vegetation growth parameter (indicated by the normalized difference vegetation index, NDVI) has the smallest relative variation and the whole linear trend of the PDP was comparatively stable across the different models. The mean relative variation of change points across the partial dependence curves of the non-dominant climate factors (i.e., radiation, windspeed, and rainfall) and vegetation growth ranged from 8.96% to 23.8%, respectively, which was much higher than those of the dominant climate factor and vegetation growth. Lastly, the model used for creating the PDP, rather than the relative importance of these climate factors, determines the fluctuation of the PDP output of these climate variables and vegetation growth. These findings have significant implications for using remote sensing data and machine learning models to investigate the quantitative relationships between the climate and terrestrial vegetation.

BackgroundEarly prediction of hospital mortality is crucial for ICU patients with sepsis. This study aimed to develop a novel blending machine learning (ML) model for hospital mortality prediction in ICU patients with sepsis.MethodsTwo ICU databases were employed: eICU Collaborative Research Database (eICU-CRD) and Medical Information Mart for Intensive Care III (MIMIC-III). All adult patients who fulfilled Sepsis-3 criteria were identified. Samples from eICU-CRD constituted training set and samples from MIMIC-III constituted test set. Stepwise logistic regression model was used for predictor selection. Blending ML model which integrated nine sorts of basic ML models was developed for hospital mortality prediction in ICU patients with sepsis. Model performance was evaluated by various measures related to discrimination or calibration.ResultsTwelve thousand five hundred fifty-eight patients from eICU-CRD were included as the training set, and 12,095 patients from MIMIC-III were included as the test set. Both the training set and the test set showed a hospital mortality of 17.9%. Maximum and minimum lactate, maximum and minimum albumin, minimum PaO2/FiO2 and age were important predictors identified by both random forest and extreme gradient boosting algorithm. Blending ML models based on corresponding set of predictors presented better discrimination than SAPS II (AUROC, 0.806 vs. 0.771; AUPRC 0.515 vs. 0.429) and SOFA (AUROC, 0.742 vs. 0.706; AUPRC 0.428 vs. 0.381) on the test set. In addition, calibration curves showed that blending ML models had better calibration than SAPS II.ConclusionsThe blending ML model is capable of integrating different sorts of basic ML models efficiently, and outperforms conventional severity scores in predicting hospital mortality among septic patients in ICU.

In-hospital mortality prediction for patients admitted to the ICU remains a critical challenge in the field of critical care medicine. This systematic review evaluates the application of artificial intelligence (AI) and machine learning (ML) models for predicting in-hospital mortality in ICU settings. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, we analyzed 15 studies published between January 2015 and April 2025 that utilized AI and ML approaches for mortality prediction in ICU populations. The most commonly employed algorithms were extreme gradient boosting (XGBoost), random forest, and logistic regression, with data predominantly sourced from two major publicly available critical care databases: the Medical Information Mart for Intensive Care (MIMIC) and the eICU Collaborative Research Database (eICU-CRD).Across all studies, AI and ML models consistently outperformed traditional clinical scoring systems such as the Acute Physiology and Chronic Health Evaluation (APACHE), Sequential Organ Failure Assessment (SOFA), and Simplified Acute Physiology Score (SAPS), demonstrating superior discriminative performance in mortality prediction. Ensemble methods, particularly random forest and XGBoost, generally achieved the highest predictive accuracy, while deep learning approaches such as recurrent neural networks showed particular promise for analyzing temporal trends in physiological data. Key predictive features identified across multiple studies included patient age, vital signs (especially heart rate and blood pressure), laboratory values (particularly markers of renal function), and metrics of neurological status such as level of consciousness. Importantly, several studies developed models that required only routinely collected clinical data available within the first 24 hours of ICU admission, demonstrating the feasibility of early risk stratification using AI and ML.Although most research remains retrospective in nature and confined to a limited number of datasets, the consistent performance advantages observed across diverse modeling approaches underscore the significant clinical potential of AI and ML in ICU mortality prediction. Future research should prioritize the use of standardized model development and reporting methodologies, prospective validation in diverse and real-world clinical settings, exploration of integration and implementation challenges, and rigorous assessments of clinical impact. Such efforts are critical to translating these promising predictive technologies into improved decision-making processes and outcomes for critically ill patients.

BackgroundThe goal of this study was to assess the effectiveness of machine learning models and create an interpretable machine learning model that adequately explained 3-year all-cause mortality in patients with chronic heart failure.MethodsThe data in this paper were selected from patients with chronic heart failure who were hospitalized at the First Affiliated Hospital of Kunming Medical University, from 2017 to 2019 with cardiac function class III-IV. The dataset was explored using six different machine learning models, including logistic regression, naive Bayes, random forest classifier, extreme gradient boost, K-nearest neighbor, and decision tree. Finally, interpretable methods based on machine learning, such as SHAP value, permutation importance, and partial dependence plots, were used to estimate the 3-year all-cause mortality risk and produce individual interpretations of the model's conclusions.ResultIn this paper, random forest was identified as the optimal aools lgorithm for this dataset. We also incorporated relevant machine learning interpretable tand techniques to improve disease prognosis, including permutation importance, PDP plots and SHAP values for analysis. From this study, we can see that the number of hospitalizations, age, glomerular filtration rate, BNP, NYHA cardiac function classification, lymphocyte absolute value, serum albumin, hemoglobin, total cholesterol, pulmonary artery systolic pressure and so on were important for providing an optimal risk assessment and were important predictive factors of chronic heart failure.ConclusionThe machine learning-based cardiovascular risk models could be used to accurately assess and stratify the 3-year risk of all-cause mortality among CHF patients. Machine learning in combination with permutation importance, PDP plots, and the SHAP value could offer a clear explanation of individual risk prediction and give doctors an intuitive knowledge of the functions of important model components.

Machine learning (ML) has shown its effectiveness in handling multi-geoinformation. Yet, the black-box nature of ML algorithms has restricted their widespread adoption in the domain of mineral prospectivity mapping (MPM). In this paper, methods for interpreting ML model predictions are introduced to aid ML-based MPM, with the goal of extracting richer insights from the ML modeling of an exploration geochemical dataset. The partial dependence plot (PDP) and accumulated local effect (ALE) plot, along with the SHAP value analysis, were utilized to demonstrate the application of random forest (RF) modeling within both regression and classification frameworks. Initially, the random forest regression (RFR) model established the relationship between the concentrations of Au and those of elements such as As, Sb, and Hg in the study area, and from this model, the most important geochemical elements and their quantitative relationships with Au were revealed by their contributions in the modeling through PDP and ALE analyses. Secondly, the RF classification modeling established the relationships of mineralization occurrences (i.e., known mineral deposits) with geochemical elements (i.e., Au, As, Sb, Hg, Cu, Pb, Zn, and Ag), as did RFR modeling. The most important geochemical elements for indicating regional Au mineralization and the trajectories of PDP and ALE reached a consensus that As and Sb contributed the most, both in the regression and classification modeling, with regard to Au mineralization. Finally, the SHAP values illustrated the behavior of the training samples (i.e., known mineral deposits) in RF modeling, and the resulting prospectivity map was evaluated using receiver operating characteristics.

On the interpretability of machine learning methods in crash frequency modeling and crash modification factor development

Introduction: Congestive heart failure (CHF) is one of the most common diagnoses for patients admitted to an intensive care unit (ICU), and intubation or ventilation may be necessary to combat severe associated respiratory symptoms. Current guidelines give high-level recommendations to intubate patients with a Glasgow Coma Scale (GCS) score of 8 or lower. However, prior studies have shown that a large portion of patients who were ultimately intubated had GCS scores greater than 8, demonstrating the need for a better tool to predict the necessity of intubation. The present study sought to build machine-learning (ML) models that could predict the need for intubation or ventilation for CHF patients upon ICU admission. Methods: 5,623 patients (of which 1,058 were intubated and 2,816 were ventilated) in the eICU Collaborative Research Database who underwent screening using the Acute Physiology and Chronic Health Evaluation and were diagnosed with CHF were used to train (50%) and test (50%) the ML models. Gender, age, treatment status, readmission status, medication status upon admission, GCS motor score upon admission, and GCS eyes score upon admission were used as feature inputs for the ML models, which were classic three-layer neural networks. For intubation, the ML model performance was compared against using the common heuristic of intubating patients with a GCS score of 8 or less. Results: The ML model for intubation upon admission prediction had an accuracy of 84.8% and an area under the receiver operating characteristic curve (AUC) of 0.79. Meanwhile, using the aforementioned heuristic for intubation prediction yields an accuracy of 85.7% (AUC n/a). A second ML model was trained to predict the need for ventilation upon admission and had an accuracy of 71.2% and an AUC of 0.80. Conclusion: This newly developed ML model for intubation demonstrates relative success in predicting intubation needs for patients with CHF upon ICU admission. In addition, this model performs similarly to currently accepted guidelines. Further improvement and a larger dataset could better train the model to provide more accurate predictions. Such predictions could help support medical decision-making and resource allocation planning.

Pacemaker implantation (PMI) after transcatheter aortic valve implantation (TAVI) is a common complication. While computed tomography (CT) scan data are known predictors of PMI, no machine learning (ML) model integrating CT with clinical, ECG, and transthoracic echocardiography (TTE) data has been proposed. This study investigates the contribution of ML methods to predict PMI after TAVI, with a focus on the role of CT imaging data. A retrospective analysis was conducted on a cohort of 520 patients who underwent TAVI. Recursive feature elimination with SHAP values was used to select key variables from clinical, ECG, TTE, and CT data. Six ML models, including Support Vector Machines (SVM), were trained using these selected variables. The model’s performance was evaluated using AUC-ROC, F1 score, and accuracy metrics. The PMI rate was 18.8%. The best-performing model achieved an AUC-ROC of 92.1% ± 4.7, an F1 score of 71.8% ± 9.9, and an accuracy of 87.9% ± 4.7 using 22 variables, 9 of which were CT-based. Membranous septum measurements and their dynamic variations were critical predictors. Our ML model provides robust PMI predictions, enabling personalized risk assessments. The model is implemented online for broad clinical use.

BackgroundIn-hospital cardiac arrest (IHCA) is a major adverse event with a high death risk. Machine learning (ML) models of prognosis in cardiac arrest (CA) patients have been established, but there are some interferences in their clinical application. This study developed an ensemble learning (EL) model based on clinical information to predict IHCA patient death risk.Methods and resultsThis retrospective cohort study used data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and eICU Collaborative Research Database. Patients (age ≥ 18 years) with CA based on the ICD-9/10 code were included. Eight candidate ML models were selected for soft voting ensemble. Features were sequentially eliminated based on feature importance scoring to reduce input complexity without compromising model performance. The final model was externally validated with the MIMIC-IV database and deployed as a web application. Overall, 4,068 patients were included. In the internal validation cohort, the EL model exceeded single ML models with an accuracy of 0.842, precision of 0.830, recall of 0.839, F1 score of 0.835, and AUC of 0.898 and showed better calibration across the spectrum of survival probabilities. Furthermore, there is no obvious decline in the prediction performance of the EL model with the top seven features (HCO3−, Glasgow Coma Scale, white blood cell count, international normalized ratio, hematocrit, body temperature, and blood urea nitrogen) retained. In external validation, the performance slightly decreased but remained acceptable for deploying a clinically feasible web application.ConclusionThe EL model outperformed single ML models in predicting IHCA patient death risk. The identified seven key features enabled the parsimonious EL model to reliably estimate the death risk.

Ultra-high-performance concrete (UHPC) bonded to normal concrete (NC) can significantly enhance the mechanical performance of UHPC–NC composite structures, and the interface shear strength is a crucial indicator for assessing the bonding performance. In this study, interpretable machine learning (ML) methods were used to analyse the effects of different parameters on interface shear strength. A database consisting of 305 UHPC–NC shear tests was created, and the isolation forest algorithm was applied to filter outliers. Subsequently, four ML models were trained to predict the interface shear strength of UHPC–NC composite structures. Among them, the extreme gradient boosting (XGBoost) model demonstrated the highest prediction accuracy, achieving an R2 value of 0.95. Shapley additive explanations (SHAP), partial dependence plots (PDP) and individual conditional expectation (ICE) were used for feature importance analysis, aiding in the interpretation of the ‘black box’ nature of the ML models. The results demonstrate that the normal compressive stress at the interface is the most influential factor affecting interfacial shear strength. Finally, a physically meaningful predictive equation for the interface shear strength of UHPC–NC composite structures was proposed based on the XGBoost model combined with curve fitting. This equation enhances the prediction accuracy of interface shear strength for UHPC–NC structures and offers deeper insights into the model’s decision making process.

Machine learning-based predictive models for perioperative major adverse cardiovascular events in patients with stable coronary artery disease undergoing noncardiac surgery.

P1923Deep and machine learning models to improve risk prediction of cardiovascular disease using data extraction from electronic health records