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

This study aimed to construct machine learning (ML) models for predicting prolonged length of stay (pLOS) in intensive care units (ICU) among general ICU patients. A multicenter database called eICU (Collaborative Research Database) was used for model derivation and internal validation, and the Medical Information Mart for Intensive Care (MIMIC) III database was used for external validation. We used four different ML methods (random forest, support vector machine, deep learning, and gradient boosting decision tree (GBDT)) to develop prediction models. The prediction performance of the four models were compared with the customized simplified acute physiology score (SAPS) II. The area under the receiver operation characteristic curve (AUROC), area under the precision-recall curve (AUPRC), estimated calibration index (ECI), and Brier score were used to measure performance. In internal validation, the GBDT model achieved the best overall performance (Brier score, 0.164), discrimination (AUROC, 0.742; AUPRC, 0.537), and calibration (ECI, 8.224). In external validation, the GBDT model also achieved the best overall performance (Brier score, 0.166), discrimination (AUROC, 0.747; AUPRC, 0.536), and calibration (ECI, 8.294). External validation showed that the calibration curve of the GBDT model was an optimal fit, and four ML models outperformed the customized SAPS II model. The GBDT-based pLOS-ICU prediction model had the best prediction performance among the five models on both internal and external datasets. Furthermore, it has the potential to assist ICU physicians to identify patients with pLOS-ICU risk and provide appropriate clinical interventions to improve patient outcomes.

Insulin plays a central role in the regulation of energy and glucose homeostasis, and insulin resistance (IR) is widely considered as the "common soil" of a cluster of cardiometabolic disorders. Assessment of insulin sensitivity is very important in preventing and treating IR-related disease. This study aims to develop and validate machine learning (ML)-augmented algorithms for insulin sensitivity assessment in the community and primary care settings. We analyzed the data of 9358 participants over 40 years old who participated in the population-based cohort of the Hubei center of the REACTION study (Risk Evaluation of Cancers in Chinese Diabetic Individuals). Three non-ensemble algorithms and four ensemble algorithms were used to develop the models with 70 non-laboratory variables for the community and 87 (70 non-laboratory and 17 laboratory) variables for the primary care settings to screen the classifier of the state-of-the-art. The models with the best performance were further streamlined using top-ranked 5, 8, 10, 13, 15, and 20 features. Performances of these ML models were evaluated using the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPR), and the Brier score. The Shapley additive explanation (SHAP) analysis was employed to evaluate the importance of features and interpret the models. The LightGBM models developed for the community (AUROC 0.794, AUPR 0.575, Brier score 0.145) and primary care settings (AUROC 0.867, AUPR 0.705, Brier score 0.119) achieved higher performance than the models constructed by the other six algorithms. The streamlined LightGBM models for the community (AUROC 0.791, AUPR 0.563, Brier score 0.146) and primary care settings (AUROC 0.863, AUPR 0.692, Brier score 0.124) using the 20 top-ranked variables also showed excellent performance. SHAP analysis indicated that the top-ranked features included fasting plasma glucose (FPG), waist circumference (WC), body mass index (BMI), triglycerides (TG), gender, waist-to-height ratio (WHtR), the number of daughters born, resting pulse rate (RPR), etc. The ML models using the LightGBM algorithm are efficient to predict insulin sensitivity in the community and primary care settings accurately and might potentially become an efficient and practical tool for insulin sensitivity assessment in these settings.

Artificial intelligence-enhanced ECG analysis shows promise to detect ventricular dysfunction and remodeling in adult populations. However, its application to pediatric populations remains underexplored. A convolutional neural network was trained on paired ECG-echocardiograms (≤2 days apart) from patients ≤18 years of age without major congenital heart disease to detect human expert-classified greater than mild left ventricular (LV) dysfunction, hypertrophy, and dilation (individually and as a composite outcome). Model performance was evaluated on single ECG-echocardiogram pairs per patient at Boston Children's Hospital and externally at Mount Sinai Hospital using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). The training cohort comprised 92 377 ECG-echocardiogram pairs (46 261 patients; median age, 8.2 years). Test groups included internal testing (12 631 patients; median age, 8.8 years; 4.6% composite outcomes), emergency department (2830 patients; median age, 7.7 years; 10.0% composite outcomes), and external validation (5088 patients; median age, 4.3 years; 6.1% composite outcomes) cohorts. Model performance was similar on internal test and emergency department cohorts, with model predictions of LV hypertrophy outperforming the pediatric cardiologist expert benchmark. Adding age and sex to the model added no benefit to model performance. When using quantitative outcome cutoffs, model performance was similar between internal testing (composite outcome: AUROC, 0.88, AUPRC, 0.43; LV dysfunction: AUROC, 0.92, AUPRC, 0.23; LV hypertrophy: AUROC, 0.88, AUPRC, 0.28; LV dilation: AUROC, 0.91, AUPRC, 0.47) and external validation (composite outcome: AUROC, 0.86, AUPRC, 0.39; LV dysfunction: AUROC, 0.94, AUPRC, 0.32; LV hypertrophy: AUROC, 0.84, AUPRC, 0.25; LV dilation: AUROC, 0.87, AUPRC, 0.33), with composite outcome negative predictive values of 99.0% and 99.2%, respectively. Saliency mapping highlighted ECG components that influenced model predictions (precordial QRS complexes for all outcomes; T waves for LV dysfunction). High-risk ECG features include lateral T-wave inversion (LV dysfunction), deep S waves in V1 and V2 and tall R waves in V6 (LV hypertrophy), and tall R waves in V4 through V6 (LV dilation). This externally validated algorithm shows promise to inexpensively screen for LV dysfunction and remodeling in children, which may facilitate improved access to care by democratizing the expertise of pediatric cardiologists.

BackgroundTuberculosis is a chronic infectious disease caused by mycobacterium tuberculosis (MTB) and is the ninth leading cause of death worldwide. It is still difficult to distinguish active TB from latent TB,but it is very important for individualized management and treatment to distinguish whether patients are active or latent tuberculosis infection.MethodsA total of 220 subjects, including active TB patients (ATB, n = 97) and latent TB patients (LTB, n = 113), were recruited in this study .46 features about blood routine indicators and the VCS parameters (volume, conductivity, light scatter) of neutrophils(NE), monocytes(MO), and lymphocytes(LY) were collected and was constructed classification model by four machine learning algorithms(logistic regression(LR), random forest(RF), support vector machine(SVM) and k-nearest neighbor(KNN)). And the area under the precision-recall curve (AUPRC) and the area under the receiver operating characteristic curve (AUROC) to estimate of the model’s predictive performance for dentifying active and latent tuberculosis infection.ResultsAfter verification,among the four classifications, LR and RF had the best performance (AUROC = 1, AUPRC = 1), followed by SVM (AUROC = 0.967, AUPRC = 0.971), KNN (AUROC = 0.943, AUPRC = 0.959) in the training set. And LR had the best performance (AUROC = 0.977, AUPRC = 0.957), followed by SVM (AUROC = 0.962, AUPRC = 0.949), RF (AUROC = 0.903, AUPRC = 0.922),KNN(AUROC = 0.883, AUPRC = 0.901) in the testing set.ConclusionsThe machine learning algorithm classifier based on leukocyte VCS parameters is of great value in identifying active and latent tuberculosis infection.

Establishment of noninvasive diabetes risk prediction model based on tongue features and machine learning techniques

Surgical site infections (SSIs), among the most frequent healthcare-associated infections, require surveillance, but traditional methods are labour-intensive. We developed machine learning (ML) and rule-based models for the semi-automated detection of deep and organ/space SSIs using data from a prospective cohort of 3931 surgical patients. We assessed sensitivity and workload reduction (proportion of patients not requiring manual review) at a 0.5 decision threshold, and computed area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). The best-performing ML models (Naïve Bayes and dense neural network) achieved sensitivity up to 0.90, AUROC up to 0.968, AUPRC up to 0.248, and workload reduction over 90%. The rule-based model showed higher sensitivity (1.000) but lower AUROC, AUPRC, and workload reduction. Our findings suggest that semi-automated approaches can support efficient and accurate SSI surveillance while reducing manual workload. Further validation in other settings is warranted.

BackgroundMachine learning (ML) achieves better predictions of postoperative mortality than previous prediction tools. Free-text descriptions of the preoperative diagnosis and the planned procedure are available preoperatively. Because reading these descriptions helps anesthesiologists evaluate the risk of the surgery, we hypothesized that deep learning (DL) models with unstructured text could improve postoperative mortality prediction. However, it is challenging to extract meaningful concept embeddings from this unstructured clinical text.ObjectiveThis study aims to develop a fusion DL model containing structured and unstructured features to predict the in-hospital 30-day postoperative mortality before surgery. ML models for predicting postoperative mortality using preoperative data with or without free clinical text were assessed.MethodsWe retrospectively collected preoperative anesthesia assessments, surgical information, and discharge summaries of patients undergoing general and neuraxial anesthesia from electronic health records (EHRs) from 2016 to 2020. We first compared the deep neural network (DNN) with other models using the same input features to demonstrate effectiveness. Then, we combined the DNN model with bidirectional encoder representations from transformers (BERT) to extract information from clinical texts. The effects of adding text information on the model performance were compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Statistical significance was evaluated using P<.05.ResultsThe final cohort contained 121,313 patients who underwent surgeries. A total of 1562 (1.29%) patients died within 30 days of surgery. Our BERT-DNN model achieved the highest AUROC (0.964, 95% CI 0.961-0.967) and AUPRC (0.336, 95% CI 0.276-0.402). The AUROC of the BERT-DNN was significantly higher compared to logistic regression (AUROC=0.952, 95% CI 0.949-0.955) and the American Society of Anesthesiologist Physical Status (ASAPS AUROC=0.892, 95% CI 0.887-0.896) but not significantly higher compared to the DNN (AUROC=0.959, 95% CI 0.956-0.962) and the random forest (AUROC=0.961, 95% CI 0.958-0.964). The AUPRC of the BERT-DNN was significantly higher compared to the DNN (AUPRC=0.319, 95% CI 0.260-0.384), the random forest (AUPRC=0.296, 95% CI 0.239-0.360), logistic regression (AUPRC=0.276, 95% CI 0.220-0.339), and the ASAPS (AUPRC=0.149, 95% CI 0.107-0.203).ConclusionsOur BERT-DNN model has an AUPRC significantly higher compared to previously proposed models using no text and an AUROC significantly higher compared to logistic regression and the ASAPS. This technique helps identify patients with higher risk from the surgical description text in EHRs.

ObjectiveTo evaluate the utility of the preoperative neutrophil percentage-to-albumin ratio (NPAR) for predicting perioperative major adverse cardiovascular events (MACE) in patients with stable coronary artery disease (SCAD) undergoing non-cardiac surgery.MethodsIn this retrospective cohort study, we included all adult SCAD patients who underwent non-cardiac surgery at the Fourth Affiliated Hospital of Zhejiang University School of Medicine between October 2020 and October 2024. The primary endpoint was the occurrence of MACE during the perioperative period, defined as a composite of all-cause mortality, cardiac arrest, myocardial infarction, heart failure, or stroke occurring intraoperatively or during the postoperative hospital stay. We used multivariable logistic regression to assess the independent association between NPAR and MACE risk. To explore potential nonlinearity, we fitted smooth curves and performed threshold-effect analysis. Mediation analysis quantified the proportion of the NPAR–MACE relationship explained by estimated glomerular filtration rate (eGFR). Incremental predictive value was evaluated by comparing the area under the receiver operating characteristic curve (AUC), net reclassification improvement (NRI), and integrated discrimination improvement (IDI) before and after adding NPAR to established risk models. Feature selection was conducted using the Boruta algorithm, and predictive performance was further validated with an XGBoost model interpreted via Shapley Additive Explanations (SHAP).ResultsOf 1,771 patients, 90 (5.1%) experienced MACE. The MACE subgroup had a higher mean NPAR than those without events (19.4 ± 5.3 vs. 15.9 ± 3.5; P < 0.001). Each 1-unit increase in NPAR was associated with a 20% higher risk of MACE (OR 1.20; 95% CI 1.10–1.30). A J-shaped relationship emerged, with an inflection point at NPAR 13.7 (P_threshold = 0.005). eGFR mediated 8.4% of the NPAR–MACE association. NPAR alone yielded an AUC of 0.721. Adding NPAR to the Revised Cardiac Risk Index raised the AUC from 0.679–0.755 (NRI 0.599; IDI 0.035; all P < 0.01). The XGBoost model achieved an AUC of 0.773, and SHAP analysis identified NPAR as the most influential predictor.ConclusionsPreoperative NPAR is an independent, readily available predictor of perioperative MACE in SCAD patients. Incorporation of NPAR into existing risk models significantly enhances predictive accuracy and may inform targeted perioperative management strategies.

Evidence-based precision medicine strategies do not currently exist to guide the choice of biologics in the treatment of psoriasis. As a result, a costly and arduous trial-and-error approach is often adopted. Artificial intelligence has the potential to improve personalization through the prediction of treatment outcomes using real-world data, such as that within the British Association of Dermatologists Biologics and Immunomodulators Register (BADBIR). We aimed to develop an explainable machine learning (ML) model to predict biologic drug discontinuation in a biologic-naive psoriasis cohort using BADBIR data. BADBIR data (2007–2024) were engineered to enable readability. Adult biologic-naive patients across all biologic cohorts with &amp;gt; 6 months of follow-up data were included. Recruitment centres representing 10% of the overall cohort were randomly separated for external validation (model testing). The residual cohort was then randomly split for model training (80%) and internal validation (20%, for hyperparameter tuning). Random forest modelling was applied for imputation of missing data. Only clinical data at baseline prior to biologic initiation were used for model training to enhance future clinical utilization. The performance of several ML (XG-Boost, AdaBoost, random forest) and deep learning (simple and recurrent neural networks) algorithms was evaluated. External validation was performed with a cross-validation leave-group-out approach of individual recruitment centres. SHAP (SHapley Additive exPlanations) and permutation feature importance values were generated to understand model predictions. In total, 10 806 patients were included, in the cohorts for training (n = 7722), internal validation (n = 1930) and external validation (for final model testing: nine centres, n = 1154). Most patients (n = 7290, 67%) discontinued initial biologic therapy within their follow-up duration (median 6.6 years). Within the discontinuation cohort, adalimumab (originator and biosimilars, 57%) was most prescribed. Higher proportions of female patients (43% vs. 37%) and patients with psoriatic arthritis (21% vs. 17%) and scalp psoriasis (59% vs. 51%) were noted in the discontinuation vs. the continuation cohort, respectively. AdaBoost, an ensemble ML model, outperformed other evaluated models with regards to area under the receiver operating characteristic curve (AUROC). Model testing predicted discontinuation of biologic therapy with (mean, 95% confidence interval) precision 0.85 (0.83–0.88), recall 0.80 (0.78–0.83), F1 score 0.82, AUROC 0.76 (0.71–0.78) and area under the precision recall curve (AUPRC) 0.83 (0.81–0.86). Performance metrics following testing with cross-validation [mean (SD)] were precision 0.79 (0.09), recall 0.69 (0.2), F1 score 0.74 (0.16), AUROC 0.71 (0.06) and AUPRC 0.75 (0.11). The features contributing most significantly to model performance were initial biologic drug, baseline Psoriasis Area and Severity Index, patient age, recruitment centre and baseline white cell count. In conclusion, AdaBoost represents an explainable, ML model with potential clinical utility to predict treatment outcomes of patients with psoriasis using real-world registry data. Future work will investigate discontinuation risk across a range of individual biologic therapies.

Evidence-based precision medicine strategies do not currently exist to guide the choice of biologics in the treatment of psoriasis. As a result, a costly and arduous trial-and-error approach is often adopted. Artificial intelligence has the potential to improve personalization through the prediction of treatment outcomes using real-world data, such as that within the British Association of Dermatologists Biologics and Immunomodulators Register (BADBIR). We aimed to develop an explainable machine learning (ML) model to predict biologic drug discontinuation in a biologic-naive psoriasis cohort using BADBIR data. BADBIR data (2007–2024) were engineered to enable readability. Adult biologic-naive patients across all biologic cohorts with &amp;gt; 6 months of follow-up data were included. Recruitment centres representing 10% of the overall cohort were randomly separated for external validation (model testing). The residual cohort was then randomly split for model training (80%) and internal validation (20%, for hyperparameter tuning). Random forest modelling was applied for imputation of missing data. Only clinical data at baseline prior to biologic initiation were used for model training to enhance future clinical utilization. The performance of several ML (XGBoost, AdaBoost, random forest) and deep learning (simple and recurrent neural networks) algorithms was evaluated. External validation was performed with a cross-validation leave-group-out approach of individual recruitment centres. SHAP (SHapley Additive exPlanations) and permutation feature importance values were generated to understand model predictions. In total, 10 806 patients were included, in the cohorts for training (n = 7722), internal validation (n = 1930) and external validation (for final model testing: nine centres, n = 1154). Most patients (n = 7290, 67%) discontinued initial biologic therapy within their follow-up duration (median 6.6 years). Within the discontinuation cohort, adalimumab (originator and biosimilars, 57%) was most prescribed. Higher proportions of female patients (43% vs. 37%) and patients with psoriatic arthritis (21% vs. 17%) and scalp psoriasis (59% vs. 51%) were noted in the discontinuation vs. the continuation cohort, respectively. AdaBoost, an ensemble ML model, outperformed other evaluated models with regards to area under the receiver operating characteristic curve (AUROC). Model testing predicted discontinuation of biologic therapy with (mean, 95% CI) precision 0.85 (0.83–0.88), recall 0.80 (0.78–0.83), F1 score 0.82, AUROC 0.76 (0.71–0.78) and area under the precision recall curve (AUPRC) 0.83 (0.81–0.86). Performance metrics following testing with cross-validation [mean (SD)] were precision 0.79 (0.09), recall 0.69 (0.2), F1 score 0.74 (0.16), AUROC 0.71 (0.06) and AUPRC 0.75 (0.11). The features contributing most significantly to model performance were initial biologic drug, baseline Psoriasis Area and Severity Index, patient age, recruitment centre and baseline white cell count. In conclusion, AdaBoost represents an explainable, ML model with potential clinical utility to predict treatment outcomes of patients with psoriasis using real-world registry data. Future work will investigate discontinuation risk across a range of individual biologic therapies.

Traumatic brain injury (TBI) is a significant healthcare concern in several countries, accounting for a major burden of morbidity, mortality, disability, and socioeconomic losses. Although conventional prognostic models for patients with TBI have been validated, their performance has been limited. Therefore, we aimed to construct machine learning (ML) models to predict the clinical outcomes in adult patients with isolated TBI in Asian countries. The Pan-Asian Trauma Outcome Study registry was used in this study, and the data were prospectively collected from January 1, 2015, to December 31, 2020. Among a total of 6540 patients (≥ 15 years) with isolated moderate and severe TBI, 3276 (50.1%) patients were randomly included with stratification by outcomes and subgrouping variables for model evaluation, and 3264 (49.9%) patients were included for model training and validation. Logistic regression was considered as a baseline, and ML models were constructed and evaluated using the area under the precision-recall curve (AUPRC) as the primary outcome metric, area under the receiver operating characteristic curve (AUROC), and precision at fixed levels of recall. The contribution of the variables to the model prediction was measured using the SHapley Additive exPlanations (SHAP) method. The ML models outperformed logistic regression in predicting the in-hospital mortality. Among the tested models, the gradient-boosted decision tree showed the best performance (AUPRC, 0.746 [0.700-0.789]; AUROC, 0.940 [0.929-0.952]). The most powerful contributors to model prediction were the Glasgow Coma Scale, O2 saturation, transfusion, systolic and diastolic blood pressure, body temperature, and age. Our study suggests that ML techniques might perform better than conventional multi-variate models in predicting the outcomes among adult patients with isolated moderate and severe TBI.

Development and validation of a novel score for predicting perioperative major adverse cardiovascular events in patients with stable coronary artery disease undergoing noncardiac surgery

Hospital-onset bacteremia and fungemia (HOB) are common and potentially preventable complications of hospital care. To assess whether nonindividual patient features, which summarize interactions with other patients and health care workers (HCWs), can contribute to predictive and causal machine learning models for HOB. This prognostic study included adult patients admitted to Barnes-Jewish Hospital, an academic hospital in St Louis, Missouri, in 2021. Analyses were developed between October 2023 and August 2024 and in April 2025. Individual patient features were extracted from electronic health records and used to engineer nonpatient features, including interactions with HCWs and direct or indirect (consecutive room occupancy) patient contact. HOB was defined as a positive blood culture after the third day of hospitalization. Patients who were hospitalized for more than 3 days were considered at risk for the outcome. We developed 3 gradient boosting models: 2 predictive (with patient features only and with both patient and nonpatient features to predict the occurrence of HOB) and 1 causal to test the association of nonpatient features and HOB. Predictive performance is reported using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC), and the results of the causal model are reported as difference in average effects. Sensitivity analyses separated intensive care unit-onset and ward-onset HOB and included a methicillin-resistant Staphylococcus aureus-specific model to adjust for colonization pressure. Among the 52 442 patients, 34 855 (66.5%) had admissions longer than 72 hours and were included for analysis; of these, 556 (1.6%) developed HOB. The median age for the included patients was 60 (IQR, 44-70) years, 50.5% were female, and obesity was the most frequent comorbidity (25.0%). Nonpatient features, such as a prior occupant of the same room receiving antipseudomonal beta-lactams and the mean number of HCWs per day for the 7 days preceding HOB, improved the model's performance (AUROC, 0.88 [95% CI, 0.88-0.89]; AUPRC, 0.20 [95% CI, 0.20-0.22]) compared with the patient-only model (AUROC, 0.85 [95% CI, 0.85-0.86]; AUPRC, 0.13 [95% CI, 0.12-0.14]) (P < .001). These 2 features were also associated with a higher likelihood of HOB in the causal gradient boosting model. These findings suggest that nonindividual patient features may contribute to a comprehensive analysis of HOB when integrated with individual patient features in a machine learning model.

Background The number of non-cardiac surgeries performed worldwide has been steadily increasing, presenting a challenge for clinicians to accurately identify patients at high risk of complications and to allocate the appropriate level of perioperative care. Accurate prediction of postoperative mortality is crucial not only for successful patient care, but also for information-based shared decision-making with patients and efficient allocation of medical resources. Purpose In this study, we aimed to develop a novel predictive model using machine learning methods applied to electronic health record data. Our objective is to identify the risk factors most likely to lead to 30-day major adverse cardiac and cerebrovascular events after non-cardiac surgery Methods We conducted a retrospective analysis of data from a single tertiary care institution that included patients aged 65 years or over who underwent non-cardiac surgery from May 2003 and December 2020. The Observational Medical Outcomes Partnership (OMOP) common data model (CDM) data was used to build predictive models, which allowed for the utilization of demographic data, as well as preoperative characteristics such as diagnosis, lab results, vital signs, medications, and information on operations and procedures from the electronic health records (EHRs) in a standardized way. We employed machine learning models, which were developed and validated using the OHDSI Patient-Level-Prediction framework. Results We included a total of 47,915 patients to train (75%) and test (25%) our predictive models. To compare prediction performances, we applied gradient boosting machine (GBM), logistic regression (LR), random forest (RF), AdaBoost (AB), and decision tree (DT). Our results for a test data (Fig 1.) showed that the GBM model had the best performance in terms of the area under the receiver operating characteristic curve (AUROC) (0.903) and the area under the precision-recall curve (AUPRC) (0.395). Conclusions Our study demonstrates that applying machine learning algorithms to electronic health record data can effectively identify patients at high risk of major adverse cardiac and cerebrovascular events following non-cardiac surgery. This algorithm has the potential to support clinicians in effectively identifying patients at high risk and provide appropriate perioperative care. Further work is needed to validate and refine the proposed model to ensure its external validity and broader applicability in clinical practice.We plan to validate the proposed model externally by testing it on a cohort of approximately 280,000 patients from other tertiary care institution, and present the results at the 2023 ESC Congress.

BackgroundIn the pediatric intensive care unit (PICU), quantifying illness severity can be guided by risk models to enable timely identification and appropriate intervention. Logistic regression models, including the pediatric index of mortality 2 (PIM-2) and pediatric risk of mortality III (PRISM-III), produce a mortality risk score using data that are routinely available at PICU admission. Artificial neural networks (ANNs) outperform regression models in some medical fields.ObjectiveIn light of this potential, we aim to examine ANN performance, compared to that of logistic regression, for mortality risk estimation in the PICU.MethodsThe analyzed data set included patients from North American PICUs whose discharge diagnostic codes indicated evidence of infection and included the data used for the PIM-2 and PRISM-III calculations and their corresponding scores. We stratified the data set into training and test sets, with approximately equal mortality rates, in an effort to replicate real-world data. Data preprocessing included imputing missing data through simple substitution and normalizing data into binary variables using PRISM-III thresholds. A 2-layer ANN model was built to predict pediatric mortality, along with a simple logistic regression model for comparison. Both models used the same features required by PIM-2 and PRISM-III. Alternative ANN models using single-layer or unnormalized data were also evaluated. Model performance was compared using the area under the receiver operating characteristic curve (AUROC) and the area under the precision recall curve (AUPRC) and their empirical 95% CIs.ResultsData from 102,945 patients (including 4068 deaths) were included in the analysis. The highest performing ANN (AUROC 0.871, 95% CI 0.862-0.880; AUPRC 0.372, 95% CI 0.345-0.396) that used normalized data performed better than PIM-2 (AUROC 0.805, 95% CI 0.801-0.816; AUPRC 0.234, 95% CI 0.213-0.255) and PRISM-III (AUROC 0.844, 95% CI 0.841-0.855; AUPRC 0.348, 95% CI 0.322-0.367). The performance of this ANN was also significantly better than that of the logistic regression model (AUROC 0.862, 95% CI 0.852-0.872; AUPRC 0.329, 95% CI 0.304-0.351). The performance of the ANN that used unnormalized data (AUROC 0.865, 95% CI 0.856-0.874) was slightly inferior to our highest performing ANN; the single-layer ANN architecture performed poorly and was not investigated further.ConclusionsA simple ANN model performed slightly better than the benchmark PIM-2 and PRISM-III scores and a traditional logistic regression model trained on the same data set. The small performance gains achieved by this two-layer ANN model may not offer clinically significant improvement; however, further research with other or more sophisticated model designs and better imputation of missing data may be warranted.