Development and Validation of a Novel Machine Learning-Based Algorithm to Predict Incident Atrial Fibrillation - A Multicohort Analysis.

Background BRAF V600E mutation detection enhances diagnostic accuracy in distinguishing benign from malignant thyroid nodules. This study aims to develop and validate a predictive model for the BRAF V600E mutation in C-TIRADS 3 or higher nodules. Methods A retrospective study was conducted involving 324 patients with C-TIRADS 3 or higher thyroid nodules. Based on BRAF V600E testing from ultrasound-guided fine needle aspiration biopsy (FNAB), patients were divided into wild-type (n=263) and mutation(n=61) groups. Predictive features were independently selected from ultrasonography (US), contrast-enhanced CT (CECT), and combined imaging using Least Absolute Shrinkage and Selection Operator (LASSO) regression. Multivariate logistic regression analysis was employed to identify independent risk factors and then develop three predictive models. Model performance was evaluated through calibration curves, receiver operating characteristic (ROC) analysis, decision curve analysis (DCA), and Brier scores, respectively. The optimal model was subsequently converted into a visualized nomogram to facilitate clinical implementation. Results Ultrasonographic microcalcifications were the strongest independent predictor of BRAF V600E mutation (OR = 9.63, 95% CI: 3.62–25.63, P < 0.001). Higher C-TIRADS grades, irregular morphology on US, and blurred borders or capsule interruption on CECT were also significant independent risk factors. Notably, smaller nodule size on US correlated with higher mutation risk (OR = 0.93, 95% CI: 0.88–0.98, p=0.012). The multimodal model combining US and CECT (AUC = 0.937) outperformed individual US (AUC = 0.915) and CECT (AUC = 0.784) models. Conclusion The nomogram integrating US and CECT features shows strong predictive performance and clinical utility for identifying BRAF V600E mutations in C-TIRADS 3 or higher thyroid nodules.

Machine learning-driven prediction of readmission risk in heart failure patients with diabetes: synergistic assessment of inflammatory and metabolic biomarkers.

External evaluation of the SORG machine learning algorithm predicting 90-day and 1-year mortality in a Midwest cohort of patients with spinal metastasis.

Accuracy of the Surgical Risk Preoperative Assessment System (SURPAS) in a broad, elderly (age ≥ 65) patient population.

Machine learning outperforms deep learning in adhesive capsulitis diagnosis: a clinical-radiomics model bridging PD-T2 MRI and multimodal data fusion.

Machine learning-based predictive model for postpartum post-traumatic stress disorder: A prospective cohort study.

This research provides a reproducible comparative analysis of the performance of six independent machine learning classifiers in predicting in-hospital mortality among ICU patients from the PhysioNet/Challenge-2012 dataset. The term 'single' in the title of the former evoked the expectation that the current work would deal with various models. The paper discusses the single-model classifiers SVM, LR, RF, XGB, MLPClassifier, and a Keras-based Neural Network, comparing their performance, calibration, and interpretability against a strict set of pipelines. Finally, the most remarkable contributions include a workflow diagram that includes information on all processes; the hyperparameter search space, early-stopping hyperparameter, and random seeds; preprocessing and imputation experiments comparing the mean, median, KNN and Iterative imputation; feature selection with the help of Random-Forest RFE, using a certain stopping rule that disregards the frequency of stability, triangulation of predictor importance by SHAP and permutation importance; current confidence intervals (CIs) and significance tests; and subgroup analyses based on age, sex, and severity. Findings indicate that XGBoost has high discrimination and calibration statistics compared to the other classifiers; statistically significant ROC-AUC and Brier score improvements are obtained in favor of this algorithm. Every performance statistic is followed by 95% CIs; calibration curves, learning curves, and data regarding runtime assessment are provided.

Defining diastolic dysfunction post-Fontan: Threshold, risk factors, and associations with outcomes.

Generative artificial intelligence (AI) tools are increasingly utilized in higher education, particularly in college admissions. This study assesses OpenAI’s ChatGPT in predicting admission outcomes, comparing it to CollegeVine’s “chancing” engine. Using 64 Common App submissions from three volunteers applying to various colleges, we asked ChatGPT to predict admission probabilities and decisions based on their profiles. ChatGPT achieved 84.38% prediction accuracy, matching CollegeVine, and had a slightly better Brier Score (0.1164 vs. 0.1186), indicating better probability calibration. These results suggest that generative AI can perform similarly to dedicated admissions prediction models. However, findings are limited to a small sample size, highlighting the need for further research. We discuss the responsible integration of AI in college advising, stressing transparency, fairness, and ethical considerations.

There are no prediction models of stent outcomes for leaks after metabolic and bariatric surgery (MBS). The current study developed an artificial intelligence-based model to predict post-MBS stent failure. Prospectively maintained database of patients with post-MBS leaks was used for model development (Center I, N = 250); external validation employed patients from another hospital (Center II, N = 150). Outcome definition was failure of the first (primary/initial) stent implantation to resolve the leak, i.e., lack of primary closure. Ranking of variables was performed, 11 machine learning algorithms were tested, the best model was selected, and a stent failure point-based risk scoring system was derived, with further external validation, calibration, and decision curve analysis. The development cohort (training sample, Center I) had 27.6% failed stents/72.4% successes; the external validation cohort (Center II) had 30% failures/70% successes. The Lasso logistic regression model exhibited the best performance. Eight variables contributed to the model's predictive performance (obstructive sleep apnea, hypertension, diabetes, hepatomegaly, hyperlipidemia, body mass index, Niti-S18 stent, gastrojejunal anastomosis leak), and nine others had varying contributions (revisional surgery, Niti-S23 stent, time to stent implantation, leak size > 1cm, age, Roux-en-Y gastric bypass surgery, esophagogastric junction leak, Hanaro 21 stent, male sex). The clinical point-based stent failure risk system showed that scores ≤ 7 had very low failure risk (<1%), scores 8-47 = low risk (1-5%), 48-77 = moderate risk (5.1-15%), 78-117 = high risk (15.1-50%), and scores ≥198 were associated with extremely high failure risk (>96%). The model's external validation demonstrated excellent discriminatory power, distinguishing between patients with/without the outcome with 0.85 area under the ROC curve (95% CI: 0.76-0.93), 80% sensitivity (95% CI: 65.4-90.4%), 82.9% specificity (95% CI: 74.3-89.5%), and 66.7% positive predictive value (95% CI: 52.4-79.0%). The negative predictive value was 90.6% (95% CI: 82.9-95.6%) indicating that the model was particularly effective at identifying patients unlikely to fail. Area under the precision-recall curve was 0.81 (95% CI: 0.70-0.89) indicating strong performance in identifying true positives while minimizing false positives. Calibration was acceptable (Brier score = 0.15). Decision curve analysis demonstrated higher net benefit when used in clinical decision-making across a broad range of threshold probabilities (0.10-0.80) compared to treating all patients or treating none. A machine learning model (Alexandria-Bari-Stent) can predict post-MBS stent failure. External validation displayed high accuracy, good sensitivity/specificity, and excellent negative predictive value indicating good discriminative ability. Clinically, the model is more reliable for ruling out stent failure than confirming it, making it especially useful in reassuring low-risk post-MBS leakage patients. Patient's general status, metabolic health, and systemic factors appeared to play a more critical role than previously recognized, complementary to, not in conflict with, established technical and local factors that influence successful stent outcomes for leak management. This prompts the need for a more holistic view of leak patients who are candidates for stenting. Prospective multicenter trials are needed to confirm the performance of the Alexandria‑Bari‑Stent model and the role of metabolic stabilization and medically optimizing the patient for better outcomes.

Traditional cardiovascular disease (CVD) risk scores, such as the Framingham Risk Score (FRS), rely on laboratory and clinical measurements often unavailable in large-scale remote health surveys. Validated tools using self-reported data could expand the feasibility of risk stratification in resource-limited settings. To assess the discriminatory capacity and agreement of adapted FRS models using only self-reported data, compared to the original FRS based on clinical and laboratory inputs. This cross-sectional study used data from 6,966 Brazilian adults aged 30-74 years from the 2013 National Health Survey. Three FRS models without laboratory inputs were evaluated: FRS-BMI (with measured BMI), FRS-BMI-HwT (using self-reported BMI and hypertension diagnosis with treatment data), and FRS-BMI-HnT (using self-reported BMI and hypertension diagnosis without treatment data). The original FRS served as the reference method. Concordance was assessed for CVD-risk ≥ 5%, ≥ 10%, and ≥ 20% using Receiver Operating Characteristic (ROC) curves and a suite of statistical measures for reliability and agreement [Intraclass Correlation Coefficient (ICC), Bland-Altman, Cohen’s Kappa]. Optimal cutoffs were identified by maximizing Youden’s index, with stability assessed through bootstrap validation. Additional continuous accuracy metrics were computed, including Brier Score, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Integrated Discrimination Improvement (IDI). Prevalence of CVD-risk ≥ 5%, ≥ 10%, and ≥ 20% by FRS-Original was 55.4%, 35.9%, and 20.4%, respectively. FRS-BMI achieved the highest AUCs (0.90, 0.88, 0.83), followed by FRS-BMI-HwT and FRS-BMI-HnT (0.86-0.83 and 0.86-0.78). Concordance was substantial for CVD-risk ≥ 5% and 10% (kappa > 0.70) and moderate for ≥ 20% (kappa > 0.55). All models demonstrated excellent predictive accuracy (Brier Score < 0.01) with minimal IDI values (-0.0014 to 0.0056), indicating nearly identical discrimination between adapted and original models. Bootstrap validation confirmed excellent stability of optimal thresholds (bias < 0.02). The adapted models slightly underestimated risk (mean score difference: -0.36 to -0.61). Regression models showed consistent associations with key risk factors across all versions. Self-reported FRS models demonstrated strong discriminatory capacity and high agreement with the original FRS, supporting their use in telephone and online surveys where laboratory data are unavailable. These pragmatic tools offer reliable alternatives for CVD-risk stratification in remote, low-resource, or large-scale epidemiological research.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-27008-7.

Background Post-operative ileus (POI) is a frequent complication after radical cystectomy (RC). Conventional scores capture only linear relations and have limited accuracy. Interpretable machine learning (ML) may improve early risk stratification. Methods In a single-centre real-world cohort ( n = 1,062, 2013–2023), POI was defined by ≥2 standard clinical–radiological criteria. We extracted pre-operative comorbidities/medications, operative factors (approach, urinary diversion, lymph-node dissection, fluids, blood loss, nasogastric-tube placement) and first-day laboratory indices. After LASSO selection, five ML models were trained/validated on a stratified split; discrimination (AUC), accuracy, precision, recall and Brier score were compared. SHAP delivered global and patient-level explanations. Results POI occurred in 28.9%. The back-propagation neural network performed best (AUC 0.828; accuracy 78.4%; Brier 0.143). Intra-operative nasogastric-tube placement and surgical approach dominated feature attribution, followed by medication history, lymph-node dissection, lymphocyte count and C-reactive protein. SHAP clarified feature effects and enabled interpretable, case-level risk summaries. Conclusion An interpretable ML model based on routinely captured peri-operative variables accurately stratifies RC patients at risk for POI as early as postoperative day 0, outperforming existing nomograms and highlighting modifiable factors. Embedding this tool into electronic-health-record workflows could enable real-time alerts and risk-adapted management. Prospective multicentre validation is warranted.

Long-term survival after a lung transplant remains highly variable, and existing risk stratification tools have limited accuracy, clinical utility, and interpretability. To develop, validate, and assess the clinical utility of an interpretable hybrid machine learning model using United Network for Organ Sharing data to predict time to death or retransplant at 1, 5, and 10 years after a lung transplant. This prognostic study used data from a United Network for Organ Sharing-Organ Procurement and Transplantation Network cohort that underwent lung transplants between October 16, 1987, and March 26, 2025. The study included 51 933 adult patients (aged ≥18 years) undergoing their first lung transplant in the US. The data were temporally split into a development cohort (1987-2014; n = 26 682) and a testing cohort (2015-2025; n = 25 251). The development cohort was divided into a training set (n = 24 014) and validation set (n = 2668) for model selection and hyperparameter tuning. The outcome was the time to death or retransplant. The model was developed using the AutoScore-Survival framework, which uses a random survival forest for variable selection and Cox proportional hazards regression for scoring. Performance was assessed by discrimination (a time-dependent area under the curve [AUC], the Harrell C-index, and integrated AUC [iAUC]), calibration (plots, slope, observed-to-expected event ratio, and Brier score), and clinical utility (decision curve analysis). Among 51 933 recipients (median age, 59 years [5th-95th percentile range, 27-71 years]; 57.6% men), the median follow-up was 8.97 years (95% CI, 8.93-8.99 years), and 31 865 (61.4%) experienced an event. Nine predictors were selected for the final model: length of hospital stay, recipient age, single vs double transplant, posttransplant ventilation support, prior cardiac surgery, creatinine level at transplant, functional status, total bilirubin level, and donor age. In the unseen testing set, the model showed moderate discrimination with an iAUC of 0.61 (95% CI, 0.60-0.63) and a C-index of 0.64 (95% CI, 0.63-0.64). The time-dependent AUC was 0.61 (95% CI, 0.52-0.70) at 1 year, 0.59 (95% CI, 0.53-0.65) at 5 years, and 0.72 (95% CI, 0.55-0.85) at 10 years. The model was well calibrated, and the decision curve analysis demonstrated a consistent net benefit across threshold probabilities. In this large prognostic study, the interpretable hybrid model provided practical, personalized risk stratification for lung transplant outcomes. With moderate discrimination, good calibration, and clear clinical utility, the model supports shared decision-making and is accessible via a web-based calculator.

Policymakers in small open economies need reliable signals of incipient private consumption downturns, yet traditional indicators are revised, noisy, and often arrive too late. This study develops a Romanian-specific early warning system that combines a time-varying parameter VAR with stochastic volatility and exogenous drivers (TVP-SV-VARX) with modern machine learning classifiers. The structural layer extracts regime-dependent anomalies in the macro-financial transmission to household demand, while the learning layer transforms these anomalies into calibrated probabilities of short-term consumption declines. A strictly time-based evaluation design with rolling blocks, purge and embargo periods, and rare-event metrics (precision–recall area under the curve, PR-AUC, and Brier score) underpins the assessment. The best-performing specification, a TVP-filtered random forest, attains a PR-AUC of 0.87, a ROC-AUC of 0.89, a median warning lead of one quarter, and no false positives at the chosen operating point. A sparse logistic calibration model improves probability reliability and supports transparent communication of risk bands. The time-varying anomaly layer is critical: ablation experiments that remove it lead to marked losses in discrimination and recall. For implementation, the paper proposes a three-tier WATCH–AMBER–RED scheme with conservative multi-signal confirmation and coverage gates, designed to balance lead time against the political cost of false alarms. The framework is explicitly predictive rather than causal and is tailored to data-poor environments, offering a practical blueprint for demand-side macroeconomic early warning in Romania and, by extension, other small open economies.

This study aims to develop and validate a machine learning-based mortality risk prediction model for V-A ECMO patients to improve the precision of clinical decision-making. This multicenter retrospective cohort study included 280 patients receiving V-A ECMO from the Second Affiliated Hospital of Guangxi Medical University, Yulin First People’s Hospital, and the MIMIC-IV database. The data from the Second Affiliated Hospital of Guangxi Medical University and the MIMIC-IV database were merged and randomly divided in a 7:3 ratio into a training set and an internal validation set, respectively. The dataset from Yulin First People’s Hospital was reserved as an external validation cohort. The primary study outcome was defined as in-hospital mortality.Feature selection was conducted using Lasso regression, followed by the development of six machine learning models: Logistic Regression, Random Forest (RF), Deep Neural Network (DNN), Support Vector Machine (SVM), LightGBM, and CatBoost. Model performance was assessed using the Area Under the Curve (AUC), accuracy, sensitivity, specificity, and F1 score. Model validation was performed through calibration and decision curve analysis. Feature importance was evaluated using SHAP, and subgroup analysis was conducted to assess the model’s applicability across different clinical scenarios. In internal validation, the Logistic Regression model performed the best, with an AUC of 0.86 (95% CI: 0.77–0.93), accuracy of 0.76, sensitivity of 0.73, specificity of 0.79, and an F1 score of 0.73. It outperformed other models (RF: AUC = 0.79, DNN: AUC = 0.78, SVM: AUC = 0.76, LightGBM: AUC = 0.71, CatBoost: AUC = 0.77). External validation yielded consistent results, with the Logistic Regression model’s AUC at 0.75 (95% CI: 0.56–0.92), accuracy of 0.69, sensitivity of 0.64, specificity of 0.73, and an F1 score of 0.66. Calibration curve analysis revealed that the Logistic Regression model had the lowest Brier score (0.1496), indicating the most reliable predicted probabilities. Decision curve analysis demonstrated that the model provided the highest net benefit across most decision thresholds. SHAP analysis identified lactate, age, and albumin as key predictors of mortality, with lactate and age positively correlated, and albumin negatively correlated. Subgroup analysis revealed better performance in the cardiac arrest group (AUC = 0.81), non-sepsis group (AUC = 0.75), and non-diabetes group (AUC = 0.78). The Logistic Regression-based mortality risk prediction model for V-A ECMO patients demonstrated comparable or even favorable performance to more complex machine learning models, with the advantage of higher interpretability.By explicitly incorporating lactate, age, and albumin as the principal predictors, this model facilitates precise risk stratification and provides practical support for clinical decision-making in ECMO management.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-25423-4.

Development and External Validation of a Nomogram for Predicting Spontaneous Preterm Birth in Singleton Pregnancies with a Short Cervix.

ObjectiveTo explore the application of machine learning methods for screening risk factors for long-term adverse prognosis in neonatal bacterial meningitis, determine the final prediction model, and evaluate its predictive value.MethodsThis study included 139 full-term neonates diagnosed with neonatal bacterial meningitis in the capital institute of pediatrics between January 2019 and December 2023. Based on follow-up outcomes, they were divided into a poor prognosis group (n = 45) and a good prognosis group (n = 94). Thirty-three clinical variables were collected. Feature selection was performed using the Least Absolute Shrinkage and Selection Operator, Boruta, and Recursive Feature Elimination. Seven machine learning models were constructed. Model performance was evaluated using metrics including the area under the receiver operating characteristic curve, accuracy, and sensitivity. The Shapley Additive explanation method was used to interpret the models.ResultsAmong the seven models, The Random Forest model demonstrates the best overall predictive performance, although Logistic Regression achieved the highest discriminative ability (AUC: 0.903), Random Forest was more suitable for clinical application due to its superior accuracy (0.881), better calibration (Brier score: 0.123), and balanced sensitivity (0.887) and specificity (0.878). Shapley Additive explanation interpretability analysis further revealed that the top three important features were cerebrospinal fluid white blood cell count, cerebrospinal fluid protein levels, and seizures.ConclusionMachine learning models, particularly the superior-performing Random Forest, are proven to reliably predict long-term adverse outcomes in NBM patients, aiding in the identification of high-risk individuals. Further validation in broader cohorts is warranted to enhance generalizability and clinical applicability.

BackgroundMachine learning has emerged as a promising tool for survival prediction in various diseases; however, its application and external validation in real-world gastric cancer populations remain limited.MethodsClinical data of patients diagnosed with gastric cancer between 2000 and 2018 were obtained from the SEER database, supplemented with data from two Chinese medical centers (2005–2018). Three feature selection methods and four modeling algorithms—including Cox, RSF, CoxBoost, and Deepsurv_Cox—were employed to construct prediction models for overall survival (OS) and cancer-specific survival (CSS). Model performance was evaluated using the concordance index (C-index), integrated Brier score (IBS), and mean area under the curve (AUC). The two best-performing base models were subsequently integrated into a stacked model and compared against the traditional TNM staging system using decision curve analysis (DCA) and time-dependent ROC curves at 3, 5, and 10 years.ResultsA total of 21,559 patients from the SEER database and 3,805 patients from two Chinese centers were included. In independent testing, the integrated model achieved a C-index/IBS/mean AUC of 0.693/0.158/0.829 for OS and 0.719/0.171/0.819 for CSS. For 3-, 5-, and 10-year survival prediction, the AUCs were 0.705/0.747/0.851 for OS and 0.734/0.779/0.830 for CSS, outperforming the TNM staging system across all metrics. Superior calibration and clinical utility of the integrated model were further confirmed by calibration curves and DCA.ConclusionThe integrated machine learning model outperformed both traditional TNM staging and deep learning approaches, offering improved predictive accuracy for survival outcomes in patients with gastric cancer.

ObjectiveTo develop a CT-based decision tree model integrating clinical and imaging features for the preoperative differentiation of gastric ectopic pancreas (GEPs) and gastrointestinal stromal tumors (GISTs) with a maximum diameter of less than 3 cm. MethodsThis retrospective study included 86 patients with pathologically confirmed GEPs (n = 26) and GISTs (n = 60), all with lesions smaller than 3 cm. Clinical information and CT features were collected. The dataset was divided into training and testing sets. A decision tree classification model was constructed using key variables selected from the training set via univariate analyses and logistic regression. The decision tree's hyperparameters were optimised using five-fold cross-validation. Diagnostic performance was evaluated on an independent test set, including plotting ROC curves to calculate AUC values, sensitivity, and specificity, alongside using calibration curves to assess goodness-of-fit. Furthermore, the SHAP method was employed to provide visual explanations for the final model's predictions.ResultsThe decision tree model identified four key variables: age (clinical factor) and three CT features: ratio of lesion-to-pancreas attenuation in the arterial phase(A2), lesion long-to-short diameter ratio (LD/SD ratio), and intralesional low attenuation (ILA). The model, based on these four features, achieved an AUC of 0.744(95% CI:0.589–0.950), with sensitivity of 76.9% and specificity of 84.6%. Concurrently, calibration analysis substantiated the model's exceptional predictive precision. The Brier score (0.0648) and the Hosmer–Lemeshow test (χ2 = 5.365, df = 8, P = 0.718) both demonstrated a high degree of agreement between the model's predicted probabilities and the actual observed values.ConclusionsThe CT-based decision tree model, integrating four clinical and CT features, provides a reliable and visualized tool for differentiating GEPs from GISTs with a maximum diameter of less than 3 cm, demonstrating strong diagnostic performance.