Development and internal validation of an explainable machine learning model for predicting textbook outcome after free flap reconstruction in oral cancer.

Reproducible bicarbonate thresholds predict critically ill patient mortality in an international personalized survival study.

Explainable 3D VGG-style convolutional neural network for pediatric hydrocephalus detection on computed tomography: A segmentation-free and fully volumetric deep learning framework.

Dual-center external validation of the ARCH score: Predictive accuracy and calibration for hypothermic circulatory arrest aortic arch surgery.

Comparative validation of risk scores for in-hospital complications following chronic total occlusion percutaneous coronary intervention.

Following therapeutic advancements, recompensation has gained increasing recognition in patients on waitlist for liver transplantation (LT). Identifying key predictors of waitlist removal because of improvement can enhance prognostication and resource allocation. We hence examined predictors of improvement-related waitlist removal using a machine learning-based approach with data from the United Network for Organ Sharing database. In this retrospective cohort study, adult LT waitlist candidates from 2000 to 2025 in the United Network for Organ Sharing registry were included. A random survival forest model was applied to examine key predictors associated with improvement-related waitlist removal, while accounting for death and LT as competing risks. Variable importance (VIMP) measure and minimal depth were used to guide variable selection. Model performance was evaluated using the concordance index, Brier scores, and time-dependent area under the curve. The cohort included 127 978 individuals listed for LT. Eight thousand four hundred ninety-three (6.6%) were delisted because of clinical improvement. The random survival forest model demonstrated strong performance and discriminatory ability overall at 1, 5, and 15 y (concordance index was 0.777, 0.771, and 0.781; time-dependent area under the curve was 0.78, 0.78, and 0.80). Brier scores were reduced relative to the reference. Strong predictors of recovery highlighted in both VIMP and minimal depth-based assessments of VIMP included diagnosis, age, and serum albumin. Identified variables could inform the development of robust predictive models to guide individualized decision-making for LT. With further validation and integration into clinical workflows, such models could enhance prognostication of patient trajectory on the LT waitlist and facilitate appropriate resource allocation.

Predictive accuracy of a perioperative hemodynamic indices-based prediction model for moderate-to-severe acute kidney injury after orthotopic heart transplantation.

Dynamic prediction of early discharge after major pancreatic surgery: Derivation and cross-validation of an automatable electronic medical record-based model.

This study aims to investigate the association between triglyceride-glucose (TyG) index and risk of CVD using explainable survival analysis method based on 2011 to 2020 China Health and Retirement Longitudinal Study (CHARLS) data. We enrolled 7,721 participants in a prospective cohort. A Lasso Cox model was implemented to select covariates for constructing the multivariate Cox regression model. Restricted cubic splines (RCS) analysis was performed to explore dose-response relationship. In addition, we utilized explainable machine learning methods to analyse the Cox model. During 9-year follow-up, 1,895 (24.5%) participants developed CVD. Nine variables including TyG index, age, BMI, WC, SBP, history of hypertension, liver disease and kidney disease and antihypertensive medication were retained to establish the Cox model. HRs (95% CIs) for CVD were 1.21 (1.06-1.39), 1.26 (1.10-1.44), and 1.22 (1.06-1.40) for Q2 to Q4 groups compared with Q1. Result of RCS showed that the Q1 group had lowest risk. Time-dependent feature importance analysis showed that age and history of hypertension were two most important risk factors based on Brier score and C/D AUC. All variables in the coxph model gain importance over time. Routine measurement of the TyG index may aid in the early detection and risk stratification of CVD in the aging population.

Interpretable and non-linear machine learning models for quantitative ICG perfusion assessment in colorectal cancer anastomosis: A feasibility study.

The American Heart Association (AHA) recently proposed the concept of Cardiovascular-Kidney-Metabolic (CKM) syndrome, highlighting the strong pathophysiological links among metabolic disorders, chronic kidney disease, and cardiovascular disease (CVD). Insulin resistance (IR) is regarded as a central mechanism underlying CKM syndrome. However, studies comparing the predictive value of different IR surrogate markers for incident CVD are still limited. This study aimed to evaluate the associations between multiple IR surrogate markers and incident cardiovascular disease and to further assess their predictive performance using machine learning approaches. Using data from the China Health and Retirement Longitudinal Study (CHARLS), this prospective cohort study included 5,528 participants. Twelve IR surrogate indices were assessed, including TyG-related indices, TG/HDL-C, METS-IR, CTI, CHG, and eGDR. Incident CVD was defined as self-reported physician-diagnosed heart disease or stroke during follow-up. Cox proportional hazards models were used to estimate associations between standardized IR indices and incident CVD. Restricted cubic splines, weighted quantile sum regression, and quantile g-computation were used to examine dose-response patterns and the relative contribution of correlated IR indices. Predictive performance was evaluated using ROC analysis, calibration, Brier score, decision curve analysis, NRI, IDI, and machine-learning models. During a median follow-up of 7.0 years, 741 participants developed incident CVD. In fully adjusted Cox models, several IR surrogate indices were associated with incident CVD. TyG-related composite indices incorporating adiposity-related information, particularly TyG-WC, TyG-CVAI, TyG-WHtR, and TyG-BMI, showed stronger positive associations with CVD risk, whereas eGDR showed an inverse association. Restricted cubic spline analyses showed significant overall associations for most indices, with nonlinear patterns observed for METS-IR, CTI, and eGDR. Mixture-based analyses suggested relatively larger contributions of CTI, TyG-BMI, and TyG-WC. Among individual indices, eGDR showed the highest discrimination for incident CVD, followed by TyG-CVAI and TyG-WC. Adding selected IR indices, particularly eGDR, to the covariate-based model modestly improved discrimination and reclassification. Among adults with CKM syndrome stages 0-3, several IR surrogate indices were prospectively associated with incident CVD, with stronger and more consistent associations observed for TyG-based indices incorporating adiposity-related measures and for eGDR. These results suggest that the combined assessment of metabolic dysfunction, adiposity, and insulin sensitivity may provide useful information for identifying individuals at higher cardiovascular risk in early-stage CKM syndrome.

The incorporation of side branches in vessel geometry influences wall shear stress (WSS) distribution. However, complete vessel reconstruction is time-consuming, and there is no evidence that its WSS estimations better predict atherosclerotic disease progression compared with the output of the conventional single-vessel reconstruction (SVR). Patients who had baseline and 1-year follow-up intravascular ultrasound imaging (n=40 vessels), and patients with neoatherosclerotic lesions (n=13 vessels) on optical coherence tomography were included. All the studied vessels had at least one side branch with a diameter >1 mm; 3-dimensional complete vessel reconstruction and SVR were performed, and the time-averaged WSS and multidirectional WSS were computed. The performance of both methods in predicting disease progression in intravascular ultrasound and optical coherence tomography models was assessed. The incorporation of side branches in 3-dimensional geometry resulted in lower minimum predominant time-averaged WSS in the intravascular ultrasound (1.09 versus 1.58 Pa, P<0.001) and optical coherence tomography-based reconstructions (0.68 versus 1.33 Pa, P<0.001) and influenced the multidirectional WSS distribution. In native segments, complete vessel reconstruction-derived WSS metrics demonstrated superior predictive performance for disease progression-defined as lumen area reduction and plaque burden increase-compared with SVR, as evidenced by improved out-of-sample accuracy (leave-one-out information criterion: 429 versus 551), discrimination (C statistic: 0.725 versus 0.651), calibration (Brier score: 0.172 versus 0.226), and explained variance (27.8% versus 20.7%). Consistent findings were observed in stented segments, where complete vessel reconstruction-derived WSS metrics more accurately predicted neointimal proliferation than SVR-derived metrics. Incorporating side branches into vessel reconstruction influences WSS distribution and enables more accurate prediction of atherosclerotic disease progression in native and stented segments than SVR.

Perioperative myocardial injury (PMI) is a frequent and often asymptomatic complication after non-cardiac surgery and is associated with increased short- and long-term mortality. Conventional risk scores, such as the Revised Cardiac Risk Index (RCRI), have limited predictive accuracy and are infrequently used in clinical practice. We aimed to develop and temporally validate an interpretable machine learning model using Explainable Boosting Machines (EBMs) to predict PMI from routine pre-operative data. In this retrospective cohort study at a tertiary care centre in Germany, we included 9323 adult patients undergoing 9824 non-cardiac surgical procedures between 2014 and 2023 who received post-operative high-sensitivity cardiac troponin testing as part of routine care. PMI was defined as a post-operative elevation of high-sensitivity cardiac troponin above the upper reference limit. An EBM was trained on structured pre-operative data from 2014 to 2021 and evaluated in a temporally independent test cohort from 2022 to 2023, with performance compared with logistic regression, random forest, XGBoost, and a modified RCRI. Model discrimination, calibration, and Brier scores were assessed. Feature contributions were examined using internal shape functions and SHAP values. PMI occurred in 2804 procedures (28.5%). The EBM achieved the highest predictive performance (AUROC 0.730, 95% CI 0.720-0.740), outperforming all comparators. Calibration was robust across clinically relevant risk ranges. Key predictors included age, leukocyte count, renal function, potassium, and platelet count. The EBM identified high-risk patients more efficiently than the modified RCRI and ESC guideline-based strategies (Number Needed to Evaluate 3.0 vs. 3.5) and reduced troponin assays by 18.2% in the temporally independent cohort. An interpretable machine learning model trained on routine clinical data can accurately predict PMI and outperform existing risk scores. The EBM supports individualized risk stratification and may enhance perioperative decision-making and resource allocation within a guideline-directed testing population. Prospective and external validation is required before clinical implementation.

Sex-specific alterations of niacin flush pathway biomarkers in schizophrenia.

Development and validation of a machine learning model and a web tool for predicting the risk of Crohn's disease diagnosis in Chinese patients.

We evaluated bidirectional long short-term memory models for predicting inpatient mortality using different approaches to processing vital signs data collected during the initial 24h of intensive care unit (ICU) admissions. We compared 3 vital-sign representations: (1) raw data recorded every 5min, (2) preprocessed data averaged hourly, and (3) preprocessed data using biomarker representations that extends a digital oximetry biomarker toolbox of PhysioZoo software, applied to blood pressure, heart rate, temperature, respiratory rate, and SpO2. Across 2 large ICU datasets, HiRID and eICU, models trained on the frequency-normalized representation achieved higher discrimination and lower Brier scores than those trained on raw 5-min and hourly averaged data. The use of biomarker representations of vital signs yielded the largest improvements in discrimination and overall probabilistic performance reflected by lower Brier scores for predicting inpatient mortality by deep learning. Thus, we recommend using a similar approach to vital signs preprocessing for time-series predictive models.

To develop and internally validate a vectorcardiography (VCG)-augmented model for estimating 12-month major adverse cardiovascular events (MACE) in hospitalized patients with chronic heart failure (CHF). We conducted a single-centre retrospective cohort study of adults hospitalized with CHF between 31 May 2023 and 31 May 2024, with data lock on 31 May 2025. The prespecified endpoint was any MACE within 12months after the index hospitalization. ECG-to-VCG transformation was performed using the Kors method. The prespecified primary six-predictor model included LVEDD, NYHA class, frontal, horizontal, and sagittal QRS-T angles, and the QRS-loop reversal/U-turn sign. BNP and LVEF were evaluated in full-model, comparator-model, and incremental-value analyses. Because the prediction target was fixed 12-month risk rather than time-to-event hazard, multivariable logistic regression was used as the primary modelling approach. Internal validation was performed using 1000 bootstrap resamples, with apparent, optimism-corrected, calibration, and shrinkage-adjusted performance reported. Of 201 screened, 160 were included; MACE occurred in 68/160 (42.5%). The six-predictor model showed an apparent AUC of 0.946 (95% CI 0.914-0.978). Bootstrap internal validation yielded an AUC optimism estimate of 0.012 and an optimism-corrected AUC of 0.934. The apparent Brier score was 0.091, the apparent calibration slope was 1.000, and the apparent calibration intercept was 0.000. After bootstrap correction, the Brier score was 0.106, calibration-in-the-large was - 0.009, calibration intercept was - 0.010, calibration slope was 0.852, and the applied uniform shrinkage factor was 0.852. In formal incremental analyses, adding VCG features to a conventional base model improved AUC from 0.890 to 0.955 (ΔAUC 0.065; DeLong P = 0.001), with a lower Brier score and favourable discrimination/reclassification indices. A VCG-augmented model incorporating LVEDD, NYHA class, spatial QRS-T angles, and the QRS-loop reversal/U-turn sign showed promising internally validated performance for estimating 12-month MACE risk in hospitalized patients with CHF. External multicentre validation is required before clinical implementation.

BackgroundAcid-base disturbances are common in critically ill patients and provide immediate physiological insight and prognostic information. We aimed to characterize the epidemiology of specific acid-base patterns and to develop ABG-driven models for 28-day mortality.MethodsWe retrospectively analyzed 1150 critically ill adults with acid-base disorders identified from the first arterial blood gas (ABG) after ICU admission. The primary outcome was 28-day mortality. Three multivariable logistic regression models were developed: (1) a continuous model using physiological variables, (2) a pragmatic threshold-based model with Firth bias-reduced logistic regression, and (3) an ABG-only model. Discrimination (area under the receiver operating characteristic curve [AUC]), calibration, Brier score, and decision curve analysis (DCA) were reported with bootstrap optimism.ResultsMixed acid-base disorders were the most prevalent (35.7%) and associated with high mortality (73.2%), while respiratory alkalosis also carried high mortality (69.6%). Non-survivors showed an apparently "normal" or higher pH (7.40 vs 7.30) despite marked hypocapnia and elevated lactate (4.8 vs 2.5 mmol/L). The continuous model achieved an AUC of 0.94 (optimism-corrected 0.88). The threshold model achieved an AUC of 0.85, with lactate > 4 mmol/L (adjusted odds ratio [aOR] 7.68) and sepsis (aOR 2.78) as dominant predictors. The ABG-only model maintained high discrimination (AUC 0.92; optimism-corrected 0.88) with acceptable calibration.ConclusionsRoutine ABG combined with basic clinical data at ICU admission enables accurate early mortality risk stratification in ICU patients. An apparently normal pH may conceal severe metabolic stress, emphasizing the need for integrated acid-base assessment in critical care.

Hip fractures are a major global health issue with high mortality and morbidity, especially in older adults. One-year mortality post-surgery ranges from 22% to 36%, with many patients never regaining baseline mobility. While several predictors of mortality have been identified, their relative contribution to mortality risk within a unified survival prediction framework remains unclear. This study used machine learning to rank key perioperative predictors of mortality following hip fracture surgery. Over 11,000 patients from the Australian and New Zealand Hip Fracture Registry were analyzed. Twenty demographic, clinical, and perioperative variables were assessed using a Random Survival Forest (RSF) model. Model performance was evaluated using the concordance index and Brier scores. Permutation-based feature importance ranked predictors according to their contribution to predictive performance for mortality risk. During the follow-up period (median 630 days), 31% of patients died. The RSF model performed well (test C index: 0.7305). The four most important predictors of mortality were American Society of Anesthesiologists (ASA) grade (importance score: 0.051), pre-existing dementia (0.036), age (0.019), and preadmission walking ability (0.016). Other factors like male sex (0.009) and acute hospital stay (0.006) had weaker associations with mortality prediction in this model. Machine learning identified ASA grade, dementia, age, and mobility as the top predictors of mortality after hip fracture surgery. RSF modeling offered strong performance and better interpretability than traditional methods. These findings support individualized stratification to inform perioperative discussions and goals-of-care planning in this high-risk population.

To evaluate hippocampal subfield radiomics for identifying aggressive behavior in hospitalized patients with schizophrenia and to validate the incremental benefit and interpretability of a combined model integrating clinical variables, whole-brain structural MRI information, and hippocampal subfield radiomics. This retrospective single-center cohort included 247 hospitalized patients with schizophrenia, randomly split (7:3) into training and test sets using outcome-stratified sampling. Aggression during hospitalization before discharge was assessed with the Modified Overt Aggression Scale (MOAS); clinically significant aggression was defined as a weighted total score ≥4. Whole-brain gray matter volume (GMV) features and hippocampal subfield radiomics features segmented using the FreeSurfer pipeline were each subjected to maximum relevance minimum redundancy (mRMR) and 10-fold least absolute shrinkage and selection operator (LASSO) for feature selection to derive sMRI-Radscore and Hip-Radscore, respectively. Three extreme gradient boosting (XGBoost) models were compared. Discrimination, incremental value, calibration, and net benefit were evaluated using area under the curve (AUC), net reclassification improvement/integrated discrimination improvement (NRI/IDI), Brier score, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) was used to interpret the combined model. Five GMV and nine hippocampal subfield radiomics features were retained. sMRI-Radscore showed lower discrimination (AUC = 0.681) than Hip-Radscore (AUC = 0.729). The combined model achieved the highest AUC (0.875), outperforming both single models, with the lowest Brier score (0.119) and higher net benefit across most threshold probabilities. SHAP indicated Hip-Radscore as the top contributor. Hip-Radscore and its key constituent features were significantly associated with aggression severity. Hip-Radscore may help identify aggressive behavior risk in hospitalized patients with schizophrenia, and the combined model showed improved discriminative performance. These findings provide imaging-based evidence for early risk stratification and management during hospitalization.