2026-10-08

Background: Preoperative nodal assessment in colorectal cancer (CRC) remains imperfect, and uncertainty regarding pathological lymph node metastasis (LNM) may persist despite conventional imaging. Machine learning using routinely available preoperative information may provide complementary patient-level risk estimates without requiring radiomic feature extraction or molecular testing. Objective: This study aimed to develop and externally validate machine learning models for preoperative estimation of pathologically confirmed LNM risk in patients with newly diagnosed CRC undergoing direct radical surgery without preoperative anticancer treatment, and to assess whether multivariable prediction adds information beyond clinical tumor stage (cT) alone. Methods: This multicenter retrospective study included 2725 patients from 2 hospitals. The internal cohort (n=1074) was split into training (n=753) and internal test (n=321) sets, while 1651 patients from a second center formed the external validation cohort. Candidate predictors were routinely available preoperative clinical, laboratory, tumor-marker, imaging-based, and endoscopic biopsy variables. Predictors were selected using univariable and multivariable logistic regression (LR). Seven algorithms were evaluated: random forest (RF), support vector machine (SVM), LR, decision tree (DT), extreme gradient boosting (XGBoost), naive Bayes (NB), and light gradient boosting machine (LightGBM). Performance was assessed using discrimination, threshold-dependent classification metrics, calibration, and decision curve analysis (DCA). Shapley additive explanations (SHAP) were used for interpretation. Sensitivity analyses compared the selected predictors with all candidate predictors and cT alone. Results: Six predictors were retained: BMI, preoperative carcinoembryonic antigen, primary tumor site, cT, histological type, and tumor differentiation. RF was selected based on internal test performance. The area under the receiver operating characteristic curve (AUROC) was 0.803 (95% CI 0.756-0.849) in the internal test set and 0.776 (95% CI 0.755-0.797) in external validation. In external validation, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were 0.687, 0.677, 0.693, 0.569, and 0.781, respectively. The Brier score was 0.190, and the calibration slope was 1.061. cT contributed most in SHAP analysis; however, the 6-predictor RF model showed significantly higher discrimination than cT alone in both the internal test set (AUROC 0.803 vs 0.712; ΔAUROC=0.091; P<.001) and external validation cohort (AUROC 0.776 vs 0.707; ΔAUROC=0.069; P<.001). Discrimination did not differ significantly between the 6-predictor and all-predictor RF models. Conclusions: We developed and externally validated a 6-predictor RF model for preoperative LNM risk estimation in patients with newly diagnosed CRC undergoing radical resection with regional lymph node dissection without preoperative anticancer treatment. The model showed moderate external discrimination and provided predictive information beyond cT alone. It is intended for research-stage adjunctive risk stratification and should not replace imaging, pathological evaluation, or clinical judgment. Prospective validation in broader populations and clinical settings is required before implementation.

Background: Sepsis is a prevalent and life-threatening complication in ICU patients after intracranial hematoma evacuation. This study aimed to develop and externally validate an interpretable machine learning model for predicting sepsis risk in this high-risk neurocritical care population. Methods: A multicenter retrospective study was designed. The MIMIC-IV was based on for model development and internal testing; MIMIC-III, eICU and an independent clinical cohort from Jiangsu Provincial Hospital of Integrated Traditional Chinese and Western Medicine for external validation. Sepsis during ICU stay was defined according to the operationalized Sepsis-3 criteria. Early ICU variables were selected using Boruta and LASSO, and six machine learning algorithms were developed and compared. Model performance was evaluated by discrimination, calibration, and decision curve analysis, and its interpretability was assessed using SHAP and LIME. A sensitivity analysis was conducted by retraining all six models after excluding baseline SOFA. Results: A total of 1,729 patients were included, of whom 717 developed sepsis during ICU stay. The final model retained 13 predictors, including WBC, SOFA score, SpO2, glucose, chloride, pneumonia, aspirin use, mannitol use, vasopressor use, mechanical ventilation, RBC, sedative use, and GCS score. XGBoost showed the most balanced overall performance, with AUCs of 0.794 in the internal test set and 0.820, 0.837, and 0.796 in the MIMIC-III, eICU, and independent clinical validation cohorts, respectively. SHAP analysis identified SOFA score, vasopressor use, glucose, pneumonia, and sedative exposure as major contributors to model predictions, while LIME supported individual-level interpretability. A web-based calculator was implemented for individualized risk prediction. Sensitivity analysis showed that XGBoost retained discrimination after SOFA exclusion, with no significant AUC changes across the external cohorts. Conclusion: This study developed and externally validated an interpretable XGBoost-based model for predicting sepsis risk during ICU stay among patients after intracranial hematoma evacuation. The model showed generally consistent AUC-based discrimination across multiple validation cohorts and may support individualized risk stratification. Further prospective validation and clinical impact assessment are required before routine clinical implementation.

Background: To compare the predictive performance of computed tomography (CT) body composition indices, anthropometric indices, and laboratory indices for gallstones occurrence, and to construct machine-learning models to improve performance. Methods: The dual-center retrospective cohort enrolled patients who underwent initial abdominal CT between January 2017 and January 2023, had no gallstones detected, and completed at least 3 years of follow-up. The data analysis was performed in April 2026. They were divided into gallstone group and non‑gallstone group by follow‑up findings. A deep-learning tool, Body and Organ Analysis (BOA), was used to quantify fat, muscle, and bone at the level of the third lumbar vertebra. The area under the receiver operating characteristic curve (AUC) of these indices was compared with that of anthropometric and laboratory indices. Predictive models were developed in the training cohort. Model performance was evaluated using fivefold cross-validation and an independent test cohort. Results: 1,944 patients were evaluated, including 1,437 in the training cohort (Center 1; median age, 63 years [25th-75th percentile, 55-72]; 699 males) and 507 in the test cohort (Center 2; median age, 63 years [55-71]; 266 males). In univariate analysis, neutrophil-to-lymphocyte ratio (NLR) showed the highest AUC (0.627, 95% confidence interval [CI]: 0.586-0.668). The extreme trees (ET) model performed best, with a test-set AUC of 0.772 (95% CI: 0.714-0.822). SHapley Additive exPlanations (SHAP) analysis identified the area ratio of subcutaneous to total fat as the most important feature. Conclusions: NLR was the best single predictor but had limited standalone utility. Among models integrating the three indicator categories, the ET performed best.

Background: Early extrahepatic recurrence (eEHR) after hepatectomy for hepatocellular carcinoma (HCC) significantly affects patient prognosis; however, an accurate predictive framework integrating radiomics with clinicopathological features is currently lacking. This study aims to develop and validate a multimodal machine learning model for prediction of early postoperative extrahepatic recurrence risk, thereby offering a basis for individualized postoperative intervention strategies. Methods: A total of 208 patients who received surgical resection and had a postoperative pathological diagnosis of HCC were retrospectively enrolled from two centers between 2018 and 2023. The cohort was divided into a training set and an internal validation set at a 7:3 ratio. An external validation set was also established. Preoperative enhanced CT radiomic features and clinicopathological characteristics were extracted. After screening for key features using regression analysis, multiple machine learning prediction models were constructed via nested k-fold cross validation. The performance of the models was evaluated, and survival analysis was subsequently conducted. Results: Among the 208 patients, 55 (26.4%) had postoperative eEHR. The most common site of recurrence was the lung (26 cases, 47.3%), followed by lymph nodes (12 cases, 21.8%) and bone (10 cases, 18.2%). Three core clinicopathological features and four key radiomic features were identified. A multimodal predictive model for HCC-eEHR was developed using the Random Forest algorithm. This model achieved an AUC of 0.852 (sensitivity 0.511, specificity 0.959) in the training cohort and an AUC of 0.836 (sensitivity 0.750, specificity 0.875, accuracy 0.850) in the external validation cohort. A 24-month landmark analysis (n=156) confirmed that patients developing eEHR within 24 months had significantly worse post-landmark overall survival (median 16.3 vs 47.7 months; HR=3.342, 95%CI: 1.570-7.112, P<0.001). Furthermore, the model demonstrated effective risk stratification in both cohorts (Southwest OS: HR=3.549, 95%CI: 2.097-6.006, P<0.001; Leshan OS: HR=3.839, 95%CI: 1.171-8.369, P=0.026). Conclusion: The developed multimodal prediction model may assist in assessing the risk of eEHR following HCC hepatectomy, thereby facilitating the timely identification of high-risk patients and informing personalized postoperative surveillance and treatment strategies.

Background: Asthma is a common chronic respiratory disease that imposes a heavy burden of morbidity and mortality in most countries. Eosinophils have become major therapeutic targets for severe eosinophilic asthma and other eosinophil-associated diseases. This study investigated the role of eosinophils in asthma through three complementary approaches: (i) Mendelian randomization to test causal associations, (ii) single-cell transcriptomic analysis to explore potential biological mechanisms underlying elevated blood eosinophil counts, and (iii) clinical data analysis and machine learning modelling to evaluate their predictive value in clinical practice. These analyses demonstrated the potential value of eosinophils in clinical applications. Methods: We conducted three complementary but methodologically independent analyses in adult asthma patients. First, a two-sample Mendelian randomization analysis using publicly available GWAS summary statistics was performed to assess whether genetically predicted eosinophil count and eosinophil percentage are causally associated with asthma susceptibility. Second, a publicly available peripheral blood mononuclear cell scRNA-seq dataset (GSE248688), which contrasts asthma patients with healthy controls, was used to characterize transcriptomic differences and intercellular communication in the potential biological mechanisms for increased blood eosinophil counts in adult asthma. Third, using retrospectively collected data from two adult asthma cohorts (SWMU, n = 370, training/internal validation; Chongqing, n = 180, external validation), we developed and externally validated machine learning models for 12-month exacerbation risk. All predictors were measured at a clinically stable baseline visit preceding the 12-month outcome window. Results: The Mendelian randomization analysis provided genetic evidence that eosinophil count is causally associated with asthma susceptibility (IVW: OR 1.005, 95% CI 1.002-1.008, P < 0.001), whereas the signal for eosinophil percentage was small and borderline (IVW: OR 0.984, 95% CI 0.970-0.998, P = 0.027) and directionally inconsistent with the clinical data, most plausibly reflecting the compositional nature of a percentage measurement. Overall, eotaxins and type 2 inflammation provide a plausible explanation for increased blood eosinophil counts in asthma patients. Further analysis revealed that the platelet subcluster in the asthma group highly expressed CCL26 (eotaxin-3). Upregulated transcriptional activity of GATA3 and MAF was observed in CD14+ monocytes. Cell-cell communication analysis indicated network remodeling. Specifically, three cell types-CD14+ monocytes, platelets, and hematopoietic stem and progenitor cells (HSPCs)-simultaneously showed upregulated sender activity for MIF signaling. The five machine learning models achieved external-cohort AUCs ranging from 0.718 to 0.769 (RF 0.763, XGBoost 0.769, LR 0.755, SVM-RBF 0.905, KNN 0.733), with substantially overlapping 95% CIs; a parsimonious three-variable benchmark model achieved a comparable AUC of 0.892. Conclusion: Peripheral blood eosinophils may play an important role in asthma. They are especially promising as potential clinical predictors of asthma exacerbation. However, the mechanisms underlying their increase are complex. We speculate that three pathways may be involved: bone marrow release, peripheral survival, and peripheral activation. Therefore, both the clinical utility and the specific biological mechanisms of peripheral blood eosinophils still require further validation.