2026-09-28

Purpose: To develop and externally validate a multi-phase CT deep learning framework for preoperative risk stratification of 2-5 cm gastric gastrointestinal stromal tumors (GISTs). Methods: This retrospective multicenter study included 427 patients with pathologically confirmed gastric GISTs. At Center A, patients treated from January 2015 to June 2022 formed the development set (n=240), in which all model selection, tuning, and early stopping were performed by five-fold cross-validation, and patients treated from July 2022 to December 2023 formed a temporally independent internal test set (n=59); Center B (n=128) served as the external validation set. Unenhanced, arterial, and venous CT images were analyzed with the Multi-Phase Fusion Network (MPFNet), which combines parallel ResNet50 encoders with a Cross-Phase Attention Module and phase-adaptive weighting. Incremental value over a combined clinical-radiomics reference model was assessed with the difference in AUC, net reclassification improvement (NRI), integrated discrimination improvement (IDI), calibration intercept and slope, and decision curve analysis; three radiologists served as additional comparators. Results: On the internal test set, MPFNet achieved an AUC of 0.935 (95% CI: 0.852-0.981) versus 0.872 for the combined model (P=0.016; NRI 0.58, IDI 0.112). On external validation, the AUC was 0.894 (95% CI: 0.827-0.942) versus 0.831 (P=0.091; NRI 0.41, IDI 0.078), with a calibration intercept of -0.14, slope of 0.87 (95% CI: 0.64-1.10), and Brier score of 0.118; net benefit exceeded that of the combined model at threshold probabilities of 0.30-0.50. The AUC was also higher than that of three radiologists (0.735-0.812 externally). External false negatives were concentrated in intermediate-risk tumors (9 of 11), and performance was lower on the external scanner than on the two Center A scanners. Conclusion: In retrospective external validation, MPFNet provided incremental discrimination over a combined clinical-radiomics model with acceptable calibration for preoperative risk stratification of 2-5 cm gastric GISTs. Its use as a decision-support tool for planning the resection approach requires prospective evaluation.

Postoperative desaturation, a significant concern following laparoscopic and robot-assisted procedures, is associated with increased mortality rates and prolonged stays in the recovery room. Although modern anesthesia workstations offer real-time spirometry pressure-volume loops for each breath, their clinical utility remains limited. This study aimed to develop a deep learning model using a convolutional neural network to analyze spirometry pressure-volume loops in patients undergoing laparoscopic or robot-assisted surgeries and to design a clinically applicable index using gradient-weighted class activation maps to enhance interpretability. This retrospective study was conducted across two tertiary hospitals, involving data from 207 patients who underwent laparoscopic or robot-assisted abdominal surgeries for model training. Data from 45 patients at an independent hospital were used for external validation. Thus, data from a total of 252 patients were used. The primary outcome measure was the model's ability to predict postoperative desaturation. The performance of the model, along with that of the derived clinical index, was evaluated using the area under the receiver operating characteristic curve (AUROC). The deep learning model predicted postoperative desaturation with AUROC (95% confidence interval) values of 0.825 [0.744-0.905] and 0.688 [0.656-0.720] on internal and external validation datasets, respectively. The desaturation predictive index demonstrated moderate predictive performance on both internal and external datasets (AUROC: 0.774 and 0.722, respectively). These findings support the efficacy of both the deep learning model and the predictive index in forecasting postoperative desaturation, thus highlighting the potential of real-time spirometry in postoperative monitoring research.

Background: PADS combines two neural networks predicting ICU mortality and discharge within 48 h, placing critically ill patients into one of four clinically meaningful states. Developed on MIMIC-IV alone, it left open whether it generalizes to other ICUs, whether its models transfer across hospitals, and whether its predictions can be explained at the bedside. Methods: We evaluated PADS on four ICU databases from different hospitals and countries (MIMIC-IV, AmsterdamUMCdb, eICU-CRD, and HiRID), using the same routinely collected variables. Mortality is scored on the final 48-h window (terminal-window, not early-warning, discrimination). For each external database, we compared the MIMIC model used as-is, retrained from scratch, and retrained from the MIMIC weights, and added an explainability layer. Results: For mortality, reusing and retraining the MIMIC model gave the highest discrimination on every database (AUROC 0.955-0.986; terminal-window (near-outcome) discrimination) and stabilized training; used as-is, it ranged from chance (Amsterdam) to good (eICU, HiRID). For discharge, training fresh on local data matched or beat reusing MIMIC on every external database, consistent with discharge timing depending on local organization rather than physiology. The explainability layer produced clinically coherent, cross-checked explanations. Conclusions: Transportability was task-dependent: mortality transferred between hospitals, discharge did not. PADS demonstrated promising external transportability across heterogeneous ICU databases, particularly after local adaptation. Reusing and adapting the MIMIC-IV mortality model across hospitals improves accuracy. This approach also stabilizes training, providing a basis for potential federated deployment, whereas discharge is better trained locally. The mortality results reported here are terminal-window discrimination and do not support use of the framework as an early-warning model. A transparent explainability layer provides an interpretable representation of model predictions, addressing a key barrier to clinical adoption.

Acute-on-chronic liver failure (ACLF) is characterized by a high short-term mortality rate. Hepatitis B virus (HBV) infection is a major cause of ACLF in the Asia-Pacific region. This study aimed to compare the performance of multiple prediction models and identify a stable approach for predicting 90-day mortality in patients with HBV-associated ACLF (HBV-ACLF). Clinical data from 577 patients with HBV-ACLF admitted to 2 hospitals were retrospectively collected. Among them, 513 patients from 1 hospital were randomly divided into a training cohort and a testing cohort at a ratio of 6:4, while 64 patients from another hospital were included as an external validation cohort. Six prediction approaches were evaluated, including logistic regression (LR), least absolute shrinkage and selection operator regression (LASSO), support vector machine (SVM), decision tree (DT), random forest (RF), and K-nearest neighbor (KNN). Model performance was assessed using accuracy, the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value, and negative predictive value. The RF model achieved excellent discrimination in the training cohort, with an AUC of 1.000. However, its performance declined in the testing cohort and external validation cohort, with AUC values of 0.912 and 0.728, respectively. In contrast, the LR model demonstrated more consistent predictive performance across the training, testing, and external validation cohorts, with AUC values of 0.905, 0.928, and 0.849, respectively. Overall, the LR model showed more consistent point estimates of discrimination across the internal testing and external validation cohorts than the more complex models. Although several machine learning approaches demonstrated good predictive performance, the LR showed the most stable and consistent performance across the testing and external validation cohorts. These findings suggest that increased model complexity does not necessarily translate into superior generalizability and that simpler, more interpretable models may provide robust prognostic prediction for patients with HBV-ACLF.

Background: Artificial intelligence-derived parameters hold substantial promise as indicators for tumor prognosis prediction and treatment guidance. However, existing studies have not sufficiently addressed the application of these parameters across different cancer types. Methods: We employed a deep learning algorithm to conduct a multi-cancer analysis for disease-free survival (MC-DFS) prediction using 8856 cases with associated whole-slide images and clinical data. The training cohort consisted of 7392 cases from the TCGA set (24 cancer types), and the independent external validation cohort comprised 1464 cases from the CPTAC and General Hospital sets (9 cancer types). A nomogram prediction signature for disease-free survival (NOMO) was developed by integrating the MC-DFS, tumor stage, and patient age. The prognostic model's performance was validated in an independent cohort. Results: In the training and validation cohorts, the MC-DFS model achieved area under the curve (AUC) values of 0.750 and 0.682, respectively. It effectively differentiated patients with poorer disease-free survival, with hazard ratios of 4.823 (95% CI: 4.343-5.356, p < 0.0001) in the training cohort and 2.092 (95% CI: 1.472-2.971, p < 0.0001) in the validation cohort. Each cancer subtype's analysis confirmed the model's robust performance. Additionally, using nomogram analysis, we developed a multi-model prediction signature for disease-free survival across multiple cancer types based on MC-DFS and the clinicopathologic features in the training cohort. This enhanced model offers more precise risk stratification for stage I malignancies and complements the existing tumor staging systems by identifying high-risk patients. Conclusions: The newly developed MC-DFS shows marked improvements in prognostic predictions across multiple cancer types. With further validations across multiple centers, this nomogram prediction system could become a valuable practical tool for managing various cancers.

This study aimed to develop a combined model integrating clinical, computed tomography (CT), and deep learning (DL) features to discriminate malignant subsolid nodules (M-SSNs) from persistent inflammatory subsolid nodules (PI-SSNs). A total of 573 patients with subsolid nodules (SSNs) who underwent preoperative chest CT and surgical resection at Center 1 were assigned to training and internal validation sets (ratio 4:1), and an external testing cohort comprising 199 patients from Center 2 was also enrolled. A clinical model was first developed using clinical and CT features that differed significantly between M-SSNs and PI-SSNs. Subsequently, five DL models were constructed based on 3D convolutional neural network architectures (VGG11, ResNet18, SEResNet34, Res2Net50, and DenseNet121), and the best-performing DL model was selected as the feature extractor. The extracted DL features were then combined with the selected clinical-CT features to form a comprehensive feature pool. Recursive feature elimination was applied to identify the most discriminative features. Finally, a combined model was built using the random forest algorithm on the refined feature subset. The performance of the clinical model, the optimal DL model, and the combined model was compared to determine the best-performing model. The clinical model was constructed using three features selected by the random forest algorithm: well-defined boundary, lobulation, and mixed ground-glass opacity with multiple solid components. In the external testing cohort, this model achieved an area under the curve (AUC) of 0.773, with an accuracy of 0.683, sensitivity of 0.654, specificity of 0.746, precision of 0.848, and F1-score of 0.739. Among the five DL models, the 3D Res2Net50 model performed best, attaining an AUC of 0.827 in the external test set, along with an accuracy of 0.809, sensitivity of 0.809, specificity of 0.810, precision of 0.902, and F1-score of 0.853. The combined model integrated three clinical-CT features with 20 discriminative DL features derived from the 3D Res2Net50 model. It achieved an AUC of 0.875 in the external testing cohort, with accuracy, sensitivity, specificity, precision, and F1-score values of 0.814, 0.824, 0.794, 0.896, and 0.858, respectively, and the lowest Brier score (0.146) among the three models, indicating the best probability calibration. In this cohort, the combined model showed superior performance compared to both the clinical model and the 3D Res2Net50 DL model. In a reader study, AI assistance significantly improved radiologists' majority-vote AUC from 0.783 to 0.833 (P < 0.05). The decision curve analysis results demonstrated that the combined model provided a higher standardized net benefit than the clinical model, the DL model, and both the treat-all and treat-none strategies across a wide range of threshold probabilities (all P < 0.05). The combined model based on clinical-CT and DL features is valuable in differentiating PI-SSNs from M-SSNs and contributes to optimizing surgical indications and guiding personalized treatment strategies for SSNs.

Hepatocellular carcinoma (HCC) patients show heterogeneous responses to immune checkpoint inhibitors (ICIs). This study developed ECOS-Net, a transformer-based multitask network integrating CT-derived habitat and 2.5-dimensional (2.5D) deep learning features for simultaneously predicting early disease control (DC) and overall survival (OS). Of 1,234 patients with HCC enrolled from eight institutions and public databases, 832 ICI-treated patients were used for model development. ECOS-Net fused features using multi-head attention and generated early DC probabilities and OS risk scores. ECOS-DC achieved AUCs of 0.836, 0.822, and 0.817 in training, internal validation, and external test sets, outperforming clinical models (all p values < 0.05). ECOS-OS yielded C-indices of 0.730, 0.722, and 0.720, respectively. Integrated models also showed favorable external performance (early DC AUC: 0.825; OS C-index: 0.741). Patients with higher ECOS-DC probabilities had a higher likelihood of early DC, whereas those with higher ECOS-OS risk had shorter OS, with directionally consistent associations across most subgroups. Exploratory biological analyses suggested that the higher ECOS-DC probability and lower ECOS-OS risk groups were associated with immune-active tumor microenvironment features. Therefore, ECOS-Net shows potential as a non-invasive imaging-based risk stratification framework for simultaneously predicting early DC and OS in ICI-treated HCC patients.

Background: Estimated pulse wave velocity (ePWV) is a non-invasive marker of arterial stiffness with potential relevance for chronic kidney disease (CKD) risk assessment. This study aimed to investigate the association between ePWV and CKD risk in the US and Chinese populations and to evaluate its discriminative performance using machine learning approaches. Methods: Data were obtained from the National Health and Nutrition Examination Survey (NHANES, 2005-2018, weighted n ≈ 100.9 million) and the China Health and Retirement Longitudinal Study (CHARLS, 2011-2015, n = 21,853). In NHANES, CKD was defined according to the 2021 KDIGO criteria as estimated glomerular filtration rate (eGFR) < 60 mL/min/1.73 m2 or a urinary albumin-to-creatinine ratio (ACR) ≥ 30 mg/g; in CHARLS, where urinary albumin was not measured, CKD was defined by the eGFR criterion alone, and this difference in case definition was taken into account when interpreting the results. Logistic regression and restricted cubic spline models were used to examine the association between ePWV and CKD, with subgroup analyses stratified by demographic and clinical characteristics. Multiple machine learning models were developed in NHANES and externally validated in CHARLS; model discrimination was assessed using the area under the receiver operating characteristic curve (AUROC), and feature importance was interpreted using SHapley Additive exPlanations (SHAP) values. Results: Higher ePWV was consistently associated with higher odds of CKD in both cohorts (NHANES: odds ratio [OR] = 1.505 per 1 m/s, 95% confidence interval [CI] = 1.459-1.552; CHARLS: OR = 1.434 per 1 m/s, 95% CI = 1.368-1.503; both p < 0.0001), with dose-response relationships observed. Formal interaction tests confirmed significant effect modification by sex (both cohorts) and by BMI and diabetes (NHANES only). Point estimates were higher in men and in NHANES obese and diabetic subgroups, but interactions across smoking and alcohol strata were not statistically significant. Among the machine learning models evaluated, discrimination was moderate and comparable across algorithms (LightGBM: AUROC = 0.804 in internal validation and 0.793 in external validation), and DeLong tests showed no significant difference between LightGBM and XGBoost (p = 0.0655 and 0.0684, respectively). SHAP analysis identified ePWV as the highest-ranking feature, surpassing uric acid, lipid levels, and diabetes history. Using the Youden index, the optimal ePWV cutoff for identifying CKD was 10.15 m/s in NHANES (sensitivity 0.658, specificity 0.695) and 10.568 m/s in CHARLS (sensitivity 0.658, specificity 0.718). Conclusions: Elevated ePWV is significantly associated with eGFR-defined CKD across the US and Chinese populations studied. These findings support ePWV as a potentially useful marker for CKD risk stratification; however, given the cross-sectional design of both cohorts, its predictive value requires confirmation in prospective studies. It is important to emphasize that no non-invasive calculated metric can replace direct measurement of serum creatinine and urinalysis for identifying individuals at risk of CKD in routine clinical practice.