2026-10-26

Rationale: Obstructive sleep apnea (OSA) is highly prevalent yet largely under-diagnosed. Current screening strategies rely on effortful questionnaires with modest accuracy, and prior machine learning models often require resource-intensive inputs and lack external validation. Objectives: To develop and validate machine learning models to predict OSA (apnea-hypopnea index [AHI4%]≥5) and moderate-severe OSA (AHI4%≥15) using only routinely collected electronic health record (EHR) data, test cross-system transferability and compare performance with clinical screeners and a Bayesian risk-updating framework. Methods: Three tree-ensemble families (XGBoost, random forests, and histogram-based gradient boosting) were trained on an adult cohort with sleep-study results at Kaiser Permanente (KPSC), using full (106-feature) and minimal (age, sex, body-mass index, race/ethnicity) predictor sets and tuned by nested cross-validation. XGBoost models were selected, internally and temporally validated, and externally validated at University of Kansas Medical Center (KUMC). Measurements And Main Results: On our internal hold-out set (n=53,048), the full XGBoost model reached a ROC-AUC of 0.791 for any OSA and 0.768 for moderate-severe OSA. The minimal model retained most performance (0.764 and 0.736). Temporal performance was stable. Externally, frozen models showed lower AUCs (0.718 and 0.690 for full models; 0.725 and 0.706 for minimal models) but similar to KUMC-native models and improved further with transfer learning. The models significantly outperformed STOP-BANG (AUC difference 0.06-0.09 for predicting any OSA; p<0.001). Bayesian updating with select patient-reported items further improved discrimination. Conclusions: Machine learning models based on EHR data can predict OSA probability that outperform existing screeners, transfers across systems, and offers potential for EHR implementation. At A Glance Commentary: Scientific Knowledge on the Subject: Obstructive sleep apnea is common and heavily underdiagnosed, yet upstream patient identification still depends on symptom and anthropometry questionnaires such as STOP-BANG, which are effort-dependent and achieve modest performance. Prior machine learning models for OSA are heterogeneous, often developed on small or demographically narrow cohorts, frequently rely on physiologic signals that reintroduce a testing step, and are rarely validated across independent health systems.What This Study Adds to the Field: We developed machine learning models from routinely collected electronic health record data, comprising a full (106-feature) and a minimal (4-feature) model, to predict any OSA and moderate-severe OSA. The study is distinguished by multiple validation efforts, including temporal validation and external validation at a second health system. We further explored adaptation strategies for external adoption using transfer learning techniques. Finally, the models outperformed existing screeners such as STOP-BANG, and we describe a Bayesian framework that optionally incorporates select patient-reported symptoms to refine individual risk estimates. Overall, these models are feasible to implement within EHR systems as a low-effort, passive screener that can support encounter-based and population health outreach screening strategies.

Background: Visceral adipose tissue (VAT) is associated with tumor progression in gastric cancer (GC), but its added value for predicting lymph node metastasis (LNM) remains unclear. Purpose: This study aimed to construct a DR-DLR model that incorporates CT-based features from both VAT and tumor, with the goal of predicting LNM in GC patients, while investigating its biological relevance. Methods: 855 GC patients were retrospectively collected across three clinical centers, divided into a primary cohort, one internal test cohort, and two independent external validation cohorts. An Attention Residual Transformation Network (ARTNet) extracted imaging representations from CT images. Feature representations and clinical variables were integrated using a multilayer perceptron (MLP) to build the DR-DLR model. Model performance, generalizability, prognostic value, and biological relevance were evaluated using internal/external testing, radiologist comparison, survival analysis, and radiogenomics analysis. Results: Incorporation of VAT-derived features improved predictive performance compared with tumor-only models. The DR-DLR model achieved areas under the receiver operating characteristic curve (AUC) of 0.898 (95 % CI: 0.817-0.946), 0.834 (95 % CI: 0.743-0.897), and 0.848 (95 % CI: 0.775-0.901) in the internal and two external test cohorts, respectively. It outperformed junior and senior radiologists in LNM assessment. Overall survival was shorter in high-risk patients than in low-risk. Radiogenomics analysis revealed that the high-risk phenotype was associated with activation of immune-related pathways and increased infiltration of CD4 + T cells, plasma cells, dendritic cells, and mast cells. Conclusion: The DR-DLR model improved LNM prediction, with its high-risk phenotype linked to immune-related molecular alterations, supporting its value for preoperative risk stratification.

Background: Spontaneous intracerebral hemorrhage (ICH) carries a 30-day mortality exceeding 40%. The glucose-to-albumin ratio (GAR) reflects metabolic stress and nutritional status. Although previous studies have demonstrated its prognostic value in ICH, its utility specifically in ICU-admitted patients and its comparative performance against other albumin-derived indices remain incompletely characterized. Methods: We included 1,625 ICU ICH patients from MIMIC-IV (derivation) and 861 from two Chinese centers (external validation). Log-transformed GAR (LogGAR) was the exposure; 28-day ICU mortality was the primary outcome. Cox regression, restricted cubic splines, Kaplan-Meier, and Fine-Gray competing-risk models assessed LogGAR's independent prognostic value, compared head-to-head with five albumin-derived indices. Lasso-selected features informed a multilayer perceptron (MLP), benchmarked against severity scores and a penalized logistic regression on identical features. Results: LogGAR was independently associated with 28-day ICU mortality (derivation: HR = 1.81, 95% CI 1.33-2.47, p < 0.001; external: HR = 3.48, 95% CI 2.54-4.78, p < 0.001), with linear dose-response (RCS P-nonlinear = 0.315). Fine-Gray competing-risk analysis confirmed robustness (subdistribution HR = 2.01, 95% CI 1.58-2.57, p < 0.001). Adding LogGAR to six severity scores improved discrimination (external ΔAUC = 0.085-0.107, all p < 0.001). Among six indices, LogGAR was comparable to logLDAR in the derivation cohort (0.679 vs. 0.681, p = 0.78) and highest in the external cohort (0.691). Subgroup analyses identified significant effect modification only for AKI (FDR = 0.029). The Lasso-MLP (8 features) and penalized logistic regression achieved comparable performance (validation AUC 0.718 for both; external AUC 0.759 vs. 0.783), indicating no added value from nonlinearity. Calibration slopes were 0.66 (validation) and 0.91 (external). SHAP identified LogGAR as the top predictor. Conclusion: LogGAR is independently associated with 28-day ICU mortality in critically ill patients with ICH and shows discriminative performance comparable to logLDAR and superior to the remaining four indices. A penalized logistic regression on Lasso-selected features matched the MLP, suggesting the predictive signal lies in features rather than model complexity-a pragmatic consideration for clinical implementation. As a readily available biomarker, LogGAR may support early ICU risk stratification.

Background: Early identification of high-risk small bowel obstruction remains clinically challenging, especially when prompt intervention is needed to prevent bowel ischemia and related adverse outcomes. Methods: In this multicenter retrospective study, 274 patients with small bowel obstruction from three hospitals were included. Patients from Hospital A were randomly divided into training and test cohorts, and patients from the other two hospitals formed an external validation cohort. Radiomics features were extracted from venous-phase contrast-enhanced CT images, and logistic regression, naive Bayes, Extra Trees, Light Gradient Boosting Machine (LightGBM), and multilayer perceptron models were developed. Independent clinical predictors were identified and integrated with radiomics features to build a combined model. Results: LightGBM showed the most stable internal performance, with AUC values of 0.934 and 0.870 in the training and test cohorts, respectively. Lower diastolic blood pressure and bowel wall thickening were independent predictors of high-risk status. The combined model achieved AUC values of 0.942, 0.867, and 0.803 in the training, test, and external validation cohorts, respectively, with an external validation accuracy of 0.833 and sensitivity of 0.946. Conclusion: CT-based radiomics can identify patients with small bowel obstruction at high risk, and a combined model integrating radiomics and clinical predictors may support earlier risk stratification and intervention.

Background: Although combining systemic and locoregional therapies has improved outcomes for hepatocellular carcinoma (HCC), identifying appropriate candidates remains challenging because no reliable tool predicts response. Methods: This retrospective multicenter study enrolled 334 HCC patients who received first-line locoregional-systemic therapy from nine centers in China. Following automated tumor segmentation using nnU-Net on pretreatment contrast-enhanced magnetic resonance imaging, predefined radiomics features were computationally extracted, and deep learning features were obtained via a ResNet convolutional neural network. Three types of predictive models for tumor response (radiomics, deep learning, and combined) were developed and validated on the training cohort (n = 213), internal validation cohort (IVC, n = 53), and external validation cohort (EVC, n = 68). Performance was assessed using area under the receiver operating characteristic curve (AUC) and net reclassification improvement (NRI). Associations of model with progression-free survival (PFS) and overall survival (OS) were evaluated. Results: Among the three models (multi-sequence radiomics, multi-sequence deep learning, and combined), the combined model showed the best predictive performance in both the training (AUC: 0.997, 95% CI: 0.994-1.000) and validation cohorts (AUC of IVC 0.900, 95% CI: 0.806-0.994; AUC of EVC: 0.871, 95% CI: 0.790-0.951). The combined model demonstrated significantly improved predictive performance compared with the multi-sequence radiomic model (training NRI = 43.9%, p < 0.001; IVC NRI = 38.7%, p = 0.038; EVC NRI = 3.2%, p = 0.806) and multi-sequence deep learning model (training NRI = 80.8%, IVC NRI = 72.7%, EVC NRI = 50.4%; all p < 0.05). Specifically, the combined model was significantly associated with both PFS and OS. Conclusions: The combined radiomics-deep learning model outperformed both single-modality models in predicting objective response and stratifying survival outcomes in unresectable HCC patients receiving combined systemic and locoregional therapy.

Objectives: To develop a deep learning multimodal fusion model based on ultrasound (US) images and clinical-US texts to differentiate benign and malignant renal tumors. Methods: The retrospective dataset from January 2012 to February 2023 in a tertiary hospital was used for model development and randomly allocated to training and validation sets. An independent prospective test set was obtained from March 2023 to October 2024. We developed Model-G using grayscale US (GUS) images, Model-GC using GUS and color Doppler flow imaging images, Model-T using clinical-US texts, and Model-UT using US images and clinical-US texts. ResNet50 was used for image feature extraction, BERT for text encoding, and ViT for multimodal classification. Results: In total, 1008 patients (740 in training set, 186 validation set, and 82 prospective test set) with a single renal tumor were included. In the validation set, Model-UT achieved the highest diagnostic performance in comparison to Model-G, Model-GC, and Model-T with an area under the receiver operating characteristic curve (AUC) of 0.89 (95% CI, 0.86-0.93) (all p-values < 0.05). In the prospective test set, Model-UT yielded an AUC of 0.92 (95% CI, 0.86-0.97), outperforming high-level, mid-level, and low-level radiologists (all p-values < 0.05). Conclusions: The multimodal fusion model presents promising potential in differentiating benign and malignant renal tumors. Advances In Knowledge: Deep learning model integrating US images with clinical-US texts improved benign-versus-malignant renal tumor differentiation. Decision-curve, controlled ablation, and SHAP analyses support its potential use as a clinical decision-support tool.

Optimal patient selection for consolidative metastasis-directed radiotherapy (MDRT) after chemoimmunotherapy (CIT) remains uncertain in nasopharyngeal carcinoma (NPC) with exclusive distant metastasis, defined as distant metastasis without synchronous locoregional recurrence. This study was aimed at estimating the survival benefit of MDRT and identifying patients most likely to benefit. A total of 1083 patients from two independent centers were analyzed using a target trial emulation with cloning, artificial censoring, and inverse probability of censoring weighting. The emulation estimated a 4.2-month gain in 3-year progression-free survival (PFS) restricted mean survival time with the MDRT strategy (95% CI: 2.4-5.9; bootstrap p < 0.001). Eight survival machine-learning algorithms were developed and compared, among which the gradient boosting machine achieved 3-year PFS AUROCs of 0.875 in internal validation and 0.877 in geographical external validation. Heterogeneous treatment effects were further estimated using a T-learner framework across risk groups defined by recursive partitioning analysis. In the external cohort, model-estimated 3-year absolute risk reductions with MDRT were 21.6%, 14.7%, and 5.7% in intermediate-, low-, and high-risk groups, respectively. The framework was subsequently implemented as a web-based clinical decision-support system. These findings support individualized selection of consolidative MDRT after CIT and provide a practical framework for multidisciplinary decision-making.

Background: Depression is a common yet underrecognized comorbidity in chronic obstructive pulmonary disease (COPD). Existing prediction models often lack external validation, limiting cross-population generalizability. We aimed to develop and externally test machine learning models for predicting concurrent elevated depressive symptoms in COPD using three international aging cohorts. Methods: We included COPD patients from the Health and Retirement Study (HRS, USA, n=1,899) and the China Health and Retirement Longitudinal Study (CHARLS, China, n=2,206) for model development, and the English Longitudinal Study of Ageing (ELSA, UK, n=538) for external testing. Elevated depressive symptoms were defined using harmonized cutoffs (CES-D‑8≥3 for HRS/ELSA; CES-D‑10≥10 for CHARLS). Following recursive feature elimination, seven machine learning algorithms were trained and assessed for discrimination, calibration, and clinical utility. SHapley Additive exPlanations (SHAP) analysis was employed for model interpretability. Results: Eight variables were used to build the seven models. The XGBoost model achieved the best internal performance (training AUC 0.824, 95% CI: 0.809-0.838; validation AUC 0.757, 95% CI: 0.731-0.783) and maintained robust external validity in ELSA (test AUC 0.785, 95% CI: 0.748-0.824). Calibration and decision curve analyses supported clinical utility. SHAP analysis identified functional dependency, self‑rated health, and life satisfaction as predominant predictors. Higher dependency and poorer self-rated health had positive SHAP contributions, indicating higher model-predicted risk, whereas higher life satisfaction had negative contributions, indicating lower model-predicted risk. BMI exhibited a U‑shaped association. Conclusion: The XGBoost model generalized to an independent cohort and was deployed as a research prototype web application for risk stratification. SHAP analysis identified important model contributors for risk stratification in COPD, but these explanations describe model predictions and do not establish causal intervention targets. Prospective studies are needed to evaluate clinical utility.

Background: Early screening for depressive symptoms in middle-aged and older adults is important for timely risk identification. Low haemoglobin may identify a clinically accessible vulnerable subgroup in which chronic disease burden, fatigue, functional impairment, and reduced physiological reserve often coexist. However, it should be understood as a high-risk phenotype rather than a direct measure of nutritional deficiency or frailty. Evidence in this population remains limited. Methods: This study developed a model using data from the 2015 China Health and Retirement Longitudinal Study (CHARLS), with CHARLS 2011 and a cross-sectional cohort from Gansu Provincial People's Hospital used for temporal and external validation, respectively. Depressive symptoms were defined as a 10-item Centre for Epidemiologic Studies Depression Scale score ≥10. Candidate variables included sociodemographic, behavioural, health status, functional, cognitive, and comorbidity-related information. After least absolute shrinkage and selection operator feature selection, 16 variables were retained. Seven machine learning models were compared, and Random Forest was retained for interpretation and online tool development based on its balanced performance and interpretability. Results: A total of 2,022 participants were included in the development cohort, 1,095 in the temporal validation cohort, and 400 in the external validation cohort. The AUROC values of the Random Forest model were 0.791, 0.776, and 0.809, respectively, with Brier scores of 0.182, 0.193, and 0.182, indicating stable discrimination and acceptable calibration. Shapley additive explanations analysis showed that sleep time, instrumental activities of daily living score, and activities of daily living score were the most important contributors. Threshold sensitivity analysis showed that lower probability thresholds increased sensitivity/recall but reduced specificity. A prototype online assessment tool was developed for individualised estimation of depressive symptom probability. Conclusion: This study developed and validated an interpretable machine learning model for identifying depressive symptoms in middle-aged and older adults with low haemoglobin. The model may support early screening and risk stratification in this clinically vulnerable population. However, the classification threshold and risk-stratification cut-offs are exploratory rather than clinically validated. Further multicentre validation is warranted before broader clinical implementation.

Background: Adjuvant therapy decision in early-stage node-negative breast cancer remains one of the most challenging problems in oncology. Although most patients in this population achieve favorable outcomes with endocrine therapy alone, a subset remains at risk of relapse and may benefit from chemotherapy. Accurate identification of high-risk and low-risk patients is therefore essential to prevent both undertreatment and overtreatment. Methods: In this study, machine learning models were assessed using MSK-IMPACT (n=645) and METABRIC (n=503) cohorts comprising node-negative stage I-II breast cancer patients. Three experimental settings were evaluated: training on MSK-IMPACT and testing on METABRIC, the reverse direction, and five-fold cross-validation on the combined cohort. Sensitivity analysis and data harmonization were applied to identify and mitigate cohort-specific bias. Survival analysis, SHapley Additive exPlanations (SHAP) interpretability, and ablation analyses were performed to validate clinical relevance and characterize feature contributions. Results: The combined cohort model achieved an area under the receiver operating characteristic curve (AUC) of 0.788 ± 0.028 across five-fold cross-validation, while bidirectional cross-cohort validation yielded AUC of 0.741 when testing on MSK-IMPACT and AUC of 0.714 when testing on METABRIC. Survival validation on METABRIC out-of-fold predictions from the combined cohort model showed hazard ratios of 3.01 (95% CI: 2.28-3.98) for disease-free survival and 2.35 (95% CI: 1.79-3.09) for overall survival, consistent across Luminal A, Luminal B, and Triple-Negative subtypes. Conclusion: Validated across two geographically independent cohorts, the proposed model enables cost-effective, gene expression-free identification of patients at high-risk of distant relapse and may facilitate adjuvant chemotherapy decision-making in early-stage node-negative breast cancer.