2026-10-27

Background: Approximately 10-60% of heart failure with reduced ejection fraction (HFrEF) patients may transition to heart failure with improved ejection fraction (HFimpEF), a clinically meaningful phenotype associated with better prognosis. However, existing prediction tools demonstrate suboptimal discrimination and poor clinical utility. We aimed to develop and validate interpretable models for HFimpEF prediction by comparing logistic regression with ensemble machine learning. Methods: This retrospective cohort study included 1,257 HFrEF patients from the First Affiliated Hospital of USTC (March 2015-October 2023), split into training (n = 1,006) and internal validation (n = 251) sets. External validation used data from 310 patients across eight hospitals in Anhui Province (January 2024-January 2025). HFimpEF was defined as LVEF increase ≥10% to >40%. Five models were developed: data-driven logistic regression (Model 1); literature-based logistic regression (Model 2) with 5 common predictors validated in both BIOSTAT-CHF and ASIAN-HF cohorts; random forest (Model 3) and XGBoost (Model 4), both tuned via 5-fold cross-validation; and a simplified ensemble (Model 5) that averaged the predicted probabilities of Models 1, 3, and 4, then calibrated the average via logistic regression. Performance was assessed using AUC for discrimination, Brier score with calibration plots for calibration, decision curve analysis for clinical utility, and DeLong's test for model comparisons; SHAP analysis was applied for interpretability of Models 3 and 4. Results: Median age was 65 years (IQR 53-73); 69.9% were male. Multivariable analysis identified seven predictors: age (OR 0.97, 95% CI 0.96-0.99), BMI (OR 1.04, 95% CI 1.00-1.09), systolic blood pressure (per 10 mmHg increase, OR 1.13, 95% CI 1.01-1.25), prior PCI (OR 0.45, 95% CI 0.26-0.79), HDL-C (OR 2.48, 95% CI 1.36-4.50), FT3 (OR 1.21, 95% CI 1.02-1.45), and LVESD (OR 0.92, 95% CI 0.87-0.99). In external validation, Model 5 showed superior discrimination (AUC 0.67 vs. 0.60 for Model 2, P = 0.004) but poorer calibration (Brier score 0.34). Model 1 demonstrated balanced performance (AUC 0.66, Brier score 0.23). Conclusions: Our findings indicate that ensemble learning methods provide a moderate enhancement in discrimination for HFimpEF prediction, while logistic regression models exhibit superior robustness and interpretability. We have developed an online tool (https://alslyy-wq.shinyapps.io/USTC-HF-HFimpEF-Prediction/) that integrates both methodologies to facilitate practical HFimpEF risk assessment.

Objective: Hemorrhagic transformation (HT) remains a major complication after endovascular therapy (EVT) in patients with acute ischemic stroke (AIS). This study aimed to develop and validate a machine-learning model based on T2-FLAIR white matter hyperintensity (WMH) radiomics combined with clinical variables to predict HT after EVT. Methods: In this dual-center retrospective study, 446 patients (mean age 61.88 ± 10.98 years; 332 [74.4%] male) with AIS who underwent EVT were enrolled from two centers. Patients from the primary center were randomly assigned to a training cohort and an internal test cohort, and patients from the second center served as an independent external validation cohort. Clinical risk factors were identified by univariate and multivariate logistic regression. Preoperative T2-FLAIR images were used to segment WMH lesions and extract radiomics features. After feature selection, three predictive models were then developed: a clinical model based on the independent clinical predictors, a radiomics model based on the radiomics signature, and a combined model integrating both. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration analysis, DeLong testing, and decision curve analysis. Results: Among the 446 patients, 153 (34.3%) developed HT within 24-48 h after EVT. Multivariate analysis identified NIHSS score, occlusion site, and atrial fibrillation as independent clinical predictors of HT. In the internal test cohort, the combined model achieved the best discriminative performance (AUC = 0.818, 95% CI: 0.651-0.986), outperforming the clinical model (AUC = 0.740, 95% CI: 0.562-0.917) and the radiomics model (AUC = 0.806, 95% CI: 0.621-0.990). In the external validation cohort, the combined model maintained robust performance (AUC = 0.817, 95% CI: 0.746-0.889), which was comparable to the clinical model (AUC = 0.818, 95% CI: 0.742-0.894) and higher than the radiomics model alone (AUC = 0.764, 95% CI: 0.682-0.845). Conclusions: A combined model incorporating T2-FLAIR WMH radiomics and key clinical variables provides a feasible tool for individualized prediction of HT after EVT in patients with AIS, although its incremental discriminatory value over the clinical model alone was limited in the external validation cohort. This approach may help improve perioperative risk stratification and support personalized clinical decision-making.

Background: Acute kidney injury (AKI) is a common, life-threatening complication among hospitalized patients. Timely prediction is crucial for prevention, while accurate prognosis following onset is essential for guiding care. Although electronic health records (EHRs) provide rich longitudinal data, their high dimensionality, heterogeneity, and temporal complexity pose challenges for traditional machine learning. Objective: The aim of this study was to develop and evaluate a transformer-based model for generalizable AKI prediction and prognosis by modeling long-range temporal dependencies and complex relationships in heterogeneous EHR data and to examine its interpretability. Methods: We developed the Biomarker-Aware Transformer for AKI (BAT-AKI), which integrates multidimensional embeddings capturing temporal dynamics, textual descriptions, hierarchical ontology structures, and semantic relationships in EHRs. To enhance sensitivity to AKI-related biomarker fluctuations, BAT-AKI combines masked language modeling with a novel biomarker abnormality-aware pretraining task. The model was pretrained and fine-tuned for 4 tasks: AKI onset, early reversal, recovery at discharge, and 90-day post-AKI mortality. Pretraining and internal validation used data from the National Patient-Centered Clinical Research Network (PCORnet) site, with external validation across 3 additional sites. Results: BAT-AKI achieved the strongest overall discrimination, outperforming baseline models in area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) on most tasks. For AKI onset prediction, it achieved an AUROC of 0.8549 (95% CI 0.8433-0.8658) and an AUPRC of 0.6041 (95% CI 0.5786-0.6292), exceeding the second-best baseline by 2.8% in AUROC and 8.1% in AUPRC. Across prognostic tasks, BAT-AKI achieved AUROCs of 0.6006 (95% CI 0.5703-0.6318), 0.6080 (95% CI 0.5745-0.6397), and 0.8017 (95% CI 0.7708-0.8311) and AUPRCs of 0.7036 (95% CI 0.6690-0.7388), 0.7417 (95% CI 0.7082-0.7765), and 0.5063 (95% CI 0.4422-0.5684) for early reversal, recovery at discharge, and 90-day post-AKI mortality, respectively, corresponding to relative gains of 2.4% to 6% in AUROC and 3.7% to 4.1% in AUPRC over the second-best baseline. Performance remained stable across external cohorts, with AKI onset AUROCs ranging from 0.7754 (95% CI 0.7677-0.7827) to 0.7989 (95% CI 0.7885-0.8092) and AUPRCs ranging from 0.4777 (95% CI 0.4631-0.4924) to 0.5687 (95% CI 0.5495-0.5871), corresponding to relative improvements of 3.7% to 5.3% in AUROC and 16.8% to 25.3% in AUPRC over the transformer baseline. In a low-resource setting using only 2% of the fine-tuning data, BAT-AKI achieved an AUROC of 0.7898 (95% CI 0.7759-0.8036) and an AUPRC of 0.4903 (95% CI 0.4640-0.5178) for AKI onset prediction, exceeding the second-best baseline by 4.2% in AUROC and 10.9% in AUPRC. Visualization of embeddings, attention patterns, and integrated-gradient feature attributions highlighted clinically meaningful relationships and patient-specific trajectories. Conclusions: BAT-AKI provides generalizable AKI prediction and prognosis with consistent discrimination across institutions, while offering interpretability insights into its predictions, supporting its potential for early warning and AKI care optimization.

Background: Left ventricular dysfunction (LVD) is common and associated with adverse prognosis, even in asymptomatic individuals. Early treatment improves outcomes, underscoring the need for efficient strategies to identify patients who may benefit from echocardiographic evaluation. This study aimed to develop and externally validate an AI auscultation method to detect LVD among patients referred for clinically indicated echocardiography. Methods: In this prospective, multicenter study, adult patients referred for echocardiography were enrolled to undergo digital auscultation at four standard auscultation sites with simultaneous single-lead ECGs. LVD was defined as a left ventricular ejection fraction (LVEF) less than or equal to 40% on concurrent echocardiography. An ensemble model combining a transformer-based network and a one-dimensional convolutional neural network was trained on multicenter data and externally validated. The model's performance was compared with a seven-cardiologist reader substudy and clinical logistic regression models incorporating clinical variables, single-lead ECG-derived features, and B-type natriuretic peptide levels. Results: The multicenter training cohort included 2359 patients (mean age, 69.7±13.5 years; 40.4% female; 8.9% with reduced LVEF). The external validation cohort of 964 patients (mean age, 68.6±15.4 years; 38.9% female; 7.6% with reduced LVEF) was obtained from another independent multicenter registry. The four-site AI auscultation outperformed ECG-only models and the clinical logistic models, achieving an area under the receiver operator curve (AUROC) of 0.926 (95% confidence interval [CI], 0.911 to 0.939) in the training cohort. In the external validation cohort, AUROC was 0.904 (95% CI, 0.876 to 0.930) and sensitivity, specificity, and accuracy were 86.3%, 74.2%, and 75.1%, respectively. In a 100-patient reader substudy, AI auscultation outperformed cardiologist consensus in accuracy (82.0% vs. 62.0%; P=0.004). Model performance was consistent across prespecified subgroups. The simplified two-site auscultation model retained high accuracy in the external validation cohort (AUROC, 0.903; 95% CI, 0.878 to 0.927). Conclusions: AI auscultation accurately detected LVD among patients referred for clinically indicated echocardiography, supporting its potential as a scalable and accessible pre-echocardiographic risk-enrichment tool. Further validation is warranted before broader screening applications. (Funded by AMI Inc. and the Japan Society for the Promotion of Science KAKENHI program [grant number 25K19371].).

Background: Accurate detection and staging of osteonecrosis of the femoral head (ONFH) are essential for effective treatment yet frequently require specialist expertise and advanced equipment. To address this requirement, we aimed to develop a deep learning-based diagnostic model for the automatic detection and staging of ONFH using x-ray images, and compared its performance with clinicians to allow them to improve their diagnostic performance. Methods: In this retrospective, large-scale, multicentre study, we developed our deep learning model, Osteonecrosis Artificial Intelligence Diagnostician (OsteoAID), which was based on the You Only Look Once version 8 (YOLOv8) detection network and trained on a multicentre x-ray dataset. The dataset, dated from Jan 1, 2015, to Dec 31, 2023, and obtained from 16 399 patients across six Chinese tertiary hospitals, included 20 722 anteroposterior hip x-ray images of 28 785 hips, including hips with ONFH, healthy hips, and hips with other conditions radiographically similar to ONFH. To establish a rigorous ground truth for model training and evaluation, expert orthopaedists annotated each image to differentiate ONFH from confounding diseases and additionally classified the confirmed ONFH lesions on the basis of the Ficat and Arlet staging system. To verify the diagnostic performance of OsteoAID, we compared the model's accuracy, sensitivity, specificity, F1 score, and area under the curve (AUC) with those of the original YOLOv8 model on the internal testing set for both the staging and detection tasks. Furthermore, the performance of OsteoAID was compared with the performance of radiologists and orthopaedists on two external testing sets, and its diagnostic performance was also benchmarked against MRI-the gold standard for ONFH diagnosis-with a third external testing set in which x-ray images were staged by corresponding MRI scans. Diagnostic metrics were compared with those of clinicians using t tests and McNemar's tests, with significance set at p<0·05. Findings: On the internal testing set, OsteoAID improved the staging accuracy from 87·42% (95% CI 85·88-89·20) to 96·44% (94·73-97·17) and the AUC from 88·54% (86·74-90·37) to 96·94% (95·75-97·98) compared with the original YOLOv8 model. For detection the model had a mean average precision of 92·97% (91·73-94·43) at an intersection-over-union threshold of 50%. In comparisons with clinicians on the external testing set I, the model showed staging accuracy of 96·60% (94·27-97·81), which was significantly higher than radiologists (74·96% [73·38-76·85]; p<0·0001) and junior orthopaedists (83·98% [81·27-84·06]; p<0·0001) and similar to that of expert orthopaedists (96·54% [94·18-97·79]; p=0·20). Furthermore, model-assisted diagnostics improved the accuracy of junior orthopaedists by an absolute difference of 12·70% and of radiologists by 21·47%. Although the accuracy improvement for expert orthopaedists was non-significant (an absolute difference of 0·74%; p=0·080), the diagnostic time was substantially reduced across all groups. On external testing set II (which included challenging cases: low-quality x-ray images affected by improper exposure, poor contrast, incomplete field of view, or foreign-object shadows, as well as smartphone photographs of x-ray films), the model showed robust performance even under extreme conditions, achieving an overall staging accuracy of 93·76% (93·09-95·82). Additionally, combining x-ray imaging with the deep learning model produced diagnostic results similar to MRI on external testing set III with an overall accuracy of 96·63% (95·68-97·86) and an AUC of 97·33% (96·65-98·75). Interpretation: The developed deep learning model shows expert-level diagnostic performance in the dual tasks of detecting and staging ONFH by use of anteroposterior x-rays, which can substantially enhance both the diagnostic proficiency and efficiency of clinicians. Consequently, the model could serve as a highly accessible decision-support tool to optimise clinical triage and guide appropriate treatment strategies. Funding: National Natural Science Foundation of China, National High Level Hospital Clinical Research Funding, and Beijing Natural Science Foundation.

Background: Primary human papillomavirus (HPV) screening has enabled scalable molecular detection in cervical cancer prevention, but downstream triage of HPV-positive individuals remains dependent on manual p16/Ki-67 dual-stain (DS) cytology interpretation. Although DS is guideline-endorsed and biologically well-defined, scalable implementation is limited by reader-dependent variability and lack of standardized automation. Methods: We developed Cytoreader-Fusion, a transparent, ensemble-based AI system for automated slide-level DS classification within HPV-based screening workflows that produces well-calibrated probabilistic outputs. The model was trained on 1919 cases and evaluated under blinded honest-broker oversight on 9006 cases across four independent cohorts, including a fully independent post-lock external validation cohort (EXVC; n=4996). All preprocessing steps, aggregation rules, and decision parameters were prespecified and irreversibly locked prior to validation. Performance was evaluated against independently ascertained histologically confirmed cervical intraepithelial neoplasia grade 3 or worse (CIN 3+) (or anal intraepithelial neoplasia grade 3 or worse [AIN 3+]) end points rather than cytologic concordance. A single unified classifier was applied across both ThinPrep and SurePath preparations without platform-specific retraining. Results: Cytoreader-Fusion demonstrated stable and reproducible slide-level classification across imaging variability, preparation platforms, and clinical settings. Across blinded screening and referral cohorts, sensitivity and specificity were higher than atypical squamous cells of undetermined significance or worse (ASC-US+) Papanicolaou (Pap) cytology in HPV-based screening populations and comparable to or exceeded manual DS interpretation. In the fully independent post-lock cohort (n=4996), the locked system achieved a sensitivity of 88.3% (90% UI, 85.3%-91.3%) and a specificity of 54.4% (90% UI, 52.9%-55.9%) for detection of CIN 3+, without any model retraining, threshold tuning, or configuration adjustment after lock. Operating characteristics remained consistent across cohorts and preparation platforms. Conclusions: This study establishes a high-stringency validation framework for clinically deployable artificial intelligence in population-based cancer screening. By combining prespecified algorithm locking, independent blinded evaluation under honest-broker supervision, end point-based validation against histologically confirmed disease, and cross-platform generalization, Cytoreader-Fusion demonstrates robust and scalable automation of DS triage within HPV-based screening programs. (Funded by the U.S. National Cancer Institute and others.).

Aims: Accurate prediction of the clinical course in implantable cardioverter-defibrillator (ICD) and cardiac resynchronization therapy-defibrillator (CRT-D) patients is crucial, as their disease progression varies significantly due to comorbidities and individual characteristics. Methods: This retrospective study included 3321 patients from four hospitals in China, with 3207 for model development and internal validation and 114 for external testing. Five machine learning (ML) models were evaluated using the C-index, time-dependent AUROC, and Brier score. Decision curve analysis (DCA) assessed clinical utility, while Kaplan-Meier curves stratified patients into high-, intermediate-, and low-risk groups. SHAP analysis identified key predictors and explained model outputs. Results: The Random Survival Forest (RSF) model demonstrated favorable numerical performance across multiple evaluation metrics, achieving the highest time-dependent AUROC (0.74, 0.74) and C-index (0.71, 0.70) in internal and external test sets. It also had the lowest time-dependent Brier score (0.17) in the internal test and ranked second (0.13) in the external test. DCA demonstrated numerically higher net benefit for the RSF model across evaluated threshold probabilities. Kaplan-Meier curves revealed distinct survival trajectories, with high-risk patients showing significantly higher mortality than low-risk patients (HR: 4.50, 95% CI: 3.10-6.55, p < 0.0001), while intermediate-risk patients also had increased mortality (HR: 2.49, 95% CI: 1.69-3.68, p < 0.0001). LVEDD was the most influential predictor of mortality. Conclusions: The RSF survival model enables individualized, time-resolved prediction of all-cause mortality risk in ICD and CRT-D patients, supporting risk stratification and future evaluation of clinical utility, while requiring cautious interpretation of longer-term predictions due to limited numbers at risk.