Background: Early differentiation of pulmonary aspergillosis (PA) from bacterial pneumonia (BP) remains challenging in respiratory practice because of overlapping clinical and computed tomography (CT) features. Conventional microbiological and serological tests may be invasive, delayed, or insufficiently sensitive in selected patients. This study aimed to develop and externally validate CT-based deep learning (DL) models as adjunctive tools for differentiating PA from BP.
Methods: This multicenter retrospective study included patients diagnosed with PA or BP at The First Affiliated Hospital of Guangzhou Medical University and The Second Affiliated Hospital of Guangzhou Medical University. CT images from The First Affiliated Hospital of Guangzhou Medical University were used for model development and internal validation, while data from The Second Affiliated Hospital of Guangzhou Medical University were used for external validation. Two CT-based DL models were developed: an unannotated supervised training (UST) model using CT images without manual lesion annotation and an annotated supervised training (AST) model incorporating physician-annotated lesions. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). A standardized image-only questionnaire was also used to explore clinician performance in differentiating PA from BP.
Results: A total of 387 PA cases and 311 BP cases were included. In internal validation, the AST model achieved an AUC of 0.83 [95% confidence interval (CI): 0.76-0.90], with 81.7% sensitivity, 73.2% specificity, and 78.0% accuracy, whereas the UST model achieved an AUC of 0.78 (95% CI: 0.69-0.86). In external validation, the AST model achieved an AUC of 0.81 (95% CI: 0.71-0.92), compared with 0.56 (95% CI: 0.41-0.70) for the UST model. In the standardized image-only clinician assessment, 158 valid responses were analyzed, with an average diagnostic accuracy of 68.2%. Exploratory analysis suggested that model performance varied across disease stages, highlighting the potential influence of temporal radiological changes.
Conclusions: The annotated CT-based DL model showed potential as an adjunctive imaging tool for differentiating PA from BP. However, the external sensitivity of 60.6% indicates that the model should not be used as a stand-alone diagnostic or rule-out tool. Prospective multicenter validation and integration of clinical and microbiological data are required before clinical implementation.
Li C, Liu B, Zhang J et al. · Front Med (Lausanne)
External validation 98%
Abstract
Background: Intraoperative hypothermia (IOH, temperature <36.0 °C) is a common complication in gynecological laparoscopic surgery that leads to significant adverse outcomes that would benefit from preoperative risk prediction.
Methods: We developed and compared two predictive models-a conventional multivariable logistic regression (LR) model and a gradient boosting machine (XGBoost) model-to preoperatively predict IOH based on 20 objectively measurable preoperative clinical variables that could be directly extracted from electronic health records (EHRs). The derivation cohort consisted of 460 cases from a tertiary care center in 2024, which were randomly divided into a training set (n = 322) and a test set (n = 138) at a 7:3 ratio. For external validation, we used an independent cohort of 183 cases from a second medical institution during the same period. Model performance was evaluated using metrics including the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC), with SHapley Additive exPlanations (SHAP) method applied for feature importance analysis.
Results: We initially analyzed 20 preoperative variables. The final model incorporated eight key indicators: body mass index (BMI), ASA physical status, fasting period, basal body temperature, estimated surgical duration, hypertension status, serum albumin level, and alanine aminotransferase (ALT). The derivation cohort consisted of 460 patients, split into a training set (n = 322) and a test set (n = 138). An independent external validation cohort included 183 patients. By AUROC comparison, the XGBoost model outperformed the LR model in both the training set (0.933 vs. 0.891) and the external validation set (0.870 vs. 0.838), while performing slightly lower in the test set (0.821 vs. 0.865).
Conclusions: We developed and validated an interpretable XGBoost model using eight readily available preoperative clinical indicators. The model demonstrated good performance for the preoperative prediction of IOH in gynecological laparoscopic surgery, which may facilitate early identification of high-risk patients and timely intervention.
Background: Distinct treatment strategies for small-cell lung cancer (SCLC) and non-small-cell lung cancer (NSCLC) brain metastases require accurate noninvasive subtyping.
Purpose: To develop and externally validate a biparametric MRI-based 2.5D multi-instance learning (MIL) deep learning (DL) model for discriminating SCLC from NSCLC brain metastases, with performance compared against clinical, radiomics, and 3D-DL models.
Study Type: Retrospective.
Population: This study included 366 patients from four institutions. Center 1 provided the training set (n = 199; mean age, 63.88 ± 10.79 years; 72.36% male) and an internal tuning set (n = 85; mean age, 64.73 ± 9.53 years; 67.06% male). Centers 2-4 were pooled into a single external test set (n = 82; mean age, 62.26 ± 10.60 years; 63.41% male).
Field Strength/Sequence: 1.5 T and 3.0 T; contrast-enhanced T1-weighted imaging and fluid-attenuated inversion-recovery imaging.
Assessment: The 2.5D-MIL model was benchmarked against clinical, radiomics, and 3D-DL models and evaluated on internal tuning and pooled external test sets.
Statistical Tests: Shapiro-Wilk, Student's t-test, Mann-Whitney U, chi-square, univariable and multivariable logistic regression, and DeLong tests were used. Performance measures included the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals and classification and calibration metrics. Statistical significance was defined as two-sided p < 0.05.
Results: In the internal tuning set, AUCs were 0.842 (2.5D-MIL), 0.740 (clinical), 0.710 (radiomics), and 0.747 (3D-DL). In the external test set, AUCs were 0.826 (2.5D-MIL), 0.826 (clinical), 0.696 (radiomics), and 0.664 (3D-DL). Differences between 2.5D-MIL and the clinical (p = 1.000), radiomics (p = 0.139), and 3D-DL (p = 0.119) models were not significant.
Data Conclusion: The MRI-based 2.5D-MIL model showed potential for noninvasive differentiation of SCLC from NSCLC brain metastases. Larger prospective studies are needed before clinical use.
Evidence Level: 3.
Technical Efficacy: Stage 2.
Rationale And Objectives: Extracellular volume (ECV) quantified by cardiac magnetic resonance (CMR) is a key indicator of myocardial fibrosis and strongly linked to outcomes after ST-segment elevation myocardial infarction (STEMI). This research focused on the development and validation of an interpretable machine learning (ML) model that incorporates ECV to predict 1-year major adverse cardiac events (MACE) in patients with STEMI following percutaneous coronary intervention.
Materials And Methods: MACE prediction models were developed using clinical variables combined with CMR-ECV parameters and validated in a prospective cohort. Feature selection was conducted through Boruta algorithms and least absolute shrinkage and selection operator algorithms, and eight ML algorithms were trained. Model performance was assessed by discrimination and calibration. The top performer was interpreted by Shapley Additive exPlanations (SHAP) analysis. An interactive online nomogram was deployed to facilitate personalized perioperative risk management.
Results: Among 673 STEMI patients (derivation: n = 398; prospective validation: n = 275), five key predictors were identified: N-terminal pro-B-type natriuretic peptide (NT-proBNP), left ventricular ejection fraction (LVEF), myocardial infarction ECV (MI-ECV), age, and creatine kinase-MB. The random forest model achieved the highest discriminative performance (area under the receiver operating characteristic curve 0.825, 95% confidence interval: 0.763-0.876). SHAP analysis identified NT-proBNP as the most impactful feature, with LVEF class and MI-ECV ranking next.
Conclusion: Integration of CMR-derived ECV into an interpretable machine learning framework enables accurate 1-year MACE risk prediction following STEMI. This approach combines robust discriminative performance with clinical transparency through SHAP-based explanations, offering a proof-of-concept tool for precision cardiovascular care.
Background: Early detection of gastric cancer (GC) is limited by invasive endoscopy and poor sensitivity of blood-based biomarkers. Red blood cells (RBCs), the most abundant cells in the circulation, are anucleate but can retain DNA remnants that have been associated with pathological stress. Here, we report a non-invasive strategy for GC detection based on genome-wide profiling of DNA remnants retained in RBCs (rbcDNA).
Methods: We performed shallow whole-genome sequencing of rbcDNA from 1-2 mL of peripheral blood of 865 individuals, including 434 treatment-naive GC patients and 431 non-GC controls with benign gastric conditions. In the discovery cohort, we identified 1,000 reproducible rbcDNA read-depth features distinguishing GC from non-cancer states. A machine learning classifier based on dimensionally reduced GC-associated rbcDNA features was trained in the discovery cohort (n = 435), with the cutoff fixed at 90% specificity, and evaluated in a test cohort (n = 109) and three independent multicenter validation cohorts (n = 321).
Findings: The classifier achieved the areas under the receiver operating characteristic curve (AUCs) of 84% in nested cross-validation of the discovery cohort and 95% in the test cohort. Across three hospital-based multicenter validation cohorts, performance remained consistent, with 83%-90% overall sensitivity and 82%-89% sensitivity for stage I GC. Additionally, the classifier distinguished early-stage GC from precancerous lesions and other tumors and detected ∼80% of GC cases negative for tumor markers.
Conclusions: These findings demonstrate the potential of rbcDNA as a non-invasive biomarker for early-stage GC detection.
Funding: This work was supported by the Key Research and Development Program of Zhejiang Province (2026C02A1176 and 2024C03170).
Rationale And Objectives: Computed tomography radiomics has shown potential for predicting acute pancreatitis (AP) severity, but conventional workflows depend on labor-intensive expert pancreatic segmentation. This study evaluated a no-new-Net (nnU-Net)-based automated segmentation-radiomics workflow against an expert manual-segmentation workflow, using AP severity classification as the downstream task, and assessed the automated workflow externally.
Materials And Methods: This retrospective multicenter study included patients with AP from Center A (n = 214), Center B (n = 56), and Center C (n = 42). Expert manual pancreatic masks served as reference annotations for nnU-Net training and as the comparator for the radiomics workflow. After selecting the optimal nnU-Net configuration, radiomics features were extracted separately from manual and automated regions of interest, and AP severity classifiers were developed using the same analytical framework. Discrimination between the two workflows was compared in the independent internal test cohort using the area under the curve (AUC) and the paired DeLong test, with Benjamini-Hochberg false discovery rate (FDR) and Bonferroni corrections applied to P values.
Results: In the internal test cohort, the Light Gradient Boosting Machine classifier achieved AUCs of 0.799 (95% confidence interval (CI), 0.683-0.915) and 0.733 (95% CI, 0.593-0.873) with manual and automated segmentations, respectively. The absolute AUC difference was 0.066, and the paired DeLong comparison was not statistically significant after FDR correction (FDR-adjusted p = 0.360; Z = 1.1755, unadjusted p = 0.2398). The automated workflow achieved AUCs of 0.784 (95% CI, 0.653-0.915) in Center B and 0.715 (95% CI, 0.559-0.871) in Center C. The selected 2D nnU-Net achieved a mean Dice coefficient of 0.798 during five-fold cross-validation and 0.726 in the independent internal test cohort.
Conclusion: The nnU-Net-based automated segmentation-radiomics workflow showed no significant difference in discrimination compared with the expert manual workflow in the internal test cohort. It also maintained moderate discrimination across two external cohorts. These findings support the feasibility of reducing dependence on manual pancreatic segmentation, although the present analysis does not establish formal equivalence or interchangeability.
Background: Hepatocellular carcinoma (HCC) is one of the most common and lethal malignancies, whose prognostic prediction is particularly challenging. Histopathological biomarkers are critical for cancer prognosis assessment. Advancements in artificial intelligence (AI) have made it possible to automate the classification and detection of information from whole-slide pathology images using deep learning. This study aims to utilize AI deep learning to analyze histopathological slides of HCC patients for prognostic stratification, thereby assisting in the development of clinical treatment plans.
Methods: The Otsu thresholding method was employed to automatically segment whole-slide images (WSIs) from two HCC datasets, The Cancer Genome Atlas (TCGA) and HCC slides from Ansteel General Hospital. The pathological features of WSIs, extracted by the pretrained CONCH model were utilized as model inputs. Model performance was assessed using the area under the receiver operating characteristic curve and additional metrics. Then Attention Heatmap Visualization was employed to analyze interpretability. The Kolmogorov-Smirnov curve was utilized to stratify patients into distinct risk groups. Survival analysis was performed to evaluate differences in overall survival between distinct risk groups. Univariate and multivariate Cox regression analyses were used to identify key clinical indicators and pathological features as significant predictors. The riskRegression package generated calibration curves comparing predicted versus observed survival outcomes. Calibration curves were quantify the deviation between model-predicted probabilities and actual observed outcomes.
Results: A prognostic prediction model was developed using the CLAM self-attention framework based on TCGA dataset, whose generalization capability was verified by histologically stained slide data from a proprietary dataset. The results showed that the model could achieve better migration. Mortality rates demonstrated a progressive increase across ascending risk groups. The risk stratification system possesses potential clinical utility for guiding treatment selection. Forest plots generated demonstrated that the risk stratification system functions as an independent prognostic predictor for HCC. Calibration curve analysis confirmed the predictive accuracy of the model, and this finding from Decision Curve Analysis underscores its superior utility in guiding clinical decision-making.
Conclusion: This study developed a prognostic prediction model, whose risk stratification could be an independent prognostic predictor for HCC and a guide for therapeutic decision-making.
Background: Transthyretin amyloid cardiomyopathy (ATTR-CM) is an important cause of heart failure (HF) with increased left ventricular (LV) wall thickness that can be challenging to distinguish from other LV hypertrophic processes. Echocardiographic myocardial radiomic texture analysis may identify imperceptible amyloid-induced structural changes.
Objectives: The study aims to develop and validate an echocardiography-based machine learning model using myocardial radiomics to identify ATTR-CM.
Methods: In this retrospective and prospective multicenter study, 854 echocardiographic scans were analyzed. Conventional and novel radiomic features were extracted from parasternal short-axis views at basal, mid, and apical LV levels, with incorporated age and sex from echocardiogram metadata. A machine learning pipeline was developed using a balanced training set (150 ATTR-CM, 150 nonamyloid HF). Performance was evaluated on an independent internal validation set (50 ATTR-CM, 399 nonamyloid HF) and an external validation set (10 ATTR-CM, 10 immunoglobulin light-chain-amyloid, 10 hypertrophic cardiomyopathy, 75 nonamyloid HF).
Results: In internal validation, the model achieved 86% sensitivity (95% CI: 73.8%-93%), 92% specificity (95% CI: 88.9%-94.3%), 57.3% positive predictive value (95% CI: 46.1%-67.9%), 98.1% negative predictive value (NPV) (95% CI: 96.2%-99.1%), and receiver operating characteristic-area under the curve 0.938 (95% CI: 0.891-0.985). Performance was consistent in the external validation (100% sensitivity, 78.9% specificity, 33.3% positive predictive value, 100% NPV, and receiver operating characteristic-area under the curve of 0.917). The model correctly classified 77.3% of nonamyloid HF, 70% of AL-amyloid cases, and 100% of hypertrophic cardiomyopathy cases as non-ATTR in subgroup analysis.
Conclusions: We developed an echocardiographic radiomics-based tool to identify ATTR-CM with excellent discrimination and very high NPV in a cohort that approximates clinical experience. These preliminary findings require validation in larger, more diverse cohorts.
Jongs N, Heerspink HJL, Tangri N et al. · Kidney360
External validation 93%Prospective 61%
Abstract
Background: Early identification and treatment of high-risk CKD can lower the risk of kidney failure and cardiovascular events. A laboratory-based machine-learning model (Klinrisk) was developed to predict the risk of CKD progression in adults with, or at risk for, CKD. It has shown good accuracy in the general population and in clinical trial populations with diabetic kidney disease, but its performance in non-diabetic CKD is unclear. We evaluated Klinrisk in a clinical trial of the sodium glucose cotransporter 2 inhibitor dapagliflozin in patients with CKD, with and without type 2 diabetes.
Methods: We externally validated Klinrisk in the DAPA-CKD trial population for prediction of CKD progression (first occurrence of sustained ≥40% eGFR decline or kidney failure) up to 3 years. Performance was assessed with the area under the receiver operating characteristic curve (AUC), Brier scores and calibration plots, and compared with Kidney Disease Improving Global Outcomes (KDIGO) heat-map categories based on eGFR and urine albumin-to-creatinine ratio. Consistency of the dapagliflozin effect on CKD progression across Klinrisk risk categories was also examined.
Results: The analysis included 4,304 DAPA-CKD participants, of whom 578 experienced CKD progression events over a median follow-up of 2.4 years. Klinrisk achieved an AUC of 0.79 (95% CI 0.77-0.81) and a Brier score of 0.10 (95% CI 0.09-0.11). It outperformed KDIGO heat-map categories in subgroups of patients with and without type 2 diabetes, and in those with IgA nephropathy. Dapagliflozin reduced CKD progression compared with placebo, irrespective of baseline risk.
Conclusions: In DAPA-CKD, Klinrisk accurately predicted CKD progression and outperformed KDIGO-based risk stratification in patients with and without type 2 diabetes. Eight of the 22 model inputs were unavailable and supplied from the development model rather than measured. Dapagliflozin reduced CKD progression across the full spectrum of predicted risk compared with placebo.
Purpose: To evaluate whether pretreatment CT radiomics adds information for overall survival (OS) prediction among patients with hepatocellular carcinoma (HCC) treated with transarterial chemoembolization (TACE), and assess frozen imaging-model performance in an independent cohort.
Methods: This retrospective study included 221 WAW-TACE and 98 HCC-TACE-SEG patients with pretreatment arterial and portal venous CT. Incremental value of a tumor-radiomics Random Survival Forest score (R1) beyond cohort-specific clinical Cox models (C0) was evaluated within each cohort using 5-fold cross-validation repeated 10 times. External validation was restricted to frozen WAW-trained R1 models (Top-K = 5, 10, 15) applied to HCC-TACE-SEG without refitting; the combined C0 + R1 model was not externally validated.
Results: WAW-TACE included 162 deaths and HCC-TACE-SEG 86 deaths. In WAW-TACE, C0, R1 Top-K = 15, and C0 + R1 achieved C-indices of 0.646 ± 0.043, 0.593 ± 0.049, and 0.639 ± 0.045, respectively. Corresponding HCC-TACE-SEG estimates were 0.496 ± 0.087, 0.530 ± 0.072, and 0.495 ± 0.094. Thus, R1 did not consistently improve cohort-specific clinical reference models. In frozen external application, Top-K = 5, 10, and 15 achieved C-indices of 0.463 (95% CI 0.397-0.532), 0.535 (0.468-0.608), and 0.528 (0.460-0.598), respectively.
Conclusion: The evaluated models showed no consistent incremental value for OS prediction within these TACE cohorts and limited frozen-model discrimination in the clinically distinct HCC-TACE-SEG cohort. These findings do not support clinical use of these models or establish a general limitation of CT radiomics.
Zhou Z, Wan Y, Cheng Y et al. · Front Cardiovasc Med
External validation 91%
Abstract
Background: Insulin resistance surrogate markers, including the triglyceride-glucose (TyG) index, TyG-BMI, and TG/HDL-C ratio, have been associated with aortic stiffness in high-risk populations. Their comparative associations and incremental discrimination in asymptomatic adults remain unclear. We therefore compared these markers and evaluated a simplified model for cfPWV-defined aortic stiffness.
Methods: We conducted a cross-sectional study of 654 asymptomatic adults aged ≥18 years who underwent routine health examinations at two tertiary medical centers. After exclusions, 457 participants formed the modeling cohort and 101 formed the external validation cohort. Aortic stiffness was defined as cfPWV ≥10 m/s. We used multivariable logistic regression to estimate associations between the three surrogate markers and aortic stiffness. Restricted cubic spline regression assessed potential nonlinear associations. Incremental discrimination was evaluated using ΔAUC, NRI, and IDI. Six machine-learning models were compared, and the most interpretable model was evaluated in an independent institutional validation cohort.
Results: Aortic stiffness was present in 32.8% (150/457) of the modeling cohort. After full adjustment, the TyG index was independently associated with aortic stiffness (per-SD OR = 3.44, P = 0.002), as were TG/HDL-C (OR = 3.11, P = 0.039) and, more weakly, TyG-BMI (OR = 1.52, P = 0.043). The model including TyG showed higher discrimination (ΔAUC = 0.014, DeLong P = 0.025) and reclassification (continuous NRI = 0.692 and IDI = 0.041; both P < 0.001) than the base model. A data-derived TyG cutoff of 8.68 was identified as a potential marker for exploratory risk stratification. The five-variable model, comprising age, MAP, diabetes mellitus, smoking, and TyG, yielded AUCs of 0.931 in the modeling cohort and 0.909 in the external validation cohort. Subgroup and sensitivity analyses showed similar patterns.
Conclusion: Among the three surrogate markers, the TyG index showed the strongest adjusted association with aortic stiffness and the greatest incremental discrimination beyond conventional clinical variables. The five-variable model showed good discrimination in an independent institutional validation cohort. However, these cross-sectional findings support only potential risk-stratification value and do not establish temporal sequence, causality, or clinical utility; prospective studies are required to assess broader generalizability.