2026-11-06

In tuberculosis-endemic settings, suspected mycobacterial disease may reflect either nontuberculous mycobacteria (NTM) or Mycobacterium tuberculosis (MTB), yet routine clinical findings do not reliably distinguish them before pathogen-specific confirmation. We developed an eight-variable Extra Trees model in Hospital A (n = 1,090; 130 NTM) and independently evaluated the frozen model in Hospital B (n = 1,160; 128 NTM). Internal out-of-fold (OOF) AUC/AUPRC were 0.892/0.588; external AUC/AUPRC were 0.849/0.505. Calibrated Brier scores were 0.072 and 0.075. At the Hospital A-derived threshold of 0.13179, sensitivity/specificity were 80.0%/80.3% internally and 67.2%/83.7% externally; 42 of 128 external NTM cases were classified below threshold. Because feature selection preceded the original OOF evaluation, a post-hoc fully nested analysis was performed (AUC 0.875; AUPRC 0.525), indicating modest internal optimism. A post-hoc benchmark did not show an overall advantage for Extra Trees over other algorithms. Routine variables may help prioritize confirmatory testing, but the frozen threshold cannot support stand-alone diagnosis or exclusion of NTM. Prospective evaluation must verify patient-level measurement timing, reference diagnoses and clinical impact.

Background: Women with polycystic ovary syndrome (PCOS) who achieve clinical pregnancy following fresh in vitro fertilization (IVF)/intracytoplasmic sperm injection (ICSI) carry an elevated miscarriage risk. However, individualized risk stratification after clinical pregnancy confirmation but before outcome determination remains challenging in clinical practice. Objective: This study aimed to develop, externally and prospectively validate over time, and interpret a machine learning-based risk stratification model for women with PCOS who had achieved a confirmed clinical pregnancy after fresh IVF/ICSI. Methods: This multicenter study included 8911 clinical pregnancy IVF/ICSI cycles across development (n=7241; miscarriages: n=973), external validation (n=1170; miscarriages: n=146), and prospective temporal validation (n=500; miscarriages: n=62) cohorts. Prespecified predictors contained no observed missing values; training-fold median imputation was kept only as a leakage-safe fallback. A total of 13 candidate machine learning and traditional statistical algorithms were compared internally using repeated stratified 5-fold cross-validation (5 repeats) within the development cohort; preprocessing and hyperparameter tuning were performed exclusively in training folds. A prelocked Extreme Gradient Boosting (XGBoost) model was carried forward unchanged for external and prospective temporal validation. Cross-fitted logistic recalibration of averaged XGBoost out-of-fold probabilities was evaluated as a secondary development-sample diagnostic analysis. Model performance encompassed discrimination, precision-recall, calibration, threshold-based metrics, decision curve analysis, risk stratification, and Shapley additive explanations-based interpretability. Results: Miscarriage rates were 13.4% (973/7241), 12.5% (146/1170), and 12.4% (62/500) for the development, external validation, and prospective temporal validation cohorts, respectively. In repeated cross-validation, XGBoost (area under the receiver operating characteristic curve [AUC] 0.885, 95% CI 0.874-0.896) and LightGBM (AUC 0.882, 95% CI 0.871-0.893) showed comparable discrimination; the paired XGBoost-LightGBM DeLong test was not significant (Holm-adjusted P=.07). Raw XGBoost out-of-fold probabilities overestimated absolute risk (mean predicted risk 30.3% vs observed risk 13.4%); cross-fitted recalibration within the development cohort improved calibration but was not propagated to validation cohorts. The locked XGBoost model achieved AUCs of 0.787 (95% CI 0.748-0.827; Brier score 0.099) in external validation and 0.763 (95% CI 0.709-0.816; Brier score 0.104) in prospective temporal validation, showing cross-cohort performance attenuation. Risk-quartile stratification using fixed development-cohort cutoffs showed consistent event enrichment across validation cohorts, but positive predictive values remained modest. Shapley additive explanations analysis identified anti-Müllerian hormone, endometrial thickness on the day of human chorionic gonadotropin administration, and basal luteinizing hormone as leading predictors. Conclusions: The locked XGBoost model using routine preoutcome variables may support miscarriage risk stratification and counseling for women with PCOS with clinical pregnancy after fresh IVF/ICSI. Given cross-cohort discrimination loss and calibration drift, it should serve only as a risk-ranking tool, not for diagnosis or treatment selection. Local calibration monitoring is mandatory before any real-world clinical deployment.

Suboptimal clinical response (SCR) after sleeve gastrectomy (SG) remains a major clinical challenge, yet reliable preoperative risk stratification is lacking. Therefore, we developed and externally validated an interpretable machine-learning model for predicting SCR after SG using multicenter data from three tertiary hospitals in China. Overall, 1312 patients were included, comprising a derivation cohort (n = 1094) and two external validation cohorts (n = 84 and n = 134). SCR, defined as percentage of total weight loss < 25% at 1 year, occurred in 374 (34.2%), 34 (40.5%), and 45 (33.6%) patients, respectively. Among six machine-learning algorithms, XGB achieved the best performance. The final model incorporated eight routinely available predictors. In internal validation, the model showed excellent discrimination, with an AUC of 0.821 (95% confidence interval [CI], 0.752-0.871), sensitivity of 0.803, and specificity of 0.714. External validation confirmed satisfactory discrimination (AUC 0.761; 95% CI, 0.688-0.823), good calibration, and clinical benefit. Notably, the XGB model significantly outperformed 10 clinicians across experience levels and remained superior in clinically challenging cases. SHAP analysis enhanced interpretability, with predictor importance ranked as BMI, age, fasting glucose, type 2 diabetes, vitamin B12, calcium, total cholesterol, and gamma-glutamyl transferase. This interpretable XGB model may enable precise preoperative risk stratification and support personalized management for patients undergoing SG.

Background: AI-enabled electrocardiography (AI-ECG) can detect atrial fibrillation (AF), but safe deployment depends on calibration, subgroup reliability, and explicit error guarantees, not on discrimination alone. Objective: To test whether the conformal-prediction coverage guarantee of an AF ECG classifier survives external deployment shift, and whether a small local labeled audit restores it without retraining. Methods: A four-class 12-lead rhythm classifier trained on PTB-XL (Germany; 7,831 ECGs, 7,164 patients) was evaluated unchanged on Chinese (Chapman-Shaoxing; n=2,289) and US (Georgia; n=1,990) cohorts. We measured discrimination, expected calibration error (ECE), 90%-target split-conformal coverage, subgroup coverage, label-free shift corrections, local recalibration across audit sizes, and an AF auto-clear policy with finite-sample risk control. Results: AF AUC remained 0.998 (95% CI 0.997-0.999) in China but fell to 0.936 (0.919-0.952) in the US cohort. ECE rose from 0.051 to 0.197 and 0.237, and source-calibrated coverage fell from 0.901 to 0.758 and 0.722 (both P<0.001). Coverage was lower in Chinese men than women (0.707 vs 0.822; difference 0.115, 95% CI 0.081-0.148). Label-free corrections failed. Local recalibration restored mean coverage to 0.90 with 50 labeled ECGs but with wide variability (95% simulation interval 0.82-0.97), narrowing to 0.86-0.94 at 200 ECGs. Local AF risk control certified auto-clearing of 68% of held-out ECGs in China but abstained in the US cohort. Conclusion: AI-ECG uncertainty can fail silently despite high AUC. A local conformal audit detects when uncertainty estimates are untrustworthy and recalibrates the uncertainty layer without changing the classifier, converting trust into a measurable, subgroup-aware, task-specific guarantee.

Background And Objective: Postoperative atrial fibrillation (POAF) is a common surgical complication. We tested whether a two-stage within-patient architecture-graph attention followed by a SAINT-inspired feature Transformer-adds information beyond strong tabular models and transports across healthcare systems. Methods: We used the independent MIMIC-IV and INSPIRE databases. Sixty leakage-controlled scalar features were encoded as tokens for piecewise-linear encoding, graph attention, feature-wise Transformer blocks, and learned pooling. Reciprocal external validation was performed in both directions. Comparators were logistic regression, random forest, XGBoost, LightGBM, and a baseline-only stack. Component ablations, dense versus training-derived sparse graphs, paired bootstrap differences, and efficiency were evaluated. Results: In Experiment 1 (MIMIC-IV development, INSPIRE external), GAT-SAINT had an external AUC of 0.844 (95% CI, 0.829-0.858), versus 0.839 for logistic regression; the paired difference was 0.005 (95% CI, -0.001 to 0.011). In Experiment 2 (INSPIRE development, MIMIC-IV external), its AUC was 0.791 (95% CI, 0.768-0.813), versus 0.788 for random forest (paired difference, 0.003; 95% CI, -0.008 to 0.014) and 0.797 for the baseline-only stack. External Brier scores were 0.096 and 0.162. The Experiment 2 validation-locked threshold transported poorly (sensitivity, 0.025; specificity, 0.999). Component and graph-sensitivity effects were direction-dependent, and removal of some modules was compatible with little or no external AUC loss. Conclusions: Reciprocal external validation showed transportable rank discrimination but did not establish a consistent advantage over simpler tabular models or a transportable operating threshold. The architecture warrants further methodological study, but prospective validation, recalibration, and clinical-impact assessment are required before deployment.

Importance: Although external validation of machine learning (ML)-based delirium risk models is a key stage in the model development pipeline, few published reports exist. Moreover, ML-based delirium risk models commonly degrade in performance during external validation. Objective: To examine the performance of a fusion multimodal with natural language processing (NLP) delirium risk stratification model in live clinical practice with cross-hospital external validation. Design, Setting, And Participants: This quality improvement study evaluated live clinical deployment of the model among patients 60 years or older who were admitted to non-intensive care units at Mount Sinai Hospital (MSH), with external validation performed at Mount Sinai Morningside Hospital (MSM), from December 1, 2024, to December 31, 2025. Main Outcomes And Measures: Primary outcomes were performance of the risk stratification model, assessed by area under the receiver operating characteristic curve (AUROC), and monthly delirium detection rates. A 2 × 2 confusion matrix was created, and the 4 cells were used to calculate model performance metrics and compare associated diagnostic yield between premodel deployment and active model deployment periods. Results: The overall sample included 33 095 admissions to non-intensive care units at MSM and MSH (mean [SD] age, 74.86 (9.22) years; 15 601 women [47.1%], 17 271 men [52.2%]). A total of 9020 MSM inpatient admissions were assigned a delirium risk score by the risk stratification model during live clinical operations and were compared to 21 186 MSH inpatient admissions that were assigned a risk score by the same model. Age at admission significantly differed between patients at MSM and MSH (MSM: median [IQR] age, 75.79 [68.25-83.88] years vs MSH: 72.62 [66.64-79.65] years; P < .001). Patients from MSM and MSH also significantly differed by sex, race and ethnicity, surgical vs medical primary diagnosis, length of stay, and dementia prevalence rate. Among 11 909 MSM patients, patients in the pre-ML cohort (median [IQR] age, 78.00 [71.00-85.00] years) were significantly older than those in the post-ML cohort (median [IQR] age, 75.79 [68.25-83.88] years) (P < .001). The AUROC of the model was 0.91 (95% CI, 0.90-0.92) for MSM and 0.94 (95% CI, 0.93-0.94) for MSH. The positive predictive value of the model was 25.9% for MSM and 19.4% for MSH. For MSM, the mean monthly delirium detection rate significantly increased from 9.23% (95% CI, 8.07%-10.38%) to 19.59% (95% CI, 17.32%-21.85%) after model deployment (P < .001). Conclusions And Relevance: In this quality improvement study, an ML-based delirium risk stratification model maintained acceptable discriminative performance during external deployment without retraining and was similar to its performance at the development site without retraining. Model deployment was also associated with increased diagnostic yield but a high false positive rate. These findings provide further evidence of the potential generalizability of the model.

Objective: To develop ML models predicting postoperative CSF leakage after endoscopic transsphenoidal pituitary adenectomy (TSPA). Methods: Four hundred and twenty patients from five hospitals (January 2021-December 2022) comprised the training cohort; 140 from another hospital were the validation cohort. The training cohort included 60 CSF leak and 360 nonleak cases. Risk factors were analyzed. Decision tree, logistic regression, and SMOTE-based models were built and their diagnostic value compared. Results: Significant differences ( p < 0.05) existed in tumor diameter, tumor texture, operative time, intraoperative diaphragma sellae rupture, reoperation, and intraoperative CSF leakage grade. Logistic regression identified tumor texture, reoperation, intraoperative CSF leakage grade, and diaphragma sellae rupture as risk factors. All models had high AUCs in training: logistic regression (0.792), decision tree (0.835), and SMOTE (0.839). The Delong test showed a decision tree, and SMOTE outperformed logistic regression. SMOTE had the highest sensitivity/specificity. Validation cohort AUCs were: logistic regression (0.775), decision tree (0.809), and SMOTE (0.831), with SMOTE superior. Calibration curves indicated good agreement between predicted and actual values. Conclusion: Prediction models for CSF leakage postendoscopic TSPA were developed using decision trees, SMOTE, and logistic regression. Model efficacy was evaluated via AUC, specificity, and sensitivity. Comparing algorithms reduced single-algorithm bias, with SMOTE demonstrating the best performance.

Introduction: Perineural invasion (PNI) occurs in the majority of pancreatic ductal adenocarcinoma (PDAC) cases and is strongly associated with local recurrence, neuropathic pain, and poor survival, yet no reliable method exists for its preoperative identification. Here, we developed and externally validated a multi-sequence MRI-based intratumoral heterogeneity (ITH) scoring framework integrated with machine learning for the non-invasive prediction of PNI in PDAC. Methods: Using preoperative T2-weighted, diffusion-weighted, and contrast-enhanced T1-weighted MRI from 205 surgically resected patients across four medical centers (100 training, 43 internal test, and 28 and 34 external validation patients), we extracted two-dimensional and three-dimensional ITH features via K-means clustering-based subregion partitioning and incorporated independent clinical predictors tumor location and clinical T-stage to construct combined predictive models. Eleven machine learning algorithms were systematically evaluated. Results: The combined clinical-ITH model achieved an area under the curve (AUC) of 0.907 in the internal test set and AUCs of 0.860 and 0.869 in the two external validation cohorts. Corresponding AUCs were 0.881, 0.860, and 0.850 for the ITH model and 0.758, 0.342, and 0.700 for the clinical logistic regression model. Discussion: SHAP analysis ranked three T2WI-derived features as the largest contributors to the combined model output, feature effects varied in direction and were not uniformly monotonic. These findings demonstrate that multi-sequence MRI-based ITH quantification provides an interpretable approach to preoperative PNI assessment with consistent discrimination across the two external surgical cohorts, offering a potential tool to guide surgical planning and adjuvant therapy selection in pancreatic cancer.

Purpose: To develop and test models that estimate bone mineral density (BMD) from thoracolumbar junction CT images using Hounsfield unit (HU) assessment, radiomics, and convolutional neural networks (CNNs). Methods: We retrospectively identified patients who underwent lumbar dual-energy x-ray absorptiometry and CT covering T12 and L1 within 6 months at four hospitals between 2014 and 2025. In total, 465 patients were eligible for T12 analysis, and 448 for L1. Data from two institutions were used for training/validation and data from the other two for external testing. We developed five models: region-of-interest-based HU regression; volume-of-interest (VOI)-based HU regression; radiomics; CNN; and an ensemble combining VOI, radiomics, and CNN. Performance was assessed using mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R2), and Bland-Altman limits of agreement (LoA). Results: For T12, the CNN model performed best among the single models (MAE: 0.122, RMSE: 0.157, R2: 0.603), whereas the ensemble model showed a slightly higher MAE (0.128) but lower RMSE (0.154) and higher R2 (0.620). For L1, the radiomics model was the best single model (MAE: 0.124, RMSE: 0.155, R2: 0.582), while the ensemble showed numerically improved performance (MAE: 0.116, RMSE: 0.145, R2: 0.633). Pairwise comparisons showed no significant difference in absolute prediction errors between the ensemble and the best single model at either T12 or L1. Conclusion: We developed an ensemble framework integrating volumetric HU assessment, radiomics, and CNNs to estimate BMD from thoracolumbar junction CT images. External testing across two independent institutions demonstrated promising performance for BMD estimation.

Background: Cardiorenal syndrome (CRS) encompasses a spectrum of disorders in which acute or chronic dysfunction of the heart or kidneys induces dysfunction of the other. The heart-kidney crosstalk underlying CRS has since been extended into the broader cardiovascular-kidney-metabolic (CKM) syndrome by incorporating metabolic risk factors.CRS is associated with significantly increased mortality in intensive care unit (ICU) patients. However, dedicated tools for predicting short-term outcomes remain lacking, which limits timely risk stratification and treatment optimization. Methods: This multicenter cohort study enrolled 6,668 CRS patients from MIMIC-IV (v2.2) for model development and 3,602 from eICU-CRD (v2.0) for external validation. In patients remaining in the ICU for ≥24 hours, candidate predictors were ascertained within the first 24 hours, including early treatment exposures. Subsequent all-cause in-hospital mortality was predicted at this 24-hour landmark. Eleven machine learning algorithms were systematically compared after 1:1 random undersampling of the development cohort. SHapley Additive exPlanations (SHAP) was applied to enhance interpretability and identify key prognostic features. Results: The artificial neural network (ANN) model performed best. The final model incorporated six readily obtainable variables: AKI stage, age, mechanical ventilation, vasopressor use, β-blocker use, and ACEI/ARB use. It achieved an AUC of 0.810 (95% CI 0.757-0.860) in internal validation and 0.704 (95% CI 0.679-0.729) in external validation, indicating good internal but moderate external discrimination. Conclusions: We developed the first interpretable machine learning model specifically for predicting in-hospital mortality in critically ill CRS patients. SHAP provides transparent feature contribution analysis, and a web-based interface facilitates clinical implementation. When applied 24 hours after ICU admission, the model enables early identification of high-risk patients. It may also support risk stratification and resource allocation in the ICU.

Objective: To develop and externally evaluate machine learning models for predicting the heart failure with improved ejection fraction (HFimpEF) phenotype in patients with ischemic cardiomyopathy, using electrocardiogram (ECG) image features, all eligible clinical baseline features, and their combination. Methods: This two-center retrospective study included patients hospitalized between March 2018 and February 2025 with baseline left ventricular ejection fraction (LVEF) ≤40%. HFimpEF was defined as an increase in LVEF of at least 10 percentage points and a follow-up LVEF >40%, with at least three months between echocardiograms. The development cohort comprised 415 patients (194 with HFimpEF and 221 with persistent HFrEF); an independent cohort of 100 patients (55 and 45, respectively) was used for external validation. Four classifiers were evaluated using ECG image features alone, all eligible clinical baseline features alone, and their combination. LASSO was used separately to rank clinical variables. Results: In the reported internal evaluation, the combined random forest model achieved an AUC of 0.825, accuracy of 0.747, precision of 0.753, specificity of 0.791, and F1 score of 0.721. In external validation, the combined-model AUCs (95% confidence intervals) were 0.624 (0.513-0.734) for random forest, 0.607 (0.496-0.718) for logistic regression, 0.653 (0.545-0.761) for SVC-RBF, and 0.565 (0.454-0.676) for k-nearest neighbors. β-Blocker use ranked highest in the exploratory LASSO analysis. Combined models did not consistently outperform single-feature-set models externally. Conclusion: Models combining ECG image features and clinical baseline features showed promising internal discrimination for identifying subsequent HFimpEF in patients with ischemic cardiomyopathy. However, external discrimination was modest, and the combined models did not consistently outperform single-source models. Further model refinement and validation in larger, independent cohorts are needed before clinical application.

Background: Patients with lower extremity arterial disease (LEAD) complicated by diabetes face substantial amputation risk. Pre-amputation death acts as a critical competing event that distorts risk estimation. Validated personalized prediction tools for this population remain scarce and hinder precise clinical interventions. This study aimed to develop and validate two complementary interpretable models for 12-month amputation risk in patients with LEAD and diabetes to support clinical decision-making. Methods: This retrospective multicenter study enrolled 1,919 patients with LEAD and diabetes from four Chinese hospitals (January 2018-September 2024), with follow-up ending in September 2025. The primary outcome was 12-month amputation (major or minor). Prior to model construction, candidate predictors were screened via LASSO regression, random forest classifier and Boruta algorithm. Ten machine learning algorithms were then trained and evaluated by discrimination and calibration metrics. SHapley Additive Explanations (SHAP) was applied for interpretability. A Fine-Gray competing risk regression model was built using the same screened predictor set. Results: Patients were split into training (n = 967), internal validation (n = 414), and external validation (n = 538) cohorts. The 12-month amputation rate was 14.0%. Median follow-up was 12 months; mean age was 72.7 ± 9.5 years and 65.7% were male. The optimal GraBoost model achieved internal validation AUC = 0.860 and external validation AUC = 0.828, incorporating 14 clinical predictors, including: Rutherford classification, financial burden, age, Barthel Index, diabetic foot, statin, fibrinogen (FIB), neutrophil (NEUT), residence, prealbumin (PALB), C-reactive protein (CRP), activated partial thromboplastin time (APTT), body mass index (BMI) classification and wound. SHAP revealed higher grade Rutherford classification, lower Barthel Index, younger age, wounds, diabetic foot, and elevated FIB, APTT, CRP, NEUT correlated with higher amputation risk. Urban residence, normal BMI classification, lower financial burden, statin use and higher PALB were protective. Shared predictors showed consistent directional effects across GraBoost and Fine-Gray models. A dual-model web-based calculator was constructed for risk assessment. Conclusion: We developed and validated complementary interpretable GraBoost machine learning and Fine-Gray competing risk models for 12-month amputation risk. The dual-model web calculator may help clinicians identify high-risk patients and guide clinical management.

The integration of artificial intelligence (AI) into digital pathology has revolutionised breast cancer diagnostics, offering transformative opportunities to automate and refine molecular profiling, which is essential to precision oncology. Accurate determination of receptor status-estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2)-remains pivotal in guiding personalised treatment planning and prognostic assessment. However, traditional immunohistochemistry (IHC) and molecular assays are limited by resource intensity, lengthy turnaround times, and inter-observer variability, underscoring the need for scalable computational alternatives. This study proposes a novel integrative AI-driven framework for predicting breast cancer receptor profiles directly from hematoxylin and eosin (H&E)-stained histopathology images. The framework employs a hybrid vision transformer-convolutional neural network (ViT-CNN) architecture with a cross-attention fusion mechanism to capture both global contextual information and fine-grained local features. Experiments show AUCs of 0.967, 0.952, and 0.941 for ER, PR, and HER2 prediction, respectively, representing relative improvements of 12.4%, 9.8%, and 10.8% over the best-performing prior method (DCS-ST), re-evaluated under our identical protocol (absolute AUC differences of 0.107, 0.085 and 0.092). Furthermore, the framework enables automated molecular subtyping into luminal A, luminal B, HER2-enriched, and triple-negative categories with a macro-averaged F1-score of 95.7%. A comprehensive multi-institutional validation was performed on 2,847 external whole-slide images from three partner institutions (not including the primary BINO Hospital test set), showing consistent performance (a mean AUC reduction of 0.83 percentage points relative to the internal test set) across the staining protocols and scanners represented. Integrated explainability modules, including attention visualisations and Grad-CAM, enhance interpretability, supporting pathologist review. In a single-institution retrospective workflow audit, AI assistance was associated with an estimated 72.7% reduction in diagnostic turnaround time (from a conventional mean of 72.4 hours to 19.8 hours; Section 6.10, where approximately 65% of the AI-assisted time is un-itemised queueing and logistics) and, in a 120-WSI reader study, with higher inter-observer agreement (kappa 0.847 to 0.967). These clinical-impact estimates are hypothesis-generating and require prospective validation. The findings position the framework as a promising, explainable approach to AI-assisted digital pathology that warrants prospective, multi-centre clinical evaluation.

Background: Epidermal growth factor receptor (EGFR) mutation status is a key determinant of targeted treatment in lung adenocarcinoma, whereas co-occurring TP53 alterations may further affect therapeutic response and prognosis. This study aimed to develop a noninvasive CT-based approach for predicting EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma, thereby providing imaging support for pretreatment risk stratification and clinical decision-making. Methods: Patients with pathologically confirmed lung adenocarcinoma who underwent pretreatment chest CT and molecular testing were retrospectively enrolled from a single-center cohort. A total of 1,633 patients were included and categorized into five molecular groups: EGFR wild type, EGFR 19del, EGFR L858R, EGFR 19del with TP53 co-mutation, and EGFR L858 R with TP53 co-mutation. Based on chest CT images, key slice extraction, lesion ROI generation, automatic segmentation, and image standardization were performed. Semantic CT features, radiomics features, and deep CT representations were then derived. Under a unified data partitioning and standardized evaluation framework, four models were constructed and compared for the five-class classification task: a random forest model based on semantic CT features, a support vector machine model based on radiomics features, a Qwen-based semantic classification model, and a MedicalDINOv3-based deep learning model. To eliminate confounding bias induced by multimodal feature fusion, two ablation experiments were supplemented to split the model input features, and two MedicalDINOv3 models were built with inputs restricted to semantic features only and CT images only, respectively. Model performance was assessed using accuracy, macro-averaged area under the receiver operating characteristic curve (AUC), and macro-F1 score. External validation was further conducted using public and multicenter datasets. Results: Among the 1,633 patients in the internal cohort, 731 were EGFR wild type, 295 had EGFR 19del, 399 had EGFR L858R, 118 had EGFR 19del with TP53 co-mutation, and 90 had EGFR L858 R with TP53 co-mutation. To quantify the independent predictive value of different input modalities, ablation experiments on three variants of MedicalDINOv3 were performed. Ablation results revealed distinct performance among three input strategies. The MedicalDINOv3 with only semantic input obtained a test accuracy of 31.55% and an AUC of 0.6305 with limited predictive ability. The branch solely fed with CT images reached an accuracy of 55.61% and a macro-average AUC of 0.6347. The complete MedicalDINOv3 integrating CT deep features, clinical and semantic features performed best, with a test accuracy of 0.6905 and a macro-average AUC of 0.8788. This comparison proved that CT deep representations outperform traditional handcrafted imaging and semantic features, and multi-modal fusion further improves subtype stratification performance. The support vector machine model achieved an accuracy of 0.527, a macro-AUC of 0.792 and a macro-F1 score of 0.524; the random forest model achieved an accuracy of 0.455, a macro-AUC of 0.819 and a macro-F1 score of 0.412; and the Qwen-based model achieved an accuracy of 0.339, a macro-AUC of 0.700 and a macro-F1 score of 0.316. Receiver operating characteristic analyses showed that the deep learning model provided relatively stable discrimination for EGFR wild type and TP53 co-mutated categories, whereas overlap remained between the two major EGFR-sensitive subtypes, EGFR 19del and EGFR L858 R. In external validation, the public dataset used for EGFR binary classification yielded an overall accuracy of 67.0% and an AUC of 0.6666. The raw overall accuracy of the cohort was 75.5%, and raw overall accuracy is prone to overestimation when the cohort has severe class imbalance dominated by EGFR wild-type samples, which can only serve as an auxiliary reference. Multicenter five-class external validation yielded a balanced accuracy of 58.2% and a macro F1 of 0.574. The subtype-specific AUCs for EGFR wild type, EGFR 19del, EGFR L858R, EGFR 19del with TP53 co-mutation, and EGFR L858 R with TP53 co-mutation were 0.8342, 0.6504, 0.7221, 0.6586, and 0.7563, respectively. Conclusion: A CT-based deep learning model can provide noninvasive stratification of EGFR mutation subtypes and TP53 co-mutation status in lung adenocarcinoma. Compared with semantic CT features and radiomics features, deep CT representations showed superior overall discriminative performance. This approach may serve as an adjunctive tool to molecular testing for pretreatment assessment in lung adenocarcinoma. However, the model showed inconsistent generalization performance across different molecular subtypes, with notably limited predictive efficacy for EGFR 19del and EGFR 19del+TP53 subtypes. Therefore, predictions for these categories should be interpreted with caution, and the model requires further validation in larger, multicenter, and standardized datasets before clinical application.