2026-09-25

Background And Purpose: Prehospital delay remains a major barrier to timely reperfusion therapy in acute ischemic stroke (AIS). We externally validated eight prediction algorithms and examined whether complex machine-learning models added value beyond logistic regression. Methods: This retrospective cohort study included consecutive patients with AIS from two hospitals in Kashgar, China (January 2019-December 2024). Hospital 1 (n = 1,377) was the development cohort and Hospital 2 (n = 1,015) the independent validation cohort. Prehospital delay was defined as onset-to-door time >4.5 h. LASSO selection and tuning were restricted to Hospital 1. Sensitivity analyses used all 36 encoded candidate features and repeated all models without SMOTE. Evaluation included discrimination, calibration, post-hoc recalibration, and decision curves. Results: Delay occurred in 80.9 and 83.2% of the two cohorts. In the prespecified LASSO-plus-SMOTE analysis, XGBoost had the highest validation AUC (0.771; 95% CI, 0.734-0.807), but did not significantly outperform logistic regression (difference, 0.015; 95% CI, -0.003 to 0.034; p = 0.107). Full-feature modeling produced no consistent gain. Omitting SMOTE improved calibration and often discrimination; logistic regression without SMOTE achieved AUC 0.774 and Brier score 0.122, versus 0.756 and 0.180 with SMOTE. The original SMOTE-trained logistic model underestimated absolute risk. Rural residence, non-emergency-channel presentation, referral, and non-ambulance transport were associated with absolute risk differences of 15.8, 21.5, 9.2, and 36.3 percentage points in Hospital 2. Conclusion: Complex algorithms did not show a clear advantage over logistic regression. Because several predictors are available only after arrival, the framework is intended for retrospective health-system profiling and quality-improvement planning, not individual pre-arrival prediction. Universal interventions remain warranted given the high prevalence of delay.

Background: Early distant recurrence within 18 months reflects aggressive tumor biology in Stage III colon cancer. Current anatomical staging often fails to capture this biological heterogeneity. Objective: To develop and externally validate a simple, pathology-based model for ruling out early distant recurrence within 18 months in Stage III colon cancer. Design: Retrospective model development with external validation, following current machine-learning prediction-model reporting guidelines. Settings: Derivation at a tertiary referral center; validation at an independent tertiary institution. Patients: Three hundred thirty-one patients in the derivation cohort (62 events) and 142 in the external cohort (19 events) who underwent curative-intent colectomy for Stage III colon adenocarcinoma. Main Outcome Measures: Early distant recurrence within 18 months. Secondary outcomes included disease-free survival and calibration performance. Results: A 4-variable XGBoost model (American Joint Committee on Cancer substage, lymph node ratio, perineural invasion, and differentiation) achieved a pooled out-of-fold area under the curve of 0.680 in the derivation cohort, comparable to conventional staging (0.625). Sensitivity (77.4%) and negative predictive value (89.8%) exceeded those of conventional staging. On external validation, the model yielded an area under the curve of 0.633 (vs 0.610 for conventional staging) and a negative predictive value of 90.8%, maintaining rule out performance of approximately 90% across both cohorts despite their differing recurrence rates. High-risk patients in the external cohort had worse disease-free survival (hazard ratio 1.83; 95% confidence interval 1.06-3.16). Exploratory subgroup analyses showed directionally consistent but statistically inconclusive results. Limitations: Retrospective design, limited molecular data, and small numbers of events in the external cohort. Conclusions: A simple 4-predictor model using routine pathological variables provides consistent rule out performance for early distant recurrence within 18months across 2 independent cohorts. The model offers a practical risk-assessment tool for early postoperative counseling where molecular testing is unavailable. A web-based calculator was developed to support early postoperative risk assessment. See Video Abstract.

Objective: This study aimed to identify factors associated with adenoma recurrence within three years after endoscopic mucosal resection (EMR) and to develop an individualized predictive model. Methods: Patients undergoing their first EMR for colorectal polyps at the Affiliated Hospital of Xuzhou Medical University (September 2018-May 2025) were retrospectively included and randomly divided into training and testing cohorts (7:3). Patients from the Third Affiliated Hospital of Xuzhou Medical University served as the external validation cohort. Least absolute shrinkage and selection operator (LASSO) regression was used to select predictors. Logistic regression (LR), random forest (RF), support vector machine (SVM), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost) models were constructed. Model performance was assessed using receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA). SHapley additive explanations (SHAP) values were applied to interpret variable contributions, and an online prediction tool was developed. Results: A total of 1,454 patients were enrolled, of whom 731 developed adenoma recurrence within three years. LASSO identified ten predictors: infection of Helicobacter pylori (H. pylori), the number of adenomas, age, triglyceride level, body mass index (BMI), diarrhea, smoking history, family history, adenoma size, and fecal occult blood positivity. GBM achieved the highest mean AUC in repeated cross-validation (0.813) and maintained robust discrimination in the testing (AUC = 0.818) and external validation cohorts (AUC = 0.775), with good calibration and clinical utility. SHAP analysis identified H. pylori infection, adenoma number, and age as the leading contributors to model predictions. Conclusion: GBM demonstrated favorable discrimination, calibration, and clinical utility across validation cohorts, supporting its use for individualized risk stratification. The resulting online calculator may further facilitate risk assessment and inform follow-up strategies in clinical practice.

Objective: To develop and temporally prospectively validate an interpretable model for identifying concurrent depressive symptoms in hospitalized patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD). Methods: This study was conducted in two stages: retrospective development and prospective validation. In the development stage, 304 AECOPD patients admitted to the Fourth Affiliated Hospital of Anhui Medical University from September 2023 to February 2025 were retrospectively enrolled and randomly divided into a training set (n = 214) and an internal validation set (n = 90) at a 7:3 ratio. Candidate predictors were screened using least absolute shrinkage and selection operator (LASSO) regression, and the selected variables were subsequently incorporated into an unpenalized multivariable logistic regression model as the primary clinical model. A nomogram was developed based on the logistic regression coefficients for individual risk estimation. Additionally, 10 machine learning algorithms were compared, and a stacked ensemble strategy was explored as a supplementary analysis to evaluate potential incremental predictive value. Model discrimination, calibration, and clinical utility were assessed using multiple metrics. In the validation stage, 120 AECOPD patients admitted to the same center from October to December 2025 were prospectively enrolled for single-center temporal validation. Model interpretability was provided by the final logistic regression coefficients and the nomogram. Results: LASSO regression screened four candidate predictors: physical function (PHYS), mental health (PSYCH), environmental support (ENVIR), and quality of life. The selected variables were entered into an unpenalized multivariable logistic regression model as the final clinical tool. The final logistic regression model achieved an AUC of 0.866 (95% CI: 0.819-0.912) in the training set, 0.862 in the internal validation set, and 0.866 (95% CI: 0.772-0.960) in the prospective validation set. In exploratory comparisons, the SVM + LR stacked ensemble demonstrated comparable discrimination to the logistic regression model (AUCs: 0.863, 0.868, and 0.864 in the training, internal validation, and prospective validation cohorts, respectively), but no statistically significant improvement was observed (DeLong test, p = 0.312). Calibration analysis showed good agreement in the training and internal validation sets, whereas miscalibration was observed in the prospective cohort. Decision curve analysis suggested potential clinical utility for risk stratification. Model interpretability was provided by the final logistic regression coefficients and the nomogram. Conclusion: A four-feature interpretable model was developed to identify AECOPD patients with concurrent depressive symptoms during hospitalization. The model demonstrated satisfactory discriminative performance and may assist early risk stratification; however, multicenter external validation, local recalibration, and decision-impact analysis are required before routine clinical implementation.

Background And Objectives: Busulfan, used in conditioning regimens prior to haematopoietic stem cell transplantation, has a narrow therapeutic index; therefore, due to pharmacokinetic variability, therapeutic drug monitoring (TDM) is employed to ensure drug exposure levels associated with optimal clinical outcomes. This study aimed to facilitate dose adjustment and potentially reduce the number of samples required for TDM in clinical practice. Methods: Population pharmacokinetic (popPK) and machine learning (ML) models (Random Forest and XGBoost) were developed using a local, retrospective, extensive-sampling dataset of 67 children (637 observations, Denmark) and externally evaluated using prediction error (PE) on a TDM dataset from another centre with 171 children (1871 observations, Netherlands). Limited sampling schedules were evaluated for both methods. Results: A one-compartment model was the best data fit, showing time-varying clearance decreases of 10.4% and 13% on Days 2 and 3, respectively, compared to Day 1. The popPK, Random Forest, and XGBoost models performed well on the external dataset for concentration prediction, with median PE ranging from -4.82 to -7.44%, median absolute PE of approximately 15%, and approximately 80% of data predicted within ±30% of the PE. Sampling strategies at 0, 0.5, 1, 4 h or 0, 1, 2, 4 h post-infusion adequately predicted busulfan plasma concentration and exposure using the popPK model. Conclusions: A new popPK model was developed to support individualised dose adjustments in children using fewer blood samples. Both Random Forest and XGBoost could predict busulfan's plasma concentration with performance like that of the popPK model (population prediction). However, a larger dataset is needed for ML methods to describe a complete pharmacokinetic profile.

The absence of accurate, noninvasive, scalable screening tools keeps early esophageal cancer (EC) detection a global health challenge. Although noncontrast computed tomography (NC CT) is widely accessible, the esophagus is a hollow tubular structure prone to collapse and motion artifacts, making small early malignant lesions difficult to distinguish from normal tissue. Here we developed the Esophageal AI-Guided malignant Lesion Evaluation (EAGLE) model to detect precancerous lesions and cancer from chest NC CT, a task historically considered impossible. EAGLE was trained on 6,813 patients from two centers and validated across 12 centers in three countries involving 80,612 patients in opportunistic and population-based screening settings. For opportunistic screening on existing CT scans, multicenter external test cohorts (eight centers, n = 11,466) achieved 98.5% specificity, with 90.0% sensitivity for cancer and 52.5% for precancerous lesions; low-dose CT (LDCT) validation (two centers, n = 1,607) showed comparable performance, supporting EC screening through lung-cancer screening programs. Calibration in a real-world cohort (three centers, n = 35,402) reduced false positives by 72.7% while preserving sensitivity; prospective hospital validation (n = 17,446) achieved a 42.2% PPV, and real-world low-dose screening (n = 10,959) reached 99.94% specificity. EAGLE also detected precancerous lesions-in paired CT-endoscopy cohorts (two centers, n = 702), sensitivities were 65.0% for precancerous lesions and 78.4% for stage I EC at a higher-sensitivity operating point. Exploratory analyses of a prospectively enrolled cohort suggest that referring high-risk individuals for endoscopy could improve screening efficiency. In conclusion, EAGLE has the potential to serve as a scalable tool for early EC screening. Chictr.org.cn identifier: ChiCTR2300074806 .

Background: Early identification of high-risk sepsis patients remains challenging. The C-reactive protein-Albumin-Lymphocyte (CALLY) index, which integrates inflammation, nutritional status, and immune function, may hold prognostic value in this setting. Methods: We retrospectively analyzed sepsis patients from two Chinese hospitals (discovery cohort: n = 695; validation cohort: n = 515). Three machine learning algorithms were applied to identify key predictors. A logistic regression model was constructed, internally validated, and externally tested. Model performance was compared with APACHE II and SOFA scores using DeLong's test. Results: The non-survivors had markedly lower CALLY values (median 0.2 vs. 0.8, P < 0.001). CALLY, lactate dehydrogenase (LDH), total proteins (TP), and platelet count (PLT) were consistently selected as top predictors across all three machine learning methods. The final model demonstrated good discrimination in the development cohort (AUC = 0.877; 95% CI: 0.847-0.908), with a significantly higher AUC than APACHE II (P < 0.001) and SOFA (P = 0.001). Calibration was satisfactory (Hosmer-Lemeshow P = 0.638), and decision curve analysis indicated clinical net benefit. In external validation, the model retained predictive ability (AUC = 0.920), and CALLY remained significantly lower in non-survivors (P < 0.001). Conclusion: The CALLY index is an acceptable, readily available biomarker strongly associated with 28-day mortality in sepsis. A prediction model incorporating CALLY, LDH, TP, and PLT offers performance comparable or superior to conventional severity scores and may support early risk stratification in clinical practice, provided that calibration-in-the-large is updated when the model is applied to settings with substantially different baseline mortality.

Haematological malignancies are rare in primary care and are hard to predict. Existing prediction models often require a specific working diagnosis and are based on extensive clinical information, access to which may be limited at initial work-up in primary care. We aimed to develop a tool for estimating overall risk of haematological malignancy in primary care patients using simple data inputs. Including all full blood cell counts (FBCs) measured from 2000 to 2016 in 856 403 adult primary care patients in Eastern Denmark, we used machine learning to develop CBC-HEMA, a risk assessment tool using FBCs, age, sex and C-reactive protein measures to stratify 'any haematological malignancy' risk within 6 months, 1 year and 3 years with area under the curve (AUC) of 0.81, 0.78 and 0.71 respectively. Predictions of chronic lymphoid leukaemia and myeloproliferative disease were excellent, while CBC-HEMA was unable to accurately predict non-Hodgkin lymphoma and plasma cell dyscrasia. External validation was conducted in 123 015 primary care patients from Western Denmark with similar results and in 418 410 individuals from the background population cohort UK Biobank with inferior performance showing unfitness in screening healthy populations. The CBC-HEMA https://shiny.sund.ku.dk/CopLab/blood-cancer-risk-prediction/ may support risk assessment in primary care but cannot replace a clinical evaluation.

Objectives: To develop and externally validate machine learning models for predicting disease flares in axial spondyloarthritis (axSpA) using real-world registry data. Methods: Patients with axSpA receiving biological disease-modifying antirheumatic drugs in the METEOR-SpA registry were analysed at visit level. Flare was defined as an Axial Spondyloarthritis Disease Activity Score (ASDAS) increase ≥0.9 (or Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) increase ≥2) at the next 6-month assessment. Candidate models utilising routine clinical predictors and minimum-redundancy maximum-relevance variable selection were optimised for average precision (AP) to account for class imbalance, evaluated using patient-grouped cross-validation (CV) and externally validated in the Groningen Leeuwarden AxSpA cohort (GLAS) using discrimination, calibration and decision curve analysis (DCA). Results: After eligibility criteria, 1167 patients and 3865 visits were included from the METEOR-SpA cohort; 618 (52.9%) patients had at least one visit with known flare status. Among them, 170 (27.5%) experienced at least one flare. A ridge logistic regression was selected as the final model with an internal CV receiver operating characteristic plot and area under the curve (ROC-AUC) of 0.72 (95% CI 0.70 to 0.75) and an AP of 0.291 (0.245 to 0.337), retaining eight predictors associated with flare. These were ASDAS (current and previous), female sex, flare history, enthesitis, infliximab exposure, physician global assessment and C reactive protein. In external validation in GLAS, ROC-AUC was 0.63 (95% CI 0.59 to 0.68) and AP 0.142 (0.116 to 0.171). DCA showed a modest positive net clinical benefit at low threshold probabilities. Conclusions: A reduced logistic-regression model showed moderate internal and modest external discrimination for short-term flare risk in axSpA. While not intended to guide individual pre-emptive therapy modification, the model may support low-risk interventions such as group-level risk stratification or tighter follow-up after further validation.

Background: Anticoagulants are high-alert medications with substantial risk of serious bleeding, yet dosing and prescribing errors remain common. Although clinical decision support systems (CDSSs) can mitigate these errors, their impact is constrained by alert fatigue and limited interpretability. Objective: We developed and validated a hybrid CDSS that integrates rule-based logic with machine learning to improve the safe use of anticoagulants. Methods: This multicenter study used electronic health record data on anticoagulant prescriptions from 3 tertiary hospitals (1 for system development and internal validation and 2 for external validation). The hybrid CDSS combined a knowledge-based rule engine with a machine learning model trained to predict whether an anticoagulant prescription would require pharmacist intervention. The system was iteratively refined through pilot testing, internal validation, and external validation. Results: A total of 75,200 anticoagulant prescriptions were used for model development. The final hybrid CDSS comprised 44 patient-specific rules and 1129 drug-drug interaction rules, combined with a CatBoost classifier (version 1.2.5; Yandex) for alert prioritization. During internal validation, 88 (18.9%) alerts were generated; all were technically correct; 88.6% (n=78) were deemed clinically relevant; 86.4% (n=76) were considered clinically useful; and 13.6% (n=12) required pharmacist intervention. In external validation across 2 hospitals, alert rates ranged from 22.6% (7/31) to 32.1% (310/966), with 18.8% (22/117) to 57.1% (4/7) of alerts requiring pharmacist intervention. The hybrid CDSS showed strong discrimination (area under the receiver operating characteristic curve 0.871-0.963). No false negatives were identified, but the estimates should be interpreted cautiously given the short validation periods and limited number of intervention-requiring prescriptions. Conclusions: A hybrid CDSS integrating rule-based logic with machine learning demonstrated high technical accuracy and clinical relevance across multiple institutions, suggesting its potential as a practical tool for supporting pharmacist-led anticoagulant prescription review.

Objective: To develop and validate deep learning models to differentiate choroideremia (CHM), USH2A-associated rod-cone dystrophy, and healthy controls using macular OCT volumes. Design: Retrospective diagnostic study with external validation. Subjects: Two hundred sixty-two participants contributed 535 macular OCT volumes acquired on a Heidelberg Spectralis (10 165 B-scans): 99 participants had CHM, 96 had USH2A-associated disease, and 67 were healthy controls. Methods: Volumes acquired between 2012 and 2021 were standardized to 19 B-scans and split at the patient level into training (76%), validation (12%), and test (12%) sets. Three strategies were compared: (1) a slice-based convolutional neural network (CNN) (ResNet-50) with majority-vote aggregation; (2) a Mixture-of-Experts (MoE) model with learned slice weighting; and (3) a hybrid CNN-Transformer integrating per-slice features across the volume. Two ophthalmologists masked to clinical and genetic data graded the 80 test volumes. Main Outcome Measures: Volume-level accuracy and weighted F1-score. Results: On the independent test set (80 volumes: 33 CHM, 29 USH2A, 18 controls), all models achieved high performance (accuracy ≥0.963). The MoE achieved the highest accuracy on the internal test set (accuracy 0.988; 79/80; weighted F1-score 0.988), followed by CNN + vote (accuracy 0.975; 78/80; weighted F1-score 0.975) and the hybrid CNN-Transformer (accuracy 0.963; 77/80; weighted F1-score 0.962). Ophthalmologists achieved lower accuracy (0.875 and 0.863). External validation on an independent Rennes University Hospital data set (n = 28 volumes; CHM = 6, controls = 10, USH2A = 12), processed identically without retraining, yielded accuracy 28/28 for CNN + vote, whereas the MoE and hybrid CNN-Transformer each achieved 26/28. Conclusions: Deep learning applied to macular OCT volumes accurately distinguished CHM, USH2A-associated rod-cone dystrophy, and healthy controls, and achieved higher observed accuracy than masked human readers in this OCT-only setting. Both learned slice weighting and simple aggregation achieved strong performance, suggesting that increased architectural complexity may not be necessary for this 3-class classification task. Further validation in larger, more diverse, and clinically representative cohorts is required to determine the potential role of such models as complementary diagnostic support. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Objective: The lipid paradox suggests low LDL-C is paradoxically associated with poor outcomes, yet its joint effect with insulin resistance in acute stroke remains unclear. This study aimed to evaluate the combined association of triglyceride-glucose (TyG) index and LDL-C with mortality and days alive out of hospital (DAOH) METHODS: This multi-center cohort study used MIMIC as the derivation cohort and eICU-CRD as external validation. Multivariable regression, survival analysis, and RCS were used to assess the associations of TyG and LDL-C with mortality and DAOH90. Mediation analysis explored potential mechanisms. Boruta and LASSO selected features for ten machine learning models, with SHAP for interpretation. Results: Among 3,028 MIMIC and 2,385 eICU patients, TyG showed a positive linear association with mortality, while LDL-C exhibited an L-shaped association. The low LDL-C + high TyG group had the highest in-hospital mortality (OR = 3.36, 95%CI: 2.37-4.76) and shortest DAOH90. Inpatient statin use was associated with lower mortality. WBC statistically accounted for 15.46% of the observed association between TyG and in-hospital mortality. Among multivariable machine-learning models incorporating nine selected predictors, GBM achieved the best performance (AUC = 0.802). Conclusions: High TyG combined with low LDL-C identifies a high-risk phenotype in acute stroke, significantly associated with adverse outcomes. TyG and LDL-C may serve as accessible biomarkers for risk stratification in acute stroke.

Breast cancer is the most frequently diagnosed malignancy among women worldwide and a leading cause of cancer-related mortality. Dynamic contrast-enhanced magnetic resonance imaging plays a central role in tumor characterization and treatment monitoring, particularly in patients receiving neoadjuvant chemotherapy. However, existing artificial intelligence models for breast magnetic resonance imaging are typically developed and evaluated using heterogeneous datasets, study populations, and assessment protocols, making direct comparison difficult and limiting understanding of model robustness across institutions and clinically relevant patient subgroups. The MAMA-MIA Challenge was designed to address these challenges by providing a standardized benchmark for the joint evaluation of primary tumor segmentation and prediction of pathologic complete response using pre-treatment magnetic resonance imaging only. The training cohort comprised 1,506 patients from multiple institutions in the United States, while evaluation was conducted on an external test set of 574 patients from three independent European centers to assess cross-continental and cross-institutional generalization. A unified scoring framework combined predictive performance with subgroup consistency across age, menopausal status, and breast density. Twenty-one international teams participated in the final evaluation phase. Results demonstrate substantial performance variability under a common external evaluation framework and reveal trade-offs between overall accuracy and subgroup fairness. The challenge provides standardized datasets, evaluation protocols, and public resources to promote the development of robust and equitable artificial intelligence systems for breast cancer imaging.

Background And Purpose: In MR-guided brachytherapy, organs at risk (OAR) are manually delineated during the treatment planning process while the patient waits for the treatment to start with the applicator in place. This study evaluates the delineation time gain and bias introduction of an in-house created artificial intelligence (AI) model trained with a manually uncurated clinical dataset. Materials And Methods: An nnU-Net was trained on clinical T2w-MR scans for brachytherapy. OAR and target structures underwent automatic interpolation but no manual recontouring. Following clinical implementation, delineation times and structure sets from 32 individual patients with 64 pulsed dose rate (PDR) treatments were collected prospectively from December 2023 to December 2024. Each patient received two treatment fractions: the first was based on AI-assisted delineations corrected by the oncologist, and the second was with the oncologist delineating from scratch without AI assistance. The reduction in delineation time and differences between the uncorrected AI-based and clinically approved delineations were analysed. Results: AI-based delineations resulted in a time reduction of 10.1 min (95% CI: [7.4, 12] minutes), corresponding to a 31% decrease in delineation time. The AI-based delineations were generally rated acceptable by oncologists, requiring only minor corrections in most cases. When using AI support, the median of the mean surface distance between the AI and clinically used delineation for all OARs and targets was below 2.1 mm. Conclusion: AI-based delineation models, trained on clinical data without additional manual recontouring, can reduce delineation time. In the current clinical trial, the delineation time was reduced by 31%.

Alzheimer's disease (AD) classification from structural magnetic resonance imaging (MRI) may benefit from weak supervision that uses clinically meaningful but imperfect supervisory signals. We evaluated a weakly supervised framework in which a multilayer perceptron (MLP) trained on age, sex, and Mini-Mental State Examination (MMSE) scores generated clinical pseudo-labels to initialize a patch-based fully convolutional network (FCN). For 260 Alzheimer's Disease Neuroimaging Initiative (ADNI) training participants, subsequent refinement combined 80% of the preceding MRI-model probability with 20% of the participant's ground-truth diagnostic label. This design preserves a dominant pseudo-label/self-training component while using partial diagnostic guidance to stabilize refinement. The FCN generated whole-brain probability maps, and selected voxel probabilities were classified by a second MLP. The framework was developed using ADNI (n = 417). Using ADNI validation data only, iteration 3 and a classification threshold of 0.5 were selected and then applied unchanged to the held-out ADNI test set and the external AIBL (n = 182), FHS (n = 102), and NACC (n = 265) cohorts. The selected model achieved F1 scores of 0.853 in ADNI, 0.707 in AIBL, 0.765 in FHS, and 0.807 in NACC. These results support the feasibility and cross-cohort transferability of clinical pseudo-label-based weak supervision for MRI classification. The framework is not intended to be label-free; rather, it provides a transparent strategy for integrating imperfect clinical pseudo-labels with partially weighted diagnostic guidance during training.