2026-10-15

Background: Raised intracranial pressure is one of the most severe secondary insults after traumatic brain injury. Recognising physiological patterns before they progress to dangerous intracranial pressure elevations could enable preventive rather than reactive interventions. In our previous Uppsala single-centre study, an XGBoost machine learning model showed high discriminative performance for prediction of imminent intracranial pressure insults. The model identifies short-term physiological trajectories associated with imminent threshold crossing rather than predicting independent future clinical events. Whether these results are generalisable across centres is unknown. The aim was therefore to evaluate whether the XGBoost model trained on an Uppsala dataset maintains similar performance when applied to a multicentre TBI cohort - The CENTER-TBI High-Resolution ICU cohort. Methods: The XGBoost model trained on the Uppsala high frequency traumatic brain injury cohort was externally validated using data from the multicentre CENTER-TBI cohort. Minute-level averages of intracranial pressure and invasive arterial blood pressure were used to predict the onset of an intracranial pressure insult within 30 min on the basis of a 4h historical window. A bidirectional validation design was applied and performance was assessed using multiple metrics. Results: The cohorts comprised 602 patients from Uppsala and 255 patients from the CENTER-TBI database. When trained on the Uppsala cohort and tested on CENTER-TBI, the model achieved an area under the receiver operating characteristic curve (AUROC) of 97.4%, sensitivity of 93.8%, and specificity of 98.6%. Training on CENTER-TBI and testing on the Uppsala dataset yielded an AUROC of 96.1%, sensitivity of 81.2%, and specificity of 98.7%. Balanced accuracy exceeded 88% in both validation directions. Conclusions: The XGBoost model retained high performance under external multicentre validation, indicating that pre-insult physiological patterns preceding increased intracranial pressure are shared across cohorts. Differences in sensitivity between validation directions were observed and may reflect cohort characteristics or dataset size. These findings support the need for local evaluation before clinical implementation.

Background: Clonal haematopoiesis is a precursor of myeloid neoplasia and common in older adults, Precursor risk stratification is critical to early cancer detection, guiding surveillance and prevention efforts to high-risk populations. Existing clonal haematopoiesis risk scores use molecular features to estimate risk of myeloid neoplasia, but only around 1% of adults with detectable clonal haematopoiesis are at high risk of myeloid neoplasia transformation. Moreover, sequencing based molecular diagnostics are expensive with limited global availability highlighting the need for pre-sequencing strategies which identify populations enriched for individuals with high-risk clonal haematopoiesis. Methods: We used blood count parameters, age, and sex as input features to and develop a multi-class machine learning model that pre-classifies individuals based on their probability of having high-risk clonal haematopoiesis or developing future myeloid neoplasia. Stratified splitting was used to divide 461,576 eligible UK Biobank participants (recruited October 1, 2006 to March 13, 2010) into a training cohort for model derivation (n = 369,260) and a test cohort for internal validation (n = 92,316). Ground truth output class assignments were defined based on documented progression to incident myeloid neoplasia and presence of clonal haematopoiesis with high-risk features defined using molecular parameters of the clonal haematopoiesis risk score. Individuals were labelled positive if they had high-risk clonal haematopoiesis and progressed to myeloid neoplasia, negative if they lacked clonal haematopoiesis with high-risk features, and equivocal if they had higher risk features but had no documented incident myeloid neoplasia event. We then compared model predictions against ground truth labels for validation in the UKB test cohort as well as in independent clinical and population cohorts from Dana-Farber Cancer Institute Hematologic Malignancy Data Repository (HMDR; n = 2233; March 26, 2010, to June 5, 2024) and the All of US database (n = 14,674; May 31, 2017 to October 1, 2023), respectively. Findings: The resulting pre-sequencing risk model (PRISM) categorises all cases of high-risk clonal haematopoiesis as positive and pre-classifies 40% (2158 of 5383) of low-risk clonal haematopoiesis and 55% (47,894 of 86,429) of clonal haematopoiesis-negative cases as negative. Validation of PRISM revealed an area under the curve of 0.89 (95% CI 0.85-0.93) in the UK Biobank test cohort, 0.74 (0.61-0.84) in HMDR, and 0.87 (0.77-0.96) in All of US. Negative predictive values for negative class assignments were 0.993 (95% CI 0.992-0.993) in the UKB test cohort, 0.968 (0.943-0.981) in HMDR, and 0.998 (0.997-0.999) in All of US. Interpretation: PRISM positive class assignment defines a population most enriched for high-risk clonal haematopoiesis, while negative class predictions, mark individuals without high-risk clonal haematopoiesis. PRISM therefore provides a framework for reproducibly assigning blood count-based risk thresholds to prioritise clonal haematopoiesis screening in populations at risk for myeloid neoplasia. Funding: Damon Runyon Cancer Research Foundation, Edward P Evans Foundation for MDS, and the National Heart Lung and Blood Institute.

Background: Acute leukemias require rapid subtype identification to guide timely treatment. Differentiation of acute lymphoblastic leukemia (ALL), acute monocytic leukemia (AMoL/M5), and acute promyelocytic leukemia (APL/M3) often relies on specialized tests that may not be immediately available. This study evaluated differentiation only among patients already confirmed to have one of these three diagnoses, rather than screening an unselected population with suspected acute leukemia. Methods: This multicenter retrospective study included 9,398 newly diagnosed, previously untreated patients with established APL/M3, AMoL/M5, or ALL between May 2019 and May 2025 at four tertiary hospitals in China. Relapse, remission, follow-up, and non-initial disease episodes were excluded. Twenty-four routine hematological parameters were evaluated. One center served as the development cohort with five-fold stratified cross-validation, while three independent centers were reserved for external validation. Ten machine-learning algorithms were compared. Performance was assessed using AUC, sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and F1 score. SHapley Additive exPlanations (SHAP) supported feature selection and interpretation. Results: An 18-feature random forest model achieved a development-stage cross-validated AUC of 0.93 and external AUCs of 0.89, 0.87, and 0.90. Overall sensitivity, specificity, PPV, NPV, accuracy, and F1 score were 0.78, 0.90, 0.81, 0.91, 0.81, and 0.79, respectively. SHAP identified leukocyte, erythrocyte, and platelet indices as the most important features. A Streamlit application was developed for clinical evaluation. Using a development-derived APL alert threshold of 0.0966, APL sensitivity was 99.02% in development and 96.32%-97.94% across external cohorts, with PPVs of 54.34% and 35.58%-84.25%, respectively. External validation showed center-specific calibration shifts, indicating that predicted probabilities should not be interpreted as universally calibrated absolute diagnostic probabilities. Conclusions: Among patients already confirmed to have APL/M3, AMoL/M5, or ALL, this explainable model enabled low-cost preliminary differentiation using routine hematological data. It may support subtype prioritization and confirmatory testing but cannot replace standard diagnostics, identify other diagnoses, or screen unselected patients. Prospective multinational and workflow-based validation is required before routine implementation.

Objective: This study aimed to identify independent predictors of seizure occurrence in patients with neuropsychiatric systemic lupus erythematosus (NPSLE), develop a diagnostic model, and evaluate whether seizures are associated with adverse renal outcomes. Methods: A retrospective cohort of 169 patients with NPSLE from Xiangya Hospital was analyzed. Among 70 pre-specified clinical, laboratory, and imaging variables, 20 variables associated with seizure occurrence (p < 0.05 in univariable logistic regression) were retained after multicollinearity screening (VIF < 5.0). Multivariable logistic regression was then performed to identify independent predictors of SE-NPSLE. Machine learning algorithms with fivefold cross-validation were then employed to develop and optimize the diagnostic model. Model performance was evaluated using sensitivity, specificity, accuracy, and AUC metrics. External validation was performed using an independent cohort (n = 46) from Jiangsu Provincial People's Hospital. Renal survival rates were assessed by Kaplan-Meier analysis with log-rank tests. Results: Independent predictors of SE-NPSLE included anti-Rib-P antibody (OR = 0.315, 95% CI: 0.121-0.819), erythrocyte sedimentation rate (ESR) (OR = 0.973 per mm/h, 95% CI: 0.954-0.992), and presence of psychosis (OR = 0.194, 95% CI: 0.057-0.659). A logistic regression-based nomogram (AUC: 0.818 training, 0.800 validation) demonstrated robust performance. External validation (n = 46) yielded an AUC of 0.706. What is more, NPSLE patients with seizures (SE-NPSLE) demonstrated significantly worse renal outcomes compared to NPSLE patients without seizures (non-SE-NPSLE). Conclusion: Seizures in NPSLE are associated with poor renal prognosis. Our diagnostic model was developed to facilitate early risk assessment and may aid personalized seizure management; however, its clinical utility requires validation in larger, prospective cohorts. Key Points • A validated diagnostic model indentifies high-risk NPSLE patients for seizures. • 3 independent predictors of SE-NPSLE included anti-Rib-P antibodies, ESR and psychosis. • SE-NPSLE patients were youner, had higher SLEDAI, and exhibited more severe renal dysfunction than non-SE-NPSLE patients.

Accurate assessment of lymph node metastasis (LNM) is critical for prognostic stratification and treatment selection in colorectal cancer (CRC), yet imaging and routine pathology can yield false-negative/false-positive results. We propose DeepLNM, a multi-center foundation model for LNM prediction in CRC, evaluated with retrospective, external, prospective, and user study evidence. A multimodal dataset was assembled from three hospitals, with matched hematoxylin and eosin (H&E)-stained whole-slide images (WSI), CT scans, and text. DeepLNM achieved a mean AUC of 0.922 in the internal cohort and AUCs of 0.872 and 0.848 in two external cohorts. In a prospective cohort from a registered clinical trial (NCT07239063), DeepLNM remained robust (AUC = 0.845). In a user study, DeepLNM improved diagnostic accuracy, especially for junior pathologists (95.0%-97.3%), while reducing diagnostic time and increasing confidence. To explore the model's generalizability, it was further validated in a gastric cancer cohort, achieving an AUC of 0.793. Finally, DeepLNM identified histopathological features associated with LNM, addressing the "black-box" concern of deep learning. Together, these results position DeepLNM as a clinically preliminarily validated multimodal model for predicting LNM status in CRC, potentially facilitating the clinical translation of AI pathology.

Patients with resected pancreatic ductal adenocarcinoma (PDAC) can have different outcomes despite similar TNM stage. We developed TIGERS (Tumor Intelligence Grading and Evaluation Risk Score), a nine-feature prognostic biomarker derived from routine hematoxylin and eosin (H&E) whole-slide images. TIGERS quantifies tumor burden, glandular and nuclear morphology, chromatin texture, and tumor-stroma topology. The score was trained in 644 treatment-naive cases and evaluated in an internal validation cohort (n = 385), a separate post-neoadjuvant cohort (n = 275), two external cohorts (n = 260 and n = 235), the TCGA-PAAD cohort (n = 133), and a prospectively accrued external cohort (n = 155) with subsequent offline inference. Overall survival was the primary endpoint; the same formula and locked cutoff were applied throughout. The cutoff stratified survival in the internal validation cohort and three of four retrospective test cohorts. In the remaining external cohort, 226 of 235 patients (96.2%) were classified as high risk, showing limited transportability of the fixed threshold. Adding TIGERS to the clinicopathologic model increased Harrell's C-index in all six retrospective cohorts (range 0.015-0.047). In the prospectively accrued cohort, mortality was 47.5% in the high-risk group and 22.7% in the low-risk group; high risk remained associated with worse survival after adjustment (hazard ratio, 4.45; 95% confidence interval, 2.28-8.70; P < .001). Automated quality control passed 92.3% of screened cases, and scoring was completed in all included cases with a median turnaround time of 13.5 minutes. TIGERS assigned 43.2% of patients to a different risk category than the prespecified stage-based definition. Higher TIGERS correlated with a more basal-like proteomic profile in both the training and internal validation cohorts, supporting a molecular basis for its prognostic value. Routine H&E histology may therefore provide prognostic information complementary to conventional clinicopathologic assessment. Prospective interventional studies are needed before TIGERS can guide treatment or surveillance.

This retrospective two-center diagnostic study was aimed at developing and externally testing an interpretable MRI-based machine learning model for differentiating endometrial stromal sarcoma (ESS) from cellular leiomyoma (CL). A total of 155 women with pathologically confirmed ESS or CL were assigned to a derivation cohort (n = 111) and an external testing cohort (n = 44). Clinical variables, qualitative MRI features, and apparent diffusion coefficient (ADC) histogram parameters were evaluated. Three MRI-derived predictors-irregular tumor margin, cystic areas on T2-weighted imaging, and mean ADC-were used to train six classifiers. The final support vector machine model with a radial basis function kernel was evaluated at a fixed threshold determined in the derivation cohort using receiver operating characteristic, calibration, decision curve, and Shapley additive explanations (SHAP) analyses. A reader study assessed the effect of model assistance. The model achieved areas under the curve of 0.843 (95% CI, 0.760-0.917) in the derivation cohort and 0.835 (95% CI, 0.669-0.960) in the external testing cohort. Calibration and decision curve analyses indicated acceptable performance and potential clinical utility. In an exploratory model-assisted reader analysis, junior radiologist specificity increased from 46.4 to 82.1% and accuracy from 52.3 to 77.3%. This interpretable three-feature model demonstrated promising performance across cohorts and may serve as an adjunctive tool for preoperative MRI assessment, although prospective multicenter validation is warranted.

Knee osteoarthritis (OA) causes symptoms such as pain, stiffness, and functional limitation in the knees; plain knee radiographs are used in addition to physical examination in its diagnosis. This study aimed to compare the performance of pre-trained deep learning models in knee OA classification and to examine the effects of lateral knee radiographs and patient clinical data on the performance of these models. Eight centers participated in the study; radiographs from six centers were used for training, validation, and testing of the models, and radiographs from two centers were used for external validation. In addition to anteroposterior (AP) and lateral knee radiographs, clinical information such as age, gender, height, and weight was collected from the patients. A total of 1550 patients were included in the study. The performance of deep learning models VGG-16, ResNet-50, DenseNet-121, Inception_v3, and EfficientNetb0 was compared in the classification of knee OA. The VGG-16 model yielded significantly higher performance metrics than competing architectures (p < 0.001), with an accuracy of 0.73, sensitivity of 0.72, specificity of 0.93, F1 score of 0.72, positive predictive value (PPV) of 0.72, quadratic weighted Kappa of 0.91, and AUC of 0.93. Feature extraction was performed using the VGG-16 model, and the patient's age, gender, and BMI clinical characteristics were added to the image feature set and classified using machine learning algorithms. The XGBoost classifier showed statistically insignificant improvements of 1-2% in performance metrics. Similar performance results were obtained with the VGG-16 model in a study using only AP knee radiographs. Radiographs from two centers were used for external validation, and the VGG-16 model yielded accuracy of 0.65, sensitivity of 0.65, specificity of 0.91, PPV of 0.67, F1 score of 0.65, quadratic weighted kappa of 0.88, and AUC of 0.90. Adding age, gender, and BMI clinical characteristics to the knee radiograph image feature set did not significantly improve model performance. Furthermore, adding lateral radiographs to the AP knee radiograph set did not provide any additional contribution to KL classification performance.

This study validates a deep learning model for automated ureteral length measurement on CT using gross specimen measurements as the reference standard. A deep learning model was trained on 3629 CT series from 1502 patients (cohort 1) to segment the ureter and extract centerline length. An independent prospective cohort (cohort 2) of 70 patients undergoing radical nephroureterectomy served as the hold-out test set. Each gross specimen was measured twice by one urologist (D1_length and D2_length), with their mean as the reference length (Ref_length). Model-predicted lengths (AI_length) were compared with Ref_length using repeated-measures one-way ANOVA with Tukey post hoc tests. Determinants of discrepancy were assessed with Spearman correlation and multiple linear regression. Marginal means (± SE) were 266 ± 3.85 mm (D1_length), 273 ± 3.25 mm (D2_length), 269 ± 2.75 mm (AI_length), and 269 ± 3.45 mm (Ref_length). Measurement method had a statistically significant but negligible effect on length (F(3, 207) = 2.73, p = 0.045, η2G = 0.007). D1_length and D2_length differed significantly from Ref_length (both p = 0.002), whereas AI_length was not statistically different from Ref_length (p > 0.999). In regression analysis, reference ureteral length was the only independently associated variable (coefficient - 0.54 mm per mm of reference length, 95% CI - 0.70 to - 0.39; p < 0.001); age, sex, body height, body weight, operation side, and lesion location were not significantly associated. The proposed deep learning pipeline achieved accurate fully automated ureteral segmentation and length measurement across diverse CT acquisition phases and demonstrated close agreement with surgical gross specimen measurements in a prospective clinical cohort, supporting its potential to assist preoperative planning and reduce reliance on manual measurement. The measurement discrepancy was governed primarily by ureteral length itself-with a small systematic underestimation in longer ureters-rather than by patient or procedural factors. However, this tendency, together with the single-center design, and the modest size of the prospective cohort, warrants further multicenter validation before routine clinical deployment.