Wei Chen

Accurate prediction of recurrence risk after hepatectomy still remains a clinical challenge for hepatocellular carcinoma (HCC). We develop a deep learning radiopathomic (DLRP) signature to fuse deep features of CT images and histological whole-slide-image, aiming to predict the recurrence-free survival (RFS) in HCC patients. The multi-omics data in The Cancer Genome Atlas (TCGA) database are used to assess the potential biological interpretation. A total of 599 patients are enrolled in this study and divided into the training (n = 272), internal test (n = 120), external test (n = 174), and TCGA (n = 33) cohorts. The DLRP signature shows better prediction for RFS than radiomics signature, pathomics signature, clinical model, and Barcelona Clinic Liver Cancer stage in the external test cohort (C-index, 0.799 vs 0.541–0.738; P value range, P < 0.001). The multi-omics data indicate that DLRP signature is relevant to Wnt/β-catenin signaling pathway and tumor immune infiltration. We conclude that the DLRP signature can efficiently predict recurrence risk in HCC patients, thereby facilitating personalized precision therapy.

Objective: The implementation of comprehensive microinvasive cervical cancer (MIC) risk assessment in high-grade squamous intraepithelial lesion (HSIL) patients undergoing loop electrosurgical excision procedure (LEEP) is critical to optimize treatment strategies and improve patient outcomes. Methods: From March 2017 to January 2024, a total of 3066 eligible patients with HSIL were retrospectively enrolled from two hospitals and assigned into one training cohort (n = 2084), one internal validation cohort (579) and one external testing cohort (n = 403). Four feature selection methods (Random Forest, Lasso regression, Boruta algorithm, and Extreme Gradient Boosting) were employed to identify key predictive factors from the training cohort. Then, four machine learning models were developed and evaluated using comprehensive metrics. The optimal model was visualized through interpretable techniques and operationalized as a web-based clinical decision support system for real-world implementation. Results: Six clinical predictive variables were identified, including surgical margins, endocervical curettage (ECC), TCT status, HPV status, Transformation Zone (TZ) type and Age. The optimal model demonstrated good predictive performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.822 (95% CI: 0.793– 0.852) in the internal validation cohort and 0.802 (95% CI: 0.730– 0.874) in the external validation cohort. Conclusion: The machine learning-based model can accurately assess the risk of MIC during the treatment of HSIL with LEEP, potentially aiding in the selection of appropriate treatment and surveillance strategies in clinical practice.

Purpose: This study aimed to identify reliable risk factors for the development of Vaginal intraepithelial neoplasia (VaIN) within two years after the conization for high-grade squamous intraepithelial lesions (HSIL). We developed a prediction model to predict the risk of VaIN based on preoperative and follow-up data. Methods: We collected 5358 patients who underwent conization for HSIL, of whom 99 developed VaIN within two years after conization. We selected 495 patients as the control group by randomly pairing them 1:5, and were randomly divided into development and validation cohorts at a ratio 7:3. Random Forest (RF), Lasso, and Extreme Gradient Boosting (XGBoost) were employed to identify the most influential variables in the model development dataset. The optimal variables selected through this process were then used for model construction. Subsequently, four machine learning models were developed, and their performance was evaluated using metrics including sensitivity, specificity, accuracy, area under the curve (AUC), and the F1 score. To enhance interpretability, the prediction process was visualized using Shapley Additive Explanations (SHAP). Finally, the model was deployed as a web-based clinical decision support system for practical clinical applications. Results: Five key clinical predictive variables were identified: age, transformation zone (TZ) type, presence of VaIN before conization, follow-up cytology after conization, and follow-up HPV after conization. The optimal model demonstrated strong predictive performance, achieving AUC of 0.910 (95% CI: 0.854– 0.966) in the internal validation cohort and 0.905 (95% CI: 0.859– 0.951) in the external validation cohort. Conclusion: We established a practical and accurate prediction model deployed in the network application to predict the occurrence of VaIN within two years after conization in patients with HSIL. This tool can facilitate targeted clinical decision-making for clinicians.

Stroke-associated pneumonia (SAP) is a common and serious complication after stroke, which prolongs hospitalization, increases mortality and the difficulty of care. Currently, most of the prediction models for SAP are based on clinical data, such as patient's age, past medical history, and clinical scores, etc. Clinical data-based models have the limitations of relying too much on doc-tor's subjectivity, leading low accuracy, and is time-consuming for doctors. CT is a usual examination for stroke patients in hospital, and the prediction of SAP by imaging data has great scientific value. In this paper, deep fusion models are proposed to combine imaging data and clinical dat. Brain CT scans from 244 ICH patients (mean age, 60.24; 66 female) were divided into a training set (n=170) and a test set (n=74). The cohort included 143 SAP patients, accounting for 58.6% of the total, with 66 cases classified as moderate or above, representing 27% of the total. Compared with the model based on clinical data and the model based on imaging methodology, the proposed fusion model shows the best performance. The modes obtained AUC values of 0.83 and 0.87. The model is expected to be an effective tool for predicting and grading SAP in clinic.

Early detection of skin diseases is essential for improving patient survival rates and preventing serious harm to people's lives and health. In recent years, there have been advancements in the field of skin disease identification using artificial intelligence, particularly through algorithms like deep learning. Scientists have developed a series of skin disease diagnostic systems that can assist clinical doctors in quickly and effectively identifying and diagnosing different types of skin diseases. In this study, we selected and trained a deep learning network (MobileNetV2) and introduced a weighted cross-entropy loss function to address class imbalance in classification tasks, improving sensitivity to minority classes. Testing over 25,000 skin disease images, the results validated the effectiveness and feasibility of the proposed method, achieving an 84.7% classification accuracy on eight categories in the combined dataset, which is highly acceptable. This method is expected to help dermatologists make more accurate clinical diagnoses in the future, potentially reducing the number of skin disease patients.