Early intestinal resection in patients with Crohn’s disease (CD) is necessary due to a severe and complicating disease course. Herein, we aim to predict which patients with CD need early intestinal resection within 3 years of diagnosis, according to a tree-based machine learning technique. The single-nucleotide polymorphism (SNP) genotype data for 337 CD patients recruited from 15 hospitals were typed using the Korea Biobank Array. For external validation, an additional 126 CD patients were genotyped. The predictive model was trained using the 102 candidate SNPs and seven sets of clinical information (age, sex, cigarette smoking, disease location, disease behavior, upper gastrointestinal involvement, and perianal disease) by employing a tree-based machine learning method (CatBoost). The importance of each feature was measured using the Shapley Additive Explanations (SHAP) model. The final model comprised two clinical parameters (age and disease behavior) and four SNPs (rs28785174, rs60532570, rs13056955, and rs7660164). The combined clinical–genetic model predicted early surgery more accurately than a clinical-only model in both internal (area under the receiver operating characteristic (AUROC), 0.878 vs. 0.782; n = 51; p < 0.001) and external validation (AUROC, 0.836 vs. 0.805; n = 126; p < 0.001). Identification of genetic polymorphisms and clinical features enhanced the prediction of early intestinal resection in patients with CD.

Background/Aims: We aimed to develop a deep learning model for the prediction of the risk of advanced colorectal neoplasia (ACRN) in asymptomatic adults, based on which colorectal cancer screening could be customized. Methods: We collected data on 26 clinical and laboratory parameters, including age, sex, smoking status, body mass index, complete blood count, blood chemistry, and tumor marker, from 70,336 first-time colonoscopy screening recipients. For reference, we used a logistic regression (LR) model with nine variables manually selected from the 26 variables. A deep neural network (DNN) model was developed using all 26 variables. The area under the receiver operating characteristic curve (AUC), sensitivity, and specificity of the models were compared in a randomly split validation group. Results: In comparison with the LR model (AUC, 0.724; 95% confidence interval [CI], 0.684 to 0.765), the DNN model (AUC, 0.760; 95% CI, 0.724 to 0.795) demonstrated significantly improved performance with respect to the prediction of ACRN (p < 0.001). At a sensitivity of 90%, the specificity significantly increased with the application of the DNN model (41.0%) in comparison with the LR model (26.5%) (p < 0.001), indicating that the colonoscopy workload required to detect the same number of ACRNs could be reduced by 20%. Conclusions: The application of DNN to big clinical data could significantly improve the prediction of ACRNs in comparison with the LR model, potentially realizing further customization by utilizing large quantities and various types of biomedical information.

To train an automatic brain tumor segmentation model, a large amount of data is required. In this paper, we proposed a strategy to overcome the limited amount of clinically collected magnetic resonance image (MRI) data regarding meningiomas by pre-training a model using a larger public dataset of MRIs of gliomas and augmenting our meningioma training set with normal brain MRIs. Pre-operative MRIs of 91 meningioma patients (171 MRIs) and 10 non-meningioma patients (normal brains) were collected between 2016 and 2019. Three-dimensional (3D) U-Net was used as the base architecture. The model was pre-trained with BraTS 2019 data, then fine-tuned with our datasets consisting of 154 meningioma MRIs and 10 normal brain MRIs. To increase the utility of the normal brain MRIs, a novel balanced Dice loss (BDL) function was used instead of the conventional soft Dice loss function. The model performance was evaluated using the Dice scores across the remaining 17 meningioma MRIs. The segmentation performance of the model was sequentially improved via the pre-training and inclusion of normal brain images. The Dice scores improved from 0.72 to 0.76 when the model was pre-trained. The inclusion of normal brain MRIs to fine-tune the model improved the Dice score; it increased to 0.79. When employing BDL as the loss function, the Dice score reached 0.84. The proposed learning strategy for U-net showed potential for use in segmenting meningioma lesions.