Machine Learning-Based Prediction of Pediatric Ulcerative Colitis Treatment Response using Diagnostic Histopathology

ABSTRACT Background and Aims We previously reported clinical features associated with outcomes in pediatric ulcerative colitis (UC). Here we developed a histopathology model to predict corticosteroid-free remission (CSFR) on mesalamine therapy alone. Methods Pre-treatment rectal biopsy slides were digitized in training and validation groups of 292 and 113 pediatric UC patients, respectively. Whole slide images (WSI) underwent pre-processing. Thirteen machine learning (ML) models were trained using 250 histomic features including texture, color, histogram, and nuclei. Feature importance was determined by the Gini index with the classifier re-trained using the top features. Results 187571 informative patches from 292 training group patients (Male:53%; Age:13y (IQR:11-15); CSFR:41%) were trained on 13 ML classifiers. The best model was random forest (RF). Eighteen optimal histomic features were identified and trained, and the corresponding WSI AUROC was 0.89 (95%CI:0.71, 0.96), accuracy of 90% for CSFR. Features were re-trained on an independent real-world dataset of 113 patients and the model WSI AUROC was 0.85 (95%CI:0.75, 0.95), accuracy of 85%. Conclusion Routine histopathology obtained at diagnosis contains histomic features associated with UC treatment response.