P0377 Development of Crohn’s Aid mobile application based on AI algorithms to make an early diagnosis of Crohns Disease in TB endemic regions

Abstract Background Intestinal tuberculosis (ITB) and Crohn’s disease (CD) mimic each other clinically, endoscopically, and radiologically and are difficult to differentiate. This is a major barrier to make an early diagnosis of CD in TB endemic areas. We aimed to develop a high accuracy machine learning-based model which could be easily implemented in an innovative way. Methods We retrospectively analyzed data from 1066 patients. After data cleaning, 796 patients (514 CD & 282 ITB) and data from 28 variables were included. A super learner approach using random forest and support vector machine was used for differentiating ITB from CD. Data was divided into 80 percent training and 20 percent testing data sets followed by 10-fold cross validation. The optimal cut-off for the diagnosis was obtained using the Youden index-measure to optimize the balance between sensitivity and specificity and the model was evaluated at multiple thresholds for clinical utility. The best performing model was incorporated into a mobile phone-based application. Prospective validation on 37 patients was carried out with similar accuracy. Results The random forest model achieved a sensitivity, specificity and accuracy of 0.92, 0.83, 0.86 respectively and performed better than the support vector machine model trained with linear and radial basis functions. The random forest model was found to have the best AUROC with a cutoff of 0.4, predicting the diagnosis of CD with a sensitivity of 93%, specificity of 83%, and accuracy of 86%, positive predictive value of 76%, and negative predictive value of 95%. The random forest model was used for creation of the application. The app is being made available on smartphones free of cost for use by any physician. Conclusion Our model differentiated between ITB and CD with high accuracy and has the potential to makean early diagnosis of CD. The free to use mobile application would make implementation of thisalgorithm much easier, allowing for widespread use in clinical practice and helping make moreinformed decisions. More data from multiple centers and different geographical locationswould aid in further improving the model performance. Figure 1: Overall model training approach Figure 2: Model performance and selection of appropriate cut-off