Development and Prospective Validation of a Machine Learning‐Based Mobile Application (CROHN'S AID ) for Differentiating Crohn's Disease From Intestinal Tuberculosis in Tuberculosis‐Endemic Regions

ABSTRACT Background Differentiating Intestinal tuberculosis (ITB) from Crohn's disease is a major diagnostic challenge in endemic regions, often leading to inappropriate treatment. We aimed to develop and prospectively validate a machine learning‐based mobile application, “Crohn's Aid”, to provide an accurate, real‐time clinical decision support tool at the point of care. Methods Retrospective analysis of a prospectively maintained database of 1066 patients (CD = 650; ITB = 416) from a tertiary centre in India. Thirty clinical, endoscopic and radiological variables were used to train multiple machine learning models, including CatBoost, Random forest and logistic regression. The lead model (CatBoost) was integrated into a mobile application. Performance was evaluated using area under the receiver operating characteristic (AUROC) and prospectively validated on a cohort of patients. Explainable AI (SHAP) values were used to identify key diagnostic drivers. The model was then validated on a prospective validation cohort of 121 patients at the same centre. Results Internal cross‐validation was repeated ten times, and the CatBoost model achieved a mean AUROC of 0.886 with a mean sensitivity of 76.0% and specificity of 84.1%. In the prospective validation cohort ( n = 121), the model maintained high diagnostic accuracy—AUROC 0.921(0.870–0.963) and outperformed the standard Bayesian reference standard (AUROC 0.750). At the optimum threshold, the model demonstrated a balanced sensitivity and specificity of 86.4% and 85.5%, respectively. Key predictors included symptom duration, transverse ulcers, granuloma, rectosigmoid involvement, and perianal disease. Conclusion Crohn's Aid represents a step forward in the application of digital health to solve a long‐standing clinical dilemma by combining the power of gradient‐boosted decision trees with the ubiquity of smartphone technology to reduce misdiagnosis between ITB and CD in resource‐limited settings.