A Machine Learning Method for Differentiation Crohn's Disease and Intestinal Tuberculosis.

Area of Science:

• Medical Informatics
• Machine Learning in Medicine
• Diagnostic Accuracy

Background:

• Differentiating Crohn's disease (CD) and intestinal tuberculosis (ITB) is clinically challenging.
• Machine learning (ML) has potential for aiding in complex diagnoses.
• The diagnostic utility of ML for CD and ITB requires investigation.

Purpose of the Study:

• To evaluate the efficacy of ML models in distinguishing between CD and ITB.
• To compare ML model performance against traditional diagnostic methods.
• To identify key clinical features influencing ML-based differential diagnoses.

Main Methods:

• Collected clinical data from 241 patients with 51 parameters.
• Tested six ML algorithms: logistic regression, decision tree, k-nearest neighbor, multinomial NB, multilayer perceptron, and XGBoost.
• Utilized SHAP and LIME for model interpretability and validated performance in clinical practice against a multidisciplinary team (MDT).

Main Results:

• XGBoost demonstrated superior performance with an AUROC of 0.946 and accuracy of 0.884.
• Key predictors included T-spot, pulmonary tuberculosis, and age of onset.
• In clinical practice, the ML model achieved 0.860 accuracy, 0.833 sensitivity, and 0.871 specificity, with 90.7% agreement with MDT.

Conclusions:

• An ML model, particularly XGBoost, can effectively and efficiently aid in the differential diagnosis of ITB and CD.
• The developed ML model shows strong agreement with expert multidisciplinary team assessments.
• The model's performance in real-world clinical settings supports its utility for improving diagnostic accuracy.