Interpretable Deep Learning for Drug-Induced Liver Injury among Patients with Tuberculosis: Model Development and Validation Study (Preprint)

BACKGROUND Drug-induced liver injury (DILI) is associated with treatment discontinuation and treatment failure in patients with tuberculosis (TB). Due to the unpredictable nature of DILI, it is increasingly becoming a threat to global TB eradication initiatives. OBJECTIVE This study aimed to develop and validate an interpretable prediction model for DILI during TB treatment. METHODS Eligible adult TB patients in Ningbo City from 2015 to 2020 were included. We used the XGBoost algorithm to predict the probability of DILI occurrence, compared with the LASSO-Logistic algorithm. HDPS method was applied for feature extraction. We assessed the performance by the stratified 10-fold cross validation and the AUROC. The SHAP value was used to interpret the XGBoost model of mild and more severe DILI patients. RESULTS We included 7,071 subjects (median age 47 years; 68.0% male). 16.3% developed DILI. The XGBoost algorithm yielded the final DILI prediction model with five main risk factors: previous history of DILI, ALP, TBL, ALT and age at TB onset. It outperformed the LASSO-Logistic model in AUROC (0.90 vs. 0.76). The SHAP value showed that older patients and higher ALP, ALT and TBL were more likely to develop more severe DILI, However, previous DILI experience is inversely related to the severity of DILI. CONCLUSIONS XGBoost model demonstrated high accuracy and interpretability in predicting DILI among TB patients. Clinicians could leverage algorithmic technique and EHR data to diagnose and manage DILI during TB treatment promptly.