A recent study published in Diagnostics applied artificial intelligence (AI) to forecast the prognosis and adverse effects of tuberculosis (TB).
TB is an infectious disease and a significant cause of global morbidity and mortality. Infected patients can be treated, and the drug dose/regimen and treatment duration depends on comorbidities, infection site, and strain type.
Most TB medicines can be toxic to the liver; therefore, physicians must monitor liver enzymes and assess the risk of hepatitis. Lately, AI and machine learning (ML) models have been employed to diagnose TB, but fewer studies have used them to predict adverse outcomes.
About the study
In the present study, researchers used AI/ML models to predict outcomes in TB patients. They collected data from TB patients from three hospitals in Taiwan between January 2004 and December 2021. Data from patients below 20 years at diagnosis and those with non-tuberculous mycobacteria were excluded.
Acute hepatitis, respiratory failure, and all-cause death during treatment were the outcomes. The team included feature variables such as age, sex, TB type, and comorbidities. All variables were used to build prediction models to ensure maximum performance. Data were randomized into training and testing datasets. The synthetic minority oversampling technique (SMOTE) was applied to fix data imbalance.
Models for each outcome were generated using six ML algorithms – random forest, XGBoost, support vector machine (SVM), multilayer perceptron (MLP), light gradient boosting machine (LightGBM), and logistic regression. The testing dataset was used to evaluate models with indicators of accuracy, sensitivity, specificity, and the area under the receiver operating characteristic (AUROC) curve.
The authors identified 4,018 cases during the study period. After exclusions, 2,248 patients were selected for model building. Most subjects were males (71.7%); the mean age of participants was 67.7. The Spearman correlation analysis identified serum alanine aminotransferase, aspartate aminotransferase, and total bilirubin levels as relevant features for acute hepatitis, and blood urea nitrogen, age, and white blood cell (WBC) count for acute respiratory failure and mortality.
The MLP algorithm achieved the highest AUROC value of 0.834 for predicting mortality with a sensitivity of 0.722, specificity of 0.736, and accuracy of 0.735. Random forest had the highest value of 0.884 for acute respiratory failure prediction with a sensitivity of 0.812, specificity of 0.82, and accuracy of 0.819. XGBoost showed the highest AUROC value of 0.92 for acute hepatitis; sensitivity, specificity, and accuracy were 0.77, 0.92, and 0.86, respectively.
Taken together, the researchers applied AI/ML models for early detection of respiratory failure, hepatitis, and death in TB patients, using commonly available clinical and demographic data. Notably, the sample comprised patients from southern Taiwan, limiting the representativeness of the findings. Moreover, alcohol/smoking status was not available, given the retrospective data collection method.