Prediction of impacts on liver enzymes from the exposure of low-dose medical radiations through artificial intelligence algorithms

ABSTRACT

SUMMARY OBJECTIVES: This study aimed to develop artificial intelligence and machine learning-based models to predict alterations in liver enzymes from the exposure of low annual average effective doses in radiology and nuclear medicine personnel of Institute of Nuclear Medicine and Oncology Hospital. METHODS: Ninety workers from the Radiology and Nuclear Medicine departments were included. A high-capacity thermoluminescent was used for annual average effective radiation dose measurements. The liver function tests were conducted for all subjects and controls. Three supervised learning models (multilayer precentron; logistic regression; and random forest) were applied and cross-validated to predict any alteration in liver enzymes. The t-test was applied to see if subjects and controls were significantly different in liver function tests. RESULTS: The annual average effective doses were in the range of 0.07-1.15 mSv. Alanine transaminase was 50% high and aspartate transaminase was 20% high in radiation workers. There existed a significant difference (p=0.0008) in Alanine-aminotransferase between radiation-exposed and radiation-unexposed workers. Random forest model achieved 90-96.6% accuracies in Alanine-aminotransferase and Aspartate-aminotransferase predictions. The second best classifier model was the Multilayer perceptron (65.5-80% accuracies). CONCLUSION: As there is a need of regular monitoring of hepatic function in radiation-exposed people, our artificial intelligence-based predicting model random forest is proved accurate in prediagnosing alterations in liver enzymes.