Personalized Prediction of Patient Radiation Exposure for Therapy of Urolithiasis: An Application and Comparison of Six Machine Learning Algorithms.

Abstract

The prediction of radiation exposure is an important tool for the choice of therapy modality and becomes, as a component of patient-informed consent, increasingly important for both surgeon and patient. The final goal is the implementation of a trained and tested machine learning model in a real-time computer system allowing the surgeon and patient to better assess patient's personal radiation risk. In summary, 995 patients with ureterorenoscopy over a period from May 2016 to December 2019 were included. According to the suggestions based on actual literature evidence, dose area product (DAP) was categorized into 'low doses' ≤ 2.8 Gy·cm2 and 'high doses' > 2.8 Gy·cm2 for ureterorenoscopy (URS). To forecast the level of radiation exposure during treatment, six different machine learning models were trained, and 10-fold crossvalidated and their model performances evaluated in training and independent test samples. The negative predictive value for low DAP during ureterorenoscopy was 94% (95% CI: 92-96%). Factors influencing the radiation exposure were: age (p = 0.0002), gender (p = 0.011), weight (p < 0.0001), stone size (p < 0.000001), surgeon experience (p = 0.039), number of stones (p = 0.0007), stone density (p = 0.023), use of flexible endoscope (p < 0.0001) and preoperative stone position (p < 0.00001). The machine learning algorithm identified a subgroup of patients of 81% of the total sample, for which highly accurate predictions (94%) were possible allowing the surgeon to assess patient's personal radiation risk. Patients without prediction (19%), the medical expert can make decisions as usual. Next step will be the implementation of the trained model in real-time computer systems for clinical decision processes in daily practice.