Introduction. Chemical composition of the urinary stone, that is, of its metabolic type, are of great importance for the choice of both invasive methods of treatment of urolithiasis and methods of stone recurrence prevention. However, urinary stones are not always available for analysis, which makes it difficult to select the best treatment methods and leads us to the need to find methods for assessing the chemical composition of patient urinary stones in a patient in vivo. One of the search directions is a comprehensive analysis of lithogenic metabolic factors, the long-term effect of which leads to the urinary stone formation.
Goal. The purpose of this paper is to evaluate the possibility of some machine learning algorithms in predicting the metabolic type of urinary calculi in vivo in a set of metabolic parameters.
Material and methods. The mineral composition of 708 urinary calculi (from 305 men and 403 women aged 16 to 81 years), as well as biochemical parameters of blood serum (calcium, uric acid, phosphates, magnesium) and daily excretion in urine of calcium, uric acid, phosphates , magnesium. The specific gravity of urine, its pH, daily volume was determined. The body mass index of the patients was calculated too.
Results and discussion. Using the IBM SPSS Modeler v18.0 (software for models building) and the prepared data set (n = 708), it was possible to determine that the C5.0 algorithm is the best for accurate prediction of the chemical type of urinary stone by metabolic parameters.
The coincidence of predictions made by the model (algorithm C5.0), in comparison with the actual distribution of types of stones, was almost absolute. Incorrect definition of the type of stone was observed only in 2 cases out of 708 (0.28%).
Further testing of the C5.0 model with separation of the main data set (n = 708) into two randomized samples – Teaching (70%, n = 499) and Testing (30%, n = 209), marked the high quality of the model prediction. In the Testing sample the correct definition of the type of stone was observed in 207 cases out of 209 (99.4%).
The prediction accuracy of other machine learning models in classification of urinary stone types was significantly lower (45- 80%), which indicates significant advantages of the model, built on C5.0 algorithm. The same model showed the highest percentage of correct stone type recognition (100%), while in other machine learning models the correct predictions of types of stones were 34.8-90.2%.
For practical clinical use of C5.0 model, there is an opportunity to build a diagnostic decision tree or create a set of decision rules that allow perform a model forecast of the metabolic type of stone in vivo.
Conclusion. The proposed machine learning model based on the C5.0 algorithm and using the metabolic parameters of urine and blood of a patient with urolithiasis can be used in the clinic for sufficiently reliable diagnostic prediction in vivo the chemical composition of urinary stones of the most common metabolic types. This model can be used for evaluation the urinary stone chemical composition before performing lithotripsy, as well as monitoring patients after removing stones from them in order to predict risk of forming a specific metabolic type stone and timely to initiate targeted metaphylaxis.
Authors declare lack of the possible conflicts of interests.