Interpretation of the prognosis of early results of nephron-sparing surgery with consideration of surgical learning curve using clinical decision support systems

Aim. To assess the possibility of interpreting machine learning models to predict the early results of laparoscopic nephron-sparing surgery (NSS) in kidney tumors with consideration of surgical learning curve.

Materials and methods. The results of 320 consecutive laparoscopic NSS in patients with localized kidney tumors, performed by 4 surgeons, were analyzed. The construction of a machine learning model taking into account surgical learning curve was carried out based on the extreme gradient boosting (eXtreme Gradient Boosting). To identify significant factors and interpret the prognostic ability of the model, the SHapley Additive exPlanations method was used with a calculation of the Shapley value. Three groups of factors were chosen as an array of input data. The first group included demographic and clinical characteristics of patients, such as age, gender, Charlson comorbidity index, body mass index, preoperative glomerular filtration rate (GFR). In the second group, there were morphometric indicators of the kidney tumor, including RENAL. Nephrometry Score, PADUA (Preoperative Aspects and Dimensions Used for an Anatomical), C-index (Centrality index score), absolute tumor volume, localization of the tumor in relation to the kidney surface. In addition, factors associated with surgical learning curve, such as case number and perioperative results last 10 procedures, were analyzed. The target variables were duration of the procedure, warm ischemia time, and postoperative GFR after 24 hours.

Results. The SHAP method allows a visual interpretation of a machine learning algorithm based on the extreme gradient boosting for individual prediction of early perioperative outcomes of laparoscopic NSS in patients with renal tumors. For the calculated new features “complexity”, “slope angle” and others using the SHAP method, the high significance in building predictive models for target variables was confirmed, and an interpretation of the influence of specific features on the target variable in the constructed machine learning models was also given.

Conclusion. The SHAP method showed good practical results that coincide with the observations of specialists. The use of such solutions will allow in the future to introduce machine learning models to form clinical decision support systems.