Preoperative classification of urinary stones based on community detection.

In the treatment of urinary stones, surgical intervention is crucial. Urinary stones composition and type directly affect surgical planning. However, research on preoperative stone composition analysis is limited. This paper aimed to predict urinary stones types preoperatively using clinical data. Data from 1020 patients, including stone composition, clinical biochemical indicators, and demographic information, were collected. A stone composition graph network was constructed using cosine similarity, with stone composition as nodes and biochemical/demographic data as node features. The Louvain community detection algorithm was utilized to divide the network into distinct communities for the classification of stone types, with the effectiveness of the partitioning evaluated by the Modularity score. Stone types were classified, and their distribution across genders and age groups was described. Clinical feature averages were calculated for each community, and patients were assigned to the most similar community. Six machine learning algorithms (RandomForest, GradientBoosting, SVM, KNN, Logistic Regression, XGBoost) were trained to predict stone types. Model performance was evaluated, and the importance of clinical features for prediction was ranked. Six stone types were identified (Modularity = 0.828), namely common COM (Class I), COM with minor AU (Class II), COM with high UA (Class III), COM containing MAP (Class IV), high CAP-MAP (Class V), and high COM-CAP containing DCPD (Class VI). Among males, Class III and Class I were most prevalent; among females, Class V and Class III were most prevalent (χ