Mp63-17 Use of Artificial Intelligence in the Diagnosis and Treatment of Recurrent Nephrolithiasis

You have accessJournal of UrologyStone Disease: Basic Research & Pathophysiology (MP63)1 May 2024MP63-17 USE OF ARTIFICIAL INTELLIGENCE IN THE DIAGNOSIS AND TREATMENT OF RECURRENT NEPHROLITHIASIS Eri Osta, Ramiro Ramirez, Jakson Alan Johnson, Yusuf A. Khan, Geeta A. Joshi, Timothy Tseng, and Ronald Rodriguez Eri OstaEri Osta , Ramiro RamirezRamiro Ramirez , Jakson Alan JohnsonJakson Alan Johnson , Yusuf A. KhanYusuf A. Khan , Geeta A. JoshiGeeta A. Joshi , Timothy TsengTimothy Tseng , and Ronald RodriguezRonald Rodriguez View All Author Informationhttps://doi.org/10.1097/01.JU.0001009436.52988.91.17AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Diagnosis of stone composition is typically performed by pathologic analysis, often augmented by 24-hour urine analysis for urinary minerals and metabolites. The process is tedious, expensive, and requires referral to a specialist. We show that Artificial Intelligence (AI) can be a diagnostic tool to predict stone composition with simple clinical information (age, gender, urinalysis, metabolic panel, CBC) obtained at presentation. METHODS: We analyzed 455 patients with acute stone disease necessitating surgical intervention and correlated their clinical data with the stone composition.. Stone parameters included Calcium Oxalate (CO) Monohydrate, Calcium Oxalate Dihydrate, Calcium Phosphate (CP) Brushite, Calcium Phosphate Hydroxyapatite, Uric Acid (UA), Cystine (CYS), and Magnesium Ammonium Phosphate (MAP). RESULTS: Of the theoretical 62 possible combinations of these stone compositions, only 14 were observed. Three multiclassification models were created: a three-classification model (mono, dual, or multicomponent stone), a five-classification model (CO, CP, CO+CP, MAP, UA ), and a full 14-classification model. An XGBoost machine learning (ML) algorithm was used to predict the stone's classification. Our model achieved a balanced accuracy (BA) of 62% in the three-classification test, 74.5% BA in the five-classification test, and 53% BA in the 14-classification test. For the 14-classification test, some combinations had perfect prediction with an AUC in ROC analysis of 100% (CO+MAP, CYS, UA+MAP), while a couple had very poor AUC of between 7 and 22% (MAP, CO+CP+MAP), and the remainder had AUC of .58-0.99. The balanced accuracy was substantially higher than anticipated (roughly 8-fold over random chance). CONCLUSIONS: All the information typically obtained during an emergency department visit for acute nephrolithiasis is sufficient to predict the functionally essential components of the presenting stone, permitting highly individualized recommendations for treatment. Incorporating generative AI can create a detailed medical treatment plan and provide individualized patient discharge instructions tailored to their specific needs. Additional fine-tuning of the model will be performed to fully optimize the performance, including the incorporation of BMI, history of intestinal surgery, and ML analysis of CT scan imaging of the stones. Source of Funding: Abraham and Linda Littenberg FoundationIRACDA/SABER © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e1038 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Eri Osta More articles by this author Ramiro Ramirez More articles by this author Jakson Alan Johnson More articles by this author Yusuf A. Khan More articles by this author Geeta A. Joshi More articles by this author Timothy Tseng More articles by this author Ronald Rodriguez More articles by this author Expand All Advertisement PDF downloadLoading ...