You have accessJournal of UrologyStone Disease: Epidemiology & Evaluation II (MP54)1 Sep 2021MP54-19 MACHINE LEARNING MODELS TO PREDICT 24-HOUR URINE ABNORMALITIES FROM ELECTRONIC HEALTH RECORD-DERIVED FEATURES Nicholas Kavoussi, Abin Abraham, Wilson Sui, Cosmin Bejan, John Capra, and Ryan Hsi Nicholas KavoussiNicholas Kavoussi More articles by this author , Abin AbrahamAbin Abraham More articles by this author , Wilson SuiWilson Sui More articles by this author , Cosmin BejanCosmin Bejan More articles by this author , John CapraJohn Capra More articles by this author , and Ryan HsiRyan Hsi More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000002084.19AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: To enable earlier preventative interventions or guide empiric therapy for kidney stone disease, our objective was to demonstrate feasibility of predicting 24-hour urine abnormalities using several machine learning methods. METHODS: We trained machine learning models (XGBoost [XG] and Ensemble [EN]) to predict 24-hour urine abnormalities from electronic health record-derived demographic, laboratory, stone composition, and comorbidity data, including age, BMI, comorbidities (n=1,296). The machine learning models were compared to a logistic regression model [LR]. Models predicted binary (normal vs high) 24-hour urine values for sodium, oxalate, calcium, and uric acid; and (normal vs low) for citrate; as well as a multiclass prediction of pH (low, normal, high). We evaluated performance using area under the receiver operating curve (ROC-AUC) and accuracy and identified predictors for each task. RESULTS: Both XG and EN were able to discriminate 24-hour urine abnormalities with fair performance, with comparable performance to LR (see Figure). EN most accurately predicted abnormalities of oxalate (accuracy=65%, ROC-AUC=0.70) and citrate (65%, 0.69), while XG most accurately predicted abnormalities of uric acid (69%, 0.73) and sodium (71%, 0.70). Both models had similar accuracy for the prediction of pH (45%, 0.57) and calcium (55%, 0.59) abnormalities. Body mass index, age, and gender were the three most important features for training the models for all outcomes. CONCLUSIONS: Urine chemistry prediction for kidney stone disease appears to be feasible with machine learning methods with acceptable discrimination for several urinary parameters modifiable with diet. Further optimization of the performance could facilitate earlier pharmacologic preventative therapy. Source of Funding: none © 2021 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 206Issue Supplement 3September 2021Page: e957-e957 Advertisement Copyright & Permissions© 2021 by American Urological Association Education and Research, Inc.MetricsAuthor Information Nicholas Kavoussi More articles by this author Abin Abraham More articles by this author Wilson Sui More articles by this author Cosmin Bejan More articles by this author John Capra More articles by this author Ryan Hsi More articles by this author Expand All Advertisement Loading ...
Read full abstract