P08-01

Predicting clinical laboratory test result related to urine tests in patients with type 2 diabetes mellitus with renal complications using clinical trial data

Hiroki ADACHI *, Takuya SEKIYAMA, Taku SAKAUE, Yasuo SUGITANI

Biometrics Department, Chugai Pharmaceutical Co., Ltd. 
( * E-mail: adachi.hiroki16@chugai-pharm.co.jp )

Clinical laboratory tests are medical tests that analyze a patient's blood, urine etc., and use the results to understand the patient's characteristics and medical condition. Clinical laboratory test results are an important tool for improving the quality of medical care, as they are used to detect diseases at an early stage, determine the progression of a disease, and even determine the effectiveness of treatment. We focused on the prediction of urinalysis values that are rarely measured in routine medical care. It is known that patients with type 2 diabetes mellitus with renal complications have higher urinalysis scores due to decreased renal function. We created prediction models using urinalysis scores from clinical trial data of patients with type 2 diabetes mellitus with renal complications conducted in the past as the objective variable and the results of key laboratory tests as explanatory variables to evaluate the prediction accuracy and improve prediction accuracy. We selected clinical trials that included patients with type 2 diabetes mellitus with renal complications. We were able to create a prediction model for both multilevel and binary classification of urinary protein in a specific study that included patients with type 2 diabetes mellitus with renal complications. For multilevel classification of urine protein scores, a model using Random Forest was created and an accuracy of AUC 0.8090 obtained. For binary classification, a model using Support Vector Classifier was created and an accuracy of AUC 0.8974 was obtained. The clinical studies are designed to have a regular schedule of patient laboratory tests. Compared to actual medical care, it is thought that the tests are performed more rigorously and that there are fewer missing values. We believe that it is necessary to verify the applicability of the predictive model based on the data in actual clinical practice, and we plan to conduct such verification using actual clinical data.