SAT-084 USE OF AN AI ALGORITHM AND MACHINE LEARNING FOR SCREENING AND EARLY DETECTION OF CHRONIC KIDNEY DISEASE

      Introduction

      Globally 840 million people suffer from chronic kidney disease (CKD). However, about 96% are unaware of the diagnosis due to the initial asymptomatic disease. Current systems are expensive or incapable of early detection of CKD. And most often the diagnosis is made after considerable renal parenchymal damage. The “Healium” pilot program is aimed at addressing this gap in early detection and screening of current CKD care. The objective of this pilot study was to develop a machine learning (ML) algorithm for detecting CKD with a limited data set from routine laboratory tests, in a low resource setting and cross-validating the accuracy with the gold standard using a test and training data set.

      Methods

      Customized software programs and clinical variables to detect CKD were programmed with machine learning (ML) developed to integrate within hospital electronic health records (EHR) and the laboratory Information system (LIS), allowing real-time flagging of CKD from lab reports and ultrasound scans during unrelated hospital visit or annual medical evaluations. Ethical clearance was obtained for the study from the hospital ethics committee.
      An initial data set was used to train our AI algorithms. We did data cleaning, feature selection, data manipulation and correlation detection from the data. Pearson’s correlation, Cramer’s V, and ANOVA tests have been used to find relationships and dependency between variables. Then an algorithm modelling and selection of the best ML method for accurate detection of CKD was performed. In the modeling stage, five machine learning algorithms were applied to the dataset to assess their ability to detect CKD which included logistic regression (LR), support vector machines (SVM), ensemble methods and neural network. At the training evaluation each model generates different outputs depending on the different values of its hyper-parameters. LR Sensitivity and specificity was found most accurate and hence it was used for inferences in the test study.
      Test data included 32 patients (unlabelled and anonymised) uploaded to Healium AI with basic parameters (without albumin or serum creatinine data) and analysed by the Healium Algorithm (proprietary LR model) and results were recorded. Similarly, gold standard eGFR calculation all the patients were analysed by nephrologist and results were recorded.

      Results

      A total of 32 patients were screened using the system. Using the conventional eGFR method, 26 patients were detected CKD and 6 patients were detected non-CKD. While the Healium AI algorithm detected 24 positive CKD and 8 non-CKD patients from among the same screened patients. AI System has shown 92% of sensitivity and 75% specificity for detecting CKD in routine unrelated hospital visits.

      Conclusions

      We have demonstrated efficacy in overcoming the large gap addressing early detection and screening of current CKD care from routine hospital visits. This has encouraging implications for use across all hospitals with minimal EHR or LIS in low resource settings for early diagnosis and prevention of progression of renal disease.