Sequence Rule Mining for Insulin Dose Prediction Using Temporal Dataset

Abstract

The objective of pattern mining is to mine sensitive and non-critical information hidden from databases. Sequential rule mining is a pattern mining technique that retrieves rules in order. It has number of applications in different area including healthcare. Healthcare is the prime concern for human’s life. Insulin dependent diabetes mellitus (IDDM), which is also known as type-1 diabetes is a kind of chronic disease and patients of these disease needs to get insulin as a supplement or medicine. Technologies that predict the right amount of insulin dose for a patient is needed for correct prognosis of patients. It facilitates to better health in patients to maintain the glucose level in a specific range which produces the right energy for human cells. Our objective is to find insulin therapy plan for diabetes patients. In this work, we applied some sequential mining algorithms to predict the dosage values of regular, NPH and ultraLente insulin at three different time frames, i.e., at prebreakfast, prelunch, and presupper. Sequential rules are generated to predict the dosage ranges for different time frames with support and confidence metrics. These rules are used to recommend particular type of insulin with dosage amounts. This work also compares different sequential rule mining algorithms such as ERMiner, CMRules, CMDeo, and RuleGrowth with their memory taken and execution speeds. We concluded that by making sequences of blood glucose ranges with insulin dose ranges we can generate rules for insulin prediction as well as diabetes predictions. We can also conclude that after applying different sequential rule mining algorithms, ERMiner is faster among the algorithms but it takes more memory to execute.