OPTIMIZED PREDICTION MODEL OF DIABETIC PATIENTS WITH HIDDEN MARKOV MODELS

Abstract

The extraordinary advances in biotechnology and wellbeing sciences have prompted a unique creation of information, for example, high throughput genetic information and clinical data, produced from expansive Electronic Health Records (EHRs). To this end, utilization of machine learning and information mining techniques in biosciences are direct, like never before previously, essential and critical in endeavors to change brilliantly all accessible data into valuable information. Inthisresearchwork,wehaveusedanall-femaledataset[19]whichcompriseseightattributes, i.e., Pregnancies, Glucose, Blood Pressure, Skin Thickness, Insulin, BMI, Diabetes Pedigree Function, Age and two class labels, i.e., 1 or 0 which indicates if a person has diabetes or not. Firstly, we used supervised techniques on the given dataset and tested their performance, i.e., accuracy in the prediction of people having diabetes. The results for various supervised algorithms are: K-NN - 64.91%, SVM - 65.91%, Random Forest - 67.80%, Gaussian Naive Bayesian-64.36%,DecisionTree-63.86%,LogisticRegression-60.71%. Fromtheseresults, we conclude that these algorithms cannot predict with sufficient accuracy which stands to be at least 80%. The unique feature about the supervised technique is if these algorithms encounter a specific attribute which dominates our prediction models, these algorithms always comes out to be on top. The result above means that our dataset has many features essential for prediction. As a result, the supervised Hidden Markov Model as an alternate technique for prediction. In our prediction methodology, with the help of the Hidden Markov Model, we use Viterbi and BestFirst Search algorithms as decoding techniques. These decoding techniques come up with following subsequences after the Forward-Backward algorithm gives emission probabilities. These probabilities belong to different states of our data model. As a result of which learning by our model gets completed. We have split the dataset into two parts 90% of the dataset as a learning data and 10% of the dataset as the testing data. The accuracy of our model was found out to be 81.81% with Viterbi and 79.22% with BestFirst.Hence, supervised HMM performs better as compared to the supervised techniques used generally.