Description

AI System for Prediction and Recommendation of Diabetes

Evidence-based medicine is a powerful tool to help minimize treatment variation and unexpected costs. Large amount of healthcare data such as physician notes, medical history, medical prescription, lab and scan reports generated is useless until there is a proper method to process this data interactively in real – time. In this world filled with the latest technology, healthcare professionals feel more comfortable to utilize the social network to treat their patients effectively. To achieve this, an effective framework is needed which is capable of handling large amount of structured, unstructured and live streaming data about the patients from their social network activities. Healthcare Recommendation System (HRS) using machine learning can be developed to predict about the health condition by analyzing patient’s life style, physical health factors, mental health factors and their social network activities. For example, on training the model with the age of women and diabetes condition helps to predict the chances of getting diabetes for new women patients without detailed diagnosis.

Methodology
Healthcare Recommender System (HRS) is designed for prediction of diseases and recommendation of treatment. It depends on a set of patient’s case history, expert rules and social media data to train and build a model that can predict and recommend disease risk, diagnosis and alternative medicines. Predictions and recommendations need to be approved by doctors. HRS system requires input information to generate predictions and recommendations. In this work diabetes data will be used as case study.

Step 1: Data collection and dataset preparation

Medical case history of diabetes patients will be stored which will contain information like blood sugar, blood pressure, weight etc. Diagnosis data comprises of physician notes, lab results, and medications. The data records may have many attributes, values and doctor’s diagnosis for each case. Diagnosis scale ranges from 1 to 5 based on the severity of the disease, 5-represents critical condition, 4-represents severe requires immediate treatment, and 3- represents moderate requires further investigation, 2-represents normal, 1-represents within control. Along with this, demographic data of active patient like name, age, location, education
level, wearable device, lifestyle, food habits and type of connectivity will also be collected. A sample set of attributes of a patient are shown in figure 2.

Step 2: Developing a recommender system based on predictions using AI Prediction is expressed as a numerical value that represents the disease risk diagnosis for
future cases based on active patients. The rules for detecting various ranges for diabetes based on Fasting Plasma Glucose test (FPG), Casual Glucose Tolerance Test (CGTT) and Glycated Haemoglobin tests (HBA1C) are given below:

Rule 1: No Diabetes Range

If FPG has a level between 70 and 100 mg/dL (3.9 and 5.6 mmol/L), then it indicates- no diabetes range. If the blood glucose level below 125 mg/dL in CGTT, then it indicates- no diabetes range. If HBA1C value is below 97 mg/dL, then it indicates- no diabetes range.

Rule 2: Pre-diabetes Range

If FPG ranges from 100 mg/dl to 125 mg/dl and CGTT ranges from 140 mg/dl to 199 mg/d and  HBA1C test values lie in range 97-154 mg/dL, it indicates pre-diabetes range.

Rule 3: Diabetes Range

If FPG is 126 mg/dl or more and CGTT is 200 mg/dl or more and HBA1C is greater than 180 mg/dL, it indicates diabetes. The nest phase of recommendation which is expressed as the suggestion required by the users. For example, non-healthcare professional might be requiring alternative remedies for treating diabetes. The recommendation rules will be created by taking opinion from many doctors for different possible scenarios. Deep learning using CNN with auto-encoders or similar
will be exploited for this task

Step 3: Training and experimentation on datasets Diabetes data set can be taken from KN specialty clinic and downloaded from UCI repository. Step 4: Deployment and analysis on real life scenario The trained and tested recommender system will be developed in real-life scenario where historical medical records of diabetic patients will be collected from local hospitals. Experimental Design Dataset Diabetes Data Set (https://archive.ics.uci.edu/ml/datasets/diabetes) will be used for experimentation.