![\"Architecture\"](\"https:\/\/devblogs.microsoft.com\/cse\/wp-content\/uploads\/sites\/55\/2020\/03\/2015-07-21-Prediction-of-diabetes-hypoglycemic-events_images-image001.png\")
<\/p>\nOur customer develops connected blood glucose meters to provide innovative diabetes solutions to its patients. Using this meter, they\u2019re able to store and archive patients\u2019 data for further analysis. Furthermore, their connected device allows them to provide real-time analysis of the measured glucose values and make predictions about an upcoming hypoglycemic (hypo) event. Considering the traumatic experience of a diabetes hypo and its associated cost, being able to alert patients about a possible upcoming hypo can be of a tremendous value to them.<\/p>\n
This case study describes the approach we took to create a Microsoft Azure Machine Learning (MAML) model which predicts diabetes hypos based on blood glucose measurements only.<\/p>\n
Figure 1 shows an overview of our approach:<\/p>\n
<\/p>\n
Figure 1: overview of the approach<\/p>\n
Development was comprised of multiple iterations; in each iteration, we refined the Python script for feature and label creation, then rebuilt and re-evaluated the MAML model.<\/p>\n
Once the model was sufficiently accurate, we published it as a Web Service, which was then integrated into the real-time data pipeline.<\/p>\n
The first step was to extract the historical measurements and store them in a csv file. This file contains the following four columns:<\/p>\n
anonymized patient id, diabetes type, timestamp, glucose value\r\n<\/code><\/pre>\n<\/div>\nOur initial idea was to predict a hypo event based on previous measurements. To do so, we translated the data into a time series of measurements and used this to train our machine learning algorithm to predict the glucose value for a specific hour. Not surprisingly, this approach didn\u2019t yield useful results, mainly because we were lacking critical data: To predict the actual glucose value for a specific hour\/timeslot, we would require additional details only available from other data such as information about insulin injections and food\/drink intake.<\/p>\n
While analyzing the data, we also realized that it is better to separate the different diabetes types: While we have an average of 2-3 daily measurements for patients of diabetes type 2, many patients of diabetes type 1 measure their glucose value more frequently \u2013 which led to a dataset where type 1 measurements made of 90% of the data.<\/p>\n
We also decided to do a binary classification to predict whether a hypo event might occur within the next 24 hours (instead of the initial approach of using linear regression to predict the glucose value for a specific hour). Such a prediction is especially useful to patients of diabetes type 2, who measure their glucose value only a few times a day, so knowing they\u2019re at risk might help them to closer manage their blood glucose values over the following 24 hours.<\/p>\n
We translated the raw, extracted csv file into a new dataset, formatted to make predictions based on a time series of historical data. In our case, the final dataset contained the following information:<\/p>\n
\n- patient id, diabetes type, measured glucose value<\/li>\n
- sequence number of measurement\nstarting at 1 for each patient; this is used to split the dataset into a training and validation set without \u201cdestroying\u201d the time series (e.g. use 1-2000 for training and 2001-3000 for validation)<\/li>\n
- the label: did a hypo occur within the next 24 hours? (a hypo is defined as glucose value < 4.0)<\/li>\n
- the high, low, and average glucose values across the last 3 measurements<\/li>\n
- the time between the current and the 3rd last measurement in minutes<\/li>\n
- the high, low, and average glucose value each day for the last 7 days<\/li>\n
- the number of measurements within the last 7 days<\/li>\n<\/ul>\n
A reusable python script for creating such time series datasets has been published to GitHub<\/a>:<\/p>\n
![\"Architecture\"](\"https:\/\/devblogs.microsoft.com\/cse\/wp-content\/uploads\/sites\/55\/2020\/03\/2015-07-21-Prediction-of-diabetes-hypoglycemic-events_images-image002.png\")
<\/p>\n![\"Architecture\"](\"https:\/\/devblogs.microsoft.com\/cse\/wp-content\/uploads\/sites\/55\/2020\/03\/2015-07-21-Prediction-of-diabetes-hypoglycemic-events_images-image003.jpg\")
<\/p>\n![\"Architecture\"](\"https:\/\/devblogs.microsoft.com\/cse\/wp-content\/uploads\/sites\/55\/2020\/03\/2015-07-21-Prediction-of-diabetes-hypoglycemic-events_images-image004.jpg\")
<\/p>\n