JMIR Publications

ABSTRACT

Background:

Semi-supervised and unsupervised anomaly detection methods have been widely used in various applications to detect anomalous objects from a given dataset. Particularly, these methods have been very popular in medical domain due to their suitability for applications, where there is lack of enough dataset for the other classes. Incidence of infections in people with type 1 diabetes creates prolonged hyperglycemia and frequent insulin injections, which are significant anomalies. Despite these potential, there has been very limited study that focused on detecting infection incidences in an individual with type 1 diabetes using a dedicated personalized health model.

Objective:

The study aims to develop a personalized health model, which can automatically detect the incidence of infection in people with type 1 diabetes using BG levels and insulin to carbohydrate ratio as input variables. The model is expected to detect deviations from the norm due to infections incidences considering elevated BG level coupled with unusual change in insulin to carbohydrate ratio.

Methods:

Three group of one-class classifiers were trained on target datasets (regular/normal day measurements) and tested on dataset containing both the target and non-target (infection days). For comparison purposes, two unsupervised models were also tested. The dataset consists of a high precision self-recorded data collected from 3 real subjects with type 1 diabetes incorporating BG, insulin, diet, and self-recorded events of infection incidences. The models were evaluated on two groups of data; raw data and filtered data with a moving average. The models were compared based on their detection performance, complexity, computational time, and number of samples required. All the experiments were conducted using MATLAB® 2018b (Mathworks, Inc, Natwick, MA).

Results:

The one-class classifiers trained and tested on the dataset achieved excellent performance in describing the dataset. In comparison, the unsupervised models suffer in performance degradation mainly because of the atypical nature of the data, i.e. non-uniform data distribution. Among the one-class classifiers, the boundary and domain-based method produced better description of the data. Regarding the computational time, NN, SVDD, and SOM took considerable training time, which typically grows as the samples size increases and only LOF and COF took considerable testing time.

Conclusions:

We demonstrated the applicability of one-class classifiers and unsupervised models for the detection of infection incidences in people with type 1 diabetes. In this patient group, detecting infection can provide an opportunity to devise tailored services and also to detect potential public health threats. The proposed approaches achieved excellent performance, and in particular the boundary and domain-based method performed better. Among the respective group, particular models such as v-SVM, K-NN, and K-means achieved excellent performance in all the sample sizes and infection cases. Altogether, we foresee that the presented result could encourage researchers to examine beyond the presented features into other additional features of the self-recorded data, e.g. various CGM feature and physical activity data, on a large scale basis.