A Novel Approach for Continuous Health Status Monitoring and Automatic Detection of Infection Incidences in People...

Area of Science:

• Medical Informatics
• Machine Learning
• Diabetes Management

Background:

• Anomaly detection methods are crucial in medical applications, especially with limited data.
• Infection incidence in type 1 diabetes (T1D) often causes hyperglycemia and requires frequent insulin adjustments, representing significant anomalies.
• Few studies have focused on detecting T1D infections using personalized health models.

Purpose of the Study:

• To develop a personalized health model for automatic infection detection in T1D individuals.
• Utilize blood glucose levels and insulin-to-carbohydrate ratio as key input variables.
• Identify deviations indicative of infection, such as elevated glucose and unusual insulin ratios.

Main Methods:

• Trained three groups of one-class classifiers on regular data and tested on data including infection days.
• Compared performance against two unsupervised models using high-precision, self-recorded data from three T1D subjects.
• Evaluated models on raw and filtered data, considering performance, computational time, and sample size.

Main Results:

• One-class classifiers demonstrated excellent performance in detecting infection incidence.
• Unsupervised models showed performance degradation due to the atypical nature of infection data.
• Boundary and domain-based one-class methods provided superior data description; specific models like one-class SVM, KNN, and K-means performed excellently across sample sizes.

Conclusions:

• One-class classifiers and unsupervised models are applicable for detecting infection incidence in T1D patients.
• Early infection detection in T1D can enable tailored services and public health threat identification.
• Future research should explore additional features like continuous glucose monitoring and physical activity data for large-scale analysis.