Machine Learning-Based Anomaly Detection Algorithms to Alert Patients Using Sensor Augmented Pump of Infusion Site...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Diabetes Technology

Background:

• Personal insulin pumps enhance therapy for type 1 diabetes (T1D) but face safety limits due to infusion site failures.
• Infusion site failures can lead to serious complications like hyperglycemia and ketoacidosis.
• Abundant data from modern diabetes technologies offer opportunities for machine learning (ML) to predict these failures.

Purpose of the Study:

• To evaluate a novel unsupervised anomaly detection method for real-time identification of insulin pump infusion site failures.
• To adapt the ML method for use without closed-loop systems or meal announcements, using clinical data.

Main Methods:

• Utilized a clinical dataset of 20 patients.
• Applied unsupervised anomaly detection algorithms.
• Developed an adapted feature engineering procedure for real-time, standalone operation.

Main Results:

• Achieved 0.75 Sensitivity (detecting 15 out of 20 failures) with 0.08 False Positives per day in optimal configuration.
• Outperformed existing literature algorithms.
• The algorithm predicted infusion set replacements approximately 2 hours in advance on average.

Conclusions:

• The proposed algorithm demonstrates potential for enhancing patient safety in sensor-augmented pump systems.
• Early detection of infusion site failures can mitigate risks associated with T1D management.