Development of a Real-time Force-based Algorithm for Infusion Failure Detection.

Area of Science:

• Biomedical Engineering
• Diabetes Technology
• Machine Learning in Healthcare

Background:

• Continuous subcutaneous insulin infusion (CSII) is vital for diabetes management but susceptible to infusion failures.
• Infusion failures can lead to hyperglycemia, burnout, and hospitalizations in people with diabetes (PWD).
• Existing CSII occlusion alarm systems have limitations in detecting infusion failures effectively.

Purpose of the Study:

• To develop and evaluate a novel classifier model for detecting CSII infusion failures.
• To improve the reliability and effectiveness of CSII therapy by addressing limitations in current detection methods.

Main Methods:

• Conducted five preclinical animal studies using insulin and non-insulin boluses.
• Collected data using in-line pressure, flow rate, and CSII pump onboard force sensors.
• Developed a novel classifier model utilizing force sensor data to detect infusion failures, comparing its performance against commercial occlusion alarm thresholds.

Main Results:

• The Bagging classifier model demonstrated high performance with 96% accuracy, 94% sensitivity, and 98% specificity.
• The novel classifier exhibited lower false-positive and false-negative rates compared to traditional occlusion alarm pressure thresholds.
• The testing dataset comprised 251 boluses, validating the model's effectiveness.

Conclusions:

• A novel non-threshold classifier significantly outperforms current CSII occlusion alarm systems in detecting infusion failures.
• This advancement can potentially reduce hyperglycemia and hospitalizations, offering more reliable CSII therapy.
• Further human clinical studies are recommended to validate the classifier's real-world performance.