Autonomous Uncertainty Quantification for Computational Point-of-Care Sensors

Area of Science:

• Biomedical Engineering
• Computational Diagnostics
• Infectious Disease Detection

Background:

• Computational point-of-care (POC) sensors offer rapid diagnostics but neural networks risk erroneous predictions.
• Accurate diagnostics are crucial in resource-limited settings for timely medical intervention.
• Lyme disease, a prevalent tick-borne illness, requires accessible and reliable diagnostic tools.

Purpose of the Study:

• To develop an autonomous uncertainty quantification technique for computational POC diagnostics.
• To enhance the reliability and accuracy of neural network-based diagnostic models.
• To improve the sensitivity and robustness of POC diagnostic systems for diseases like Lyme disease.

Main Methods:

• Developed and implemented a Monte Carlo dropout (MCDO)-based uncertainty quantification approach.
• Integrated MCDO into a paper-based, computational vertical flow assay (xVFA) platform for Lyme disease diagnosis.
• Utilized a neural network inference algorithm within the xVFA system for autonomous error exclusion.

Main Results:

• The uncertainty quantification method successfully identified and excluded erroneous predictions with high uncertainty.
• Diagnostic sensitivity of the xVFA platform for Lyme disease increased from 88.2% to 95.7% in blinded testing.
• The approach improved the reliability of neural network-driven computational POC sensing systems with minimal overhead.

Conclusions:

• Autonomous uncertainty quantification significantly enhances the accuracy and reliability of computational POC diagnostics.
• The MCDO-based technique offers a robust solution for improving neural network performance in diagnostic applications.
• This advancement holds promise for more dependable diagnostics in emergency and resource-limited settings.