Deep learning-assisted sensitive detection of fentanyl using a bubbling-microchip.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Analytical Chemistry

Background:

• Deep learning in point-of-care diagnostics is limited by the need for large, annotated datasets.
• Developing diagnostic algorithms often requires extensive retrospective data, hindering novel applications.
• Adversarial neural networks offer a solution to expand model utility with non-specific data.

Purpose of the Study:

• To develop a deep learning model for point-of-care fentanyl detection using adversarial networks.
• To overcome the challenge of limited task-specific annotated datasets in developing diagnostic AI.
• To demonstrate the efficacy of a smartphone-based assay for opioid detection.

Main Methods:

• Utilized adversarial neural networks (SPyDERMAN) trained on a small dataset of specific fentanyl images and a large dataset of non-specific microchip images.
• Employed a smartphone-enabled microchip bubbling assay leveraging platinum nanoparticles (PtNPs) for fentanyl detection.
• Achieved high analytical sensitivity for fentanyl detection in various matrices like PBS, human serum, and artificial urine.

Main Results:

• The SPyDERMAN network accurately detected fentanyl using limited specific data combined with extensive non-specific data.
• The assay demonstrated high sensitivity, detecting fentanyl at low ng mL^-1 concentrations.
• The system achieved high accuracy: 93 ± 0% in human serum and 95.3 ± 1.5% in artificial human urine for fentanyl detection.

Conclusions:

• Adversarial neural networks can effectively expand the utility of non-specific data for developing deep learning diagnostic models.
• Smartphone-based assays empowered by deep learning offer a promising avenue for point-of-care opioid detection.
• This approach addresses data limitations, enabling the development of advanced diagnostics for small molecular weight drugs.