Edge-Based Deep Learning Systems for Point-of-Care Diagnostic Intelligence

Abstract

Point-of-Care Testing for Blood-Based Biomarkers of Diseases Such as Cancer, Cardiovascular Diseases, and Infectious Diseases Often Benefits from Deep Learning for Diagnostic Intelligence. However, High-Quality Testing Requires Data-Gathering Conditions That Are Difficult to Meet. Recent Developments in Edge-Based Deep Learning Systems—Systems That Perform Deep Learning Model Inference Locally, on the Device—Overcome Current Limitations of Deep Learning in Point-of-Care Testing, Supporting Anonymization of Data, Reducing Inference Time, Reducing Resource Requirements, and Enabling Offline Operation. The Solutions Follow a Conceptual Roadmap Based on Four Principles: (1) Speed-Up Deep Learning Inference by Utilizing an Edge Hardware Accelerator to Generate a Real-Time Testing Feedback; (2) Run Data Acquisition and Evaluation Models on-Device, Performing Deep Learning Inference Locally to Avoid the Need for a Data Connection during Testing; (3) Development of a Software Stack That Fuses Multi-Modal Data from All Sensors to Provide Diagnostic-Class Distinction; and (4) Design Testing and Deployment Protocols to Comply with All Regulation Requirements for Clinical Use in the Point-of-Care Testing Context. Recent Advances in Data Ecosystems for Deep Learning Also Support the Point-of-Care Testing Context. Establishing a Data Ecosystem Changes the Way Data Is Collected, Supporting Fast, Scalable, and Low-Cost Annotation of Data. Data-Constrained Domains Such as Point-of-Care Testing Can Benefit from Data Generation through Adaptation/Migration, Semi-Supervised Learning, or Data Synthesis. Practical Deployment of Advanced Edge-Based Deep Learning Point-of-Care Testing System-of-Systems Will Require Defining Clear Testing Objectives, Common Evaluation Metrics, and Generalizable Testing Pipelines, While Also Maintaining an Active Data Ecosystem That Delivers Quality Data for Continuous Improvement of All Systems..