Enhancing Peer Diagnostic Support for Rural Clinicians via Collective Intelligence and Adaptive Machine Learning

About This Grant

Modified Project Summary/Abstract Section Diagnostic errors impact 12 million Americans annually, contributing to nearly 40,000 deaths and costing the healthcare system over $100 billion yearly. Low-resource environments, including rural and inner-city communities in the US, encounter distinct challenges that adversely affect diagnostic accuracy and timeliness. These challenges include suboptimal healthcare infrastructure, physician shortages, and limited access to specialized care. These factors contribute to a high incidence of diagnostic errors in rural communities, leading to delayed diagnoses and misdiagnoses, suboptimal clinical outcomes, and increased healthcare expenditures. Artificial-intelligence tools are promising, but are untrusted, opaque (“black boxes”) and will require perpetual human-clinician oversight to avoid harm. Recognizing these issues, CollectiveGood’s overall objective is to deliver a mobile peer-consult platform that learns who the most suitable clinician reviewers are for any given case, weights their opinions intelligently, and explains the final consensus—thereby providing specialist-level diagnostic accuracy at generalist cost. Multiple studies across a range of domains show that aggregating independent clinical opinions significantly outperforms individual expert assessments—reducing error rates on the order of 33%. However, current medical crowdsourcing approaches are static, neglecting varying clinician competencies and lacking methods to optimize cost-efficiency over time. Addressing this, our platform will reduce diagnostic errors and enhance treatment outcomes by dynamically assessing clinician expertise profiles, selecting reviewers based on these profiles and case characteristics, and intelligently aggregating opinions with calibrated uncertainty estimates, thus maximizing diagnostic accuracy and minimizing resource use. In Phase I, we will enhance an already-deployed mobile peer-consult application with an attention network-powered collective-intelligence engine to develop a self-improving system that learns to predict the optimal number and composition of clinical opinions needed for different case types while continuously improving with each diagnostic decision. Specifically, we aim to 1) Create and validate a 15-minute onboarding process that assigns every new clinician a reliable “skill profile”; 2) Build a real-time “consensus engine” that converts many peer opinions into one high-quality answer; and 3) Evaluate the feasibility, acceptability, and user experience of the enhanced application in clinical settings. Completion of these aims will yield (1) a validated onboarding module, (2) a low-latency attention-based consensus engine, and (3) real-world feasibility data—all embedded in a HIPAA-compliant mobile app already in use. These assets position our university–small-business team for a multicenter Phase II trial and rapid SaaS scale-up, advancing diagnostic safety for America’s most low-resource communities.