Abstract
Advances in artificial intelligence (AI) and computer vision hold great promise for assisting medical staff, optimizing healthcare workflow, and improving patient outcomes. The COVID-19 pandemic, which caused unprecedented stress on healthcare systems around the world, presented what seems to be a perfect opportunity for AI to demonstrate its usefulness. However, of the several hundred medical imaging AI models developed for COVID-19, very few were fit for deployment in real-world settings, and some were potentially harmful. This review aims to examine the strengths and weaknesses of prior studies and provide recommendations for different stages of building useful AI models for medical imaging, among them: needfinding, dataset curation, model development and evaluation, and post-deployment considerations. In addition, this review summarizes the lessons learned to inform the scientific community about ways to create useful medical imaging AI in a future pandemic.
The Translation Gap in Pandemic AI
The rapid response of the global research community during the COVID-19 pandemic led to a surge in medical imaging AI models. Despite this effort, a significant "translation gap" emerged. Most researchers adopted a data-first approach—building models because datasets were available—rather than a needs-first approach. This resulted in models that achieved high accuracy on paper but lacked clinical utility in real-world environments like a crowded emergency room.
A primary failure mode was "shortcut learning," where AI models identified hospital-specific watermarks or patient positioning (e.g., supine vs. upright) to "diagnose" disease rather than identifying actual pathology. Such models fail immediately when deployed in new settings or with different equipment.
A Strategic Roadmap for AI Development
To address these failures, Huang et al. propose a structured four-stage framework to guide the development of clinically relevant AI tools during future health crises. This roadmap shifts the focus from statistical performance to real-world utility.
Stage 1: Needfinding and Planning
The foundation of a useful model is the Decision-Action Pair. Developers must move beyond simple diagnosis and define a specific clinical action triggered by the model's output (e.g., "If high risk is predicted, the action is to administer antiviral therapy"). Without a clear action, the AI's output remains a statistical abstraction without clinical impact.
Stage 2: Data Curation
Effective curation requires anticipating potential biases and utilizing multi-site data to ensure generalizability. Standardizing annotations and de-identification processes are critical for creating robust datasets. Large-scale, multi-institutional databases like RICORD and MIDRC have become essential benchmarks, though many early pandemic datasets failed basic risk-of-bias assessments.
Stage 3: Model Building and Evaluation
Standard metrics like AUROC can be misleading. A model might have a high AUROC but zero clinical utility if it fails at the specific operating point required for patient safety (e.g., missing a highly contagious patient). The authors advocate for utility-based evaluation, using indifference curves to determine the "net benefit" of an intervention within a specific clinical context. Additionally, technical safeguards like Federated Learning and Self-Supervised Pre-training can mitigate data scarcity; notably, self-supervised models can achieve superior performance using only 1% of training labels compared to traditional methods.
Stage 4: Post-Deployment Considerations
Pandemics are dynamic. A model trained on pre-vaccination data may dangerously overestimate risk in a post-vaccination population—a phenomenon known as distribution shift. Continuous monitoring for these shifts and defining the physician's role as an "educated consumer" are essential for maintaining safety and efficacy over time.
Conclusion
The lessons of the COVID-19 pandemic highlight that technical performance is not a proxy for clinical value. By shifting from "AI for the sake of publications" to a rigorous, needs-based framework, the scientific community can build a validated infrastructure ready to be deployed the moment the next pandemic emerges.
Summary updated