Artificial Intelligence in Health Care: Benefits and Challenges of Machine Learning Technologies for Medical Diagnostics

What GAO Discovered

A number of machine studying (ML) applied sciences can be found within the U.S. to help with the diagnostic course of. The ensuing advantages embody earlier detection of ailments; extra constant evaluation of medical knowledge; and elevated entry to care, notably for underserved populations. GAO recognized quite a lot of ML-based applied sciences for 5 chosen ailments — sure cancers, diabetic retinopathy, Alzheimer’s illness, coronary heart illness, and COVID-19 —with most applied sciences counting on knowledge from imaging corresponding to x-rays or magnetic resonance imaging (MRI). Nonetheless, these ML applied sciences have typically not been extensively adopted.

Tutorial, authorities, and personal sector researchers are working to increase the capabilities of ML-based medical diagnostic applied sciences. As well as, GAO recognized three broader rising approaches—autonomous, adaptive, and consumer-oriented ML-diagnostics—that may be utilized to diagnose quite a lot of ailments. These advances may improve medical professionals’ capabilities and enhance affected person remedies but in addition have sure limitations. For instance, adaptive applied sciences could enhance accuracy by incorporating extra knowledge to replace themselves, however computerized incorporation of low-quality knowledge could result in inconsistent or poorer algorithmic efficiency.

Spectrum of adaptive algorithms

We recognized a number of challenges affecting the event and adoption of ML in medical diagnostics:

Demonstrating real-world efficiency throughout numerous scientific settings and in rigorous research.
Assembly scientific wants, corresponding to creating applied sciences that combine into scientific workflows.
Addressing regulatory gaps, corresponding to offering clear steerage for the event of adaptive algorithms.

These challenges have an effect on numerous stakeholders together with know-how builders, medical suppliers, and sufferers, and should sluggish the event and adoption of those applied sciences.

GAO developed three coverage choices that might assist deal with these challenges or improve the advantages of ML diagnostic applied sciences. These coverage choices establish doable actions by policymakers, which embody Congress, federal companies, state and native governments, educational and analysis establishments, and trade. See under for a abstract of the coverage choices and related alternatives and issues.

Coverage Choices to Assist Deal with Challenges or Improve Advantages of ML Diagnostic Applied sciences

	Alternatives	Concerns
Analysis (report web page 28) Policymakers may create incentives, steerage, or insurance policies to encourage or require the analysis of ML diagnostic applied sciences throughout a variety of deployment situations and demographics consultant of the meant use. This coverage possibility may assist deal with the problem of demonstrating actual world efficiency.	Stakeholders may higher perceive the efficiency of those applied sciences throughout numerous situations and assist to establish biases, limitations, and alternatives for enchancment. Might inform suppliers’ adoption choices, doubtlessly resulting in elevated adoption by enhancing belief. Info from evaluations will help inform the selections of policymakers, corresponding to choices about regulatory necessities.	Could also be time-intensive, which may delay the motion of those applied sciences into {the marketplace}, doubtlessly affecting sufferers and professionals who may gain advantage from these applied sciences. Extra rigorous analysis will seemingly result in additional prices, corresponding to direct prices for funding the research. Builders will not be incentivized to conduct these evaluations if it may present their merchandise in a unfavorable mild, so policymakers may think about whether or not evaluations needs to be performed or reviewed by impartial events, in accordance with trade officers.
Knowledge Entry (report web page 29) Policymakers may develop or increase entry to high-quality medical knowledge to develop and check ML medical diagnostic applied sciences. Examples embody requirements for accumulating and sharing knowledge, creating knowledge commons, or utilizing incentives to encourage knowledge sharing. This coverage possibility may assist deal with the problem of demonstrating actual world efficiency.	Creating or increasing entry to high-quality datasets may assist facilitate coaching and testing ML applied sciences throughout numerous and consultant situations. This might enhance the applied sciences’ efficiency and generalizability, assist builders perceive their efficiency and areas for enchancment, and assist to construct belief and adoption in these applied sciences. Increasing entry may allow builders to save lots of time within the improvement course of, which may shorten the time it takes for these applied sciences to be out there for adoption.	Entities that personal knowledge could also be reluctant to share them for quite a lot of causes. For instance, these entities could think about their knowledge worthwhile or proprietary. Some entities can also be involved in regards to the privateness of their sufferers and the meant use and safety of their knowledge. Knowledge sharing mechanisms could also be of restricted use to researchers and builders relying on the standard and interoperability of those knowledge, and curating and storing knowledge may very well be costly and should require private and non-private assets.
Collaboration (report web page 30) Policymakers may promote collaboration amongst builders, suppliers, and regulators within the improvement and adoption of ML diagnostic applied sciences. For instance, policymakers may convene multidisciplinary specialists collectively within the design and improvement of those applied sciences by way of workshops and conferences. This coverage possibility may assist deal with the challenges of assembly medical wants and addressing regulatory gaps.	Collaboration between ML builders and suppliers may assist make sure that the applied sciences deal with scientific wants. For instance, collaboration between builders and medical professionals may assist builders create ML applied sciences that combine into medical professionals’ workflows, and decrease time, effort, and disruption. Collaboration amongst builders and medical suppliers may assist in the creation and entry of ML prepared knowledge, in accordance with NIH officers.	As beforehand reported, suppliers could not have time to each collaborate with builders and deal with sufferers; nonetheless, organizations can present protected time for workers to interact in innovation actions corresponding to collaboration. If builders solely collaborate with suppliers in particular settings, their applied sciences will not be usable throughout a variety of situations and settings, corresponding to throughout completely different affected person sorts or know-how methods.

Supply: GAO. | GAO-22-104629

Why GAO Did This Research

Diagnostic errors have an effect on greater than 12 million Individuals annually, with mixture prices seemingly in extra of $100 billion, in accordance with a report by the Society to Enhance Prognosis in Medication. ML, a subfield of synthetic intelligence, has emerged as a robust software for fixing complicated issues in numerous domains, together with medical diagnostics. Nonetheless, challenges to the event and use of machine studying applied sciences in medical diagnostics elevate technological, financial, and regulatory questions.

GAO was requested to conduct a know-how evaluation on the present and rising makes use of of machine studying in medical diagnostics, in addition to the challenges and coverage implications of those applied sciences. This report discusses (1) presently out there ML medical diagnostic applied sciences for 5 chosen ailments, (2) rising ML medical diagnostic applied sciences, (3) challenges affecting the event and adoption of ML applied sciences for medical analysis, and (4) coverage choices to assist deal with these challenges.

GAO assessed out there and rising ML applied sciences; interviewed stakeholders from authorities, trade, and academia; convened a gathering of specialists in collaboration with the Nationwide Academy of Medication; and reviewed stories and scientific literature. GAO is figuring out coverage choices on this report.

For extra info, contact Karen L. Howard at (202) 512-6888 or howardk@gao.gov.