By Glenn Jones, BSc, MD, FRCPC, MSc Assistant Dean of Clinical Medicine & Maurice Clifton, MD, MSEd, MBA Dean of Clinical Medicine, Saba University School of Medicine

Artificial Intelligence and Augmented Intelligence (AI) technologies are exciting. AI holds interest and is an opportunity to improve patient outcomes and business efficiencies. More intelligent work can enhance business agility against market shifts and other disruptive events. How far along is AI development, and where might it go forward?

What is healthcare AI?

People have different definitions for AI. The critical point is that AI is a new and potentially disruptive research program. AI is not business intelligence, basic situation or scenario analyses, or quantitative analytics with advanced statistics. These approaches provide enterprise data valuation, but they cannot exploit the more extensive data supplied by the Internet Of Things, or aggregated image, audio, video, and text files. Discovery in these new challenges requires new methods and complex technologies, such as machine learning models, deep learning with neural networks, robotics, and natural language processing. AI methods like these are typically more comprehensive and may be more compelling.

How much progress is there?

There are dozens of areas in healthcare. That makes for numerous applications of AI in larger data sets. Areas include, not an exhaustive listing:  patient risk assessments to focus prevention and screening strategies; diagnostic methods (e.g., radiology, pathology, dentistry, and cosmetology); making medical documentation easy and routine; sharper decision-making; designing therapeutics; monitoring patients and healthcare systems; greater quality assurance; and improving process efficiencies. Mapped out healthcare systems (e.g., oncology services) indicate that all parts affect the quality of care, costs, outcomes, and healthcare value. AI should investigate all parts and the whole set to find pathways for optimizing value.

Some areas of AI application now have randomized trials and systematic reviews of the evidence. For example, in radiology or imaging, a PubMed.gov search on August 16, 2021, revealed twenty-nine reviews in 2020, but so far thirty-four in 2021. Examples of systematically reviewed topics are as follows. In dentistry, AI models can help identify and classify diseases. Metabolic imaging in cancer helps detection and determining if a lesion is benign, and staging, monitoring, and prognosis. Such AI appears to out-perform hand-crafted radiomics and non-AI prediction models. In another medical imaging review, deep and machine learning methods performed better than experts in diagnosing metastases to the brain.

The concluding themes of many reviews are similar. Studies are about methods and development and are proof-of-concept in small patient samples. Models are typically complex and still require rigorous testing for reliability, clinical validity, and generalizability. Overall, models are not ready for clinical practice.

What are other healthcare areas ready for AI?

Priority areas to apply AI align with the value proposition. That is a scorecard for businesses, professionals, care recipients, bureaucrats, and politicians, similar to social choice theory and impact analyses. For example, improving efficiency is a systems-wide objective. An AI method could be an excellent class solution at the system level. In contrast, innovations could improve patient care and safety at the level of disease and therapy, yet the return on investment might be contextually limited.

Good areas for near-term AI progress are involved in the internet of things, medical diagnostics, and searching for predictors of clinical outcomes. Problem areas include medical documentation, requiring better methods for representing language and culture, and multi-omic exploitation. Distant applications of AI are medical and continuing education, licensing and certification processes, and managing medico-legal risk.

What is the challenge?

Research is sharpening our understanding of AI in healthcare. However, these are early days in interconnectivity and data aggregation, developing and testing methods, and exploring data valuation. Many businesses might be unable to fund and manage fundamental research to a distant time-horizon and potential profitability. Further, there are few thoroughly AI-trained individuals (e.g., in radiology), which is another constraint. Meanwhile, the full potential of non-AI strategies and statistics is still to be realised.

Presently, the field of healthcare AI is a disparate set of innovations and developmental directions. Methods will consolidate, and standards for research and reporting are evolving. Proprietary processes might be challenging to validate. Commercially available solutions will require dissemination and implementation strategies and science. Later, profit and social responsibility will be necessary.

An updated ethics framework must guide and shape research and application. There are cyber-security, patient risk-management, and liability concerns. Who is responsible when healthcare providers rely on AI results that produce false-positive and false-negative findings, risking patient safety? How fast should AI tools be deployed, and based on what criteria?  Will AI training and AI access be uniform or equitable and include minorities?  Along these lines, theories of medical choice—i.e., how healthcare choices get made—will have to incorporate AI.

Summary

Profound impacts of AI in healthcare will emerge. However, caution and sober analysis are essential at this moment. Due diligence and more research are needed. We need new frameworks, clinical trials, critical appraisal, and relevant regulations. The field of AI is exciting and promising, but some promises will take time to arrive.