New insights, derived from masses of medical data and coupled with AI-based technologies that can make sense of it, are supposed to make us all healthier.
Yet, despite all the claims about the potential of AI and the billions of dollars invested to date, the revolution we’ve been waiting for has not yet arrived. There seems to be a significant gap between the expansive vision, especially as described in research and publications, and what has actually been implemented and proven to be successful.
While AI is relatively new to medicine, quantitative risk prediction is not. In fact, classical statistical models derived from structured data sources have existed for decades. Despite the fact that these models have reported to be superior to the prognostic and predictive accuracy of physicians, their application remains limited.
If AI technologies are to be adopted, there must be enough data related to the problem; the algorithms must reveal a “signal” from the data; there must be an ROI that justifies the implementation; there has to be a fit within clinical flow; the prediction must lead to an action that can influence clinical decision; and the performance of the predictive algorithms must have a high enough PPV (positive predictive value) to justify integration within the clinical practice.
While identifying a problem and building an AI model that solves it is an incredible challenge in itself, this is just the beginning. Successful implementation in healthcare systems, in such a way that the model goes on to be used in clinical practice with measurable results, is the real test. In our experience, success depends on careful consideration and addressing of the following eight factors:
1. Workflow: A key challenge for bringing AI into everyday hospital practice lies in building the workflow and processes around it. In the busy and hectic environment of a hospital or healthcare system, the introduction of a new technology, along with a new workflow to support it, is doomed to fail. As a result, it is crucial to build solutions that seamlessly fit the existing entrenched workflow.
2. Alignment of stakeholder incentives: It takes just one stakeholder to block an implementation, but for implementations to succeed, buy-in is needed from all of them. While the disruptive changes that AI brings can elevate value-based care, most of the U.S. still relies on a pay-for-service system. Thus, the success of such technologies depends on the alignment of incentives of each stakeholder in the healthcare organization—from the CMO to the physician, the care manager to the patient, the data science team to IT support.
3. Keep it simple: Building AI-based solutions can be complex. But using them and fitting them into the clinical workflow need not be. Another key to successful adoption is the simplicity of the model, when it comes to both implementing it and communicating it throughout the organization.
4. Validation: While it is hard enough to build AI-based solutions for healthcare, it is far harder to prove empirically that they actually work. Vendors in the industry have come to realize that, even though what they are delivering is software, if they truly want it to be adopted by physicians and the healthcare market in general, they must supplement their clinical implications with time and capital-consuming scientific evidence to support them.
5. Quick dissemination of positive results: For broad adoption, AI systems need to be designed in a way that makes positive results evident immediately. These results must then be communicated to all involved.
6. Transparency: While machine learning technologies are often branded “black box” solutions, in many cases this could not be further from the truth. Communicating to physicians why a flag was raised and how it was determined is critical to their adoption. These technologies are designed to support or enhance the incredibly complicated and scientific work that physicians do. Therefore, the users — physicians and healthcare providers — must receive adequate explanations of the recommendations they receive from these technologies.
7. Easy on the IT: Healthcare organizations’ IT systems are extremely complex, consisting of dozens (sometimes hundreds) of vendors. Additionally, in many organizations, data integration remains far from where it needs to be to support the power that can be extracted from AI. The last thing a CIO needs is another niche application requiring specialized support. Thus, AI systems must be designed in such a way that they can more or less be retrofitted to existing IT infrastructure and workflow with seamless integration.
8. Experimental learning style: Physicians spend a decade learning their profession through apprentice-style education; that is, they learn new skills on the job. Therefore, training them to use new systems should also be done through experimental learning (see Kolb). In addition, since implementing AI technology is never a “one and done” deal, the vendor and champion must be relentlessly committed to the implementation, making sure that it is used, measured, communicated, and ultimately successful.
While AI is not new, it is still relatively new to healthcare. Moreover, in no other industry do the implications of AI have so much significance. Thus, successful integration of AI-based clinical insights will take time, and rightly so. However, as in any industry, as the technology matures, the processes become streamlined and the industry is educated. We can be confident that eventually this technology will deliver on changing the way medicine is practiced today.
Tomer Amit is VP Corporate Marketing for Medial EarlySign, whose advanced algorithm platform accurately detects the likelihood of disease for subpopulations using basic medical information, such as blood test results, and other EMR data. The company's predictive tools provide physicians with actionable insight, while providing insurers with effective models to flag and focus on patients at risk, helping to prioritize resources, save money and improve care. Medial EarlySign's platform addresses numerous potential clinical outcomes, including cancers, diabetes and other life-threatening illnesses.