Artificial Intelligence & Machine Learning in Clinical Practice

This is the last article of a three-part series written to help rising healthcare leaders establish a basic framework for understanding artificial intelligence and machine learning use within healthcare. The series concludes with a summary of the latest uses of AI and ML technology within the research and clinical practice realms.

Healthcare organizations and hospitals are already well into the stage at which cost, benefit, risk, and uncertainty analyses are conducted in parallel to funds designated for investment into AI- and ML-driven technologies.

Several AI and ML areas have already shown promise in clinical applications, such as radiology imaging support, electronic health records analytics, drug discovery and development, personalized medicine, virtual assistance, and remote physiological management. Additionally, there are several related, burgeoning research fields that healthcare leaders interested in AI can consider, including natural language processing, decision-making support, and the development of an ethics and regulatory framework.

The adoption of AI and ML technologies intended to improve healthcare delivery necessitates a carefully managed and conservative approach. The balancing of risk and reward is not only fundamental to practicing medicine throughout its history, but also for advancing the field overall.

Several writings on the subject contain the common criticism that providers are largely responsible for the slow pace of AI and ML adoption. This criticism is not entirely incorrect, but any vitriol they contain would be misguided. Such comments rarely come from other physicians.

While some might argue that this is entirely the result of self-interest, it must also be recognized that only a physician who has practiced medicine can truly know the unique burden of holding the health and well-being of another human being in their hands. It is a burden as unique as it is complex, and one central to the practice of medicine. That applies as much to malpractice policy calculations as to the physician expertise judgment call.

The influence of physician expertise tends to be weighted more heavily in the collaborative decision-making discussions at the highest levels of HCOs. The opinion of this writer is that such weighting is expected and is appropriate for patient safety. This may change over time as a greater number of AI and ML frameworks for best practices and standards are formalized and adopted by regulatory bodies.

Every year, a declining number of physicians report being employed by a small, physician-owned practice. As a greater number of physicians are being employed by larger and larger organizations, more and more healthcare delivery is taking place via healthcare organizations (HCOs).

This is relevant to investment decisions for AI and ML technologies in healthcare because the consolidation of leadership power at the top of these organizations is growing in influence relative to that of physicians.

HCO leadership is largely made up of non-clinical professionals (though there are exceptions). They will approach AI adoption risk differently on average, than a collective of physicians. Non-physician leadership may be more willing to take on more significant amounts of risk. They may weigh potential gains in operational efficiency that translate into bottom-line impact on a shorter timeline, without considering the full impact of the reduction in physician decision-making autonomy.

Naturally, there will be a spectrum of approaches among HCOs throughout the nation, and where any particular one falls on the spectrum is a complex calculation influenced by historical successes and failures, institutional status, and how aligned physician and non-physician leadership structures are.

HCO change management strategy will continuously evolve in alignment with expected modifications to data privacy laws and payer legislation, too, which are expected to change significantly in the coming years. For data privacy and data sharing laws, the Information Blocking Rule rolled out by HHS and OIG, and their introduction of fairly harsh financial penalties imposed for violations, are expected to enhance the use of data sets for ML algorithm training. Such datasets are sold by health systems and by medical records companies, among others.
Another powerful force in play related to this will be the trade-offs between health data “privacy” and the value generated by ensuring analytics can be performed on all of these data, as we continue navigating toward a value-based care system aligned with societal values.

Most analysts predict greater transparency of hospital pricing, performance, and other relevant metrics that will further advance value-based care initiatives to the point at which society collectively recognizes that a single-payer system is, in fact, the only way to ensure the healthcare sector remains innovative and competitive on the international playing field.

Research into AI and ML generates many of the most eye-catching headlines. Some of the areas attracting notable researchers and funding dollars include:

The need to process massive amounts of clinical healthcare data records into datasets that can be used for analytics—including the training of ML algorithms—is surprisingly underrecognized despite the sheer amount of dollars flowing into the development of these tools. This area has cross-over into the EHR data analytics field currently gaining traction in clinical practice.

AI tools that can effectively and consistently support providers at the point of care are often discussed but are not yet a part of clinical practice in any measurable way. Initial pilots have shown wildly variable results, even within the same body of disease.

The tools most rapidly advancing to clinical deployment are those targeting non-physician providers, who must operate under physician oversight (the degree to which is dependent upon the state law and institution). This research is attracting as much public funding as it does private due to the growing number of publications that demonstrate the potential cost savings and gains in operational efficiency, in addition to medical error reduction—still a leading cause of death in our country, to the surprise of many.

While the earliest research targeted a limited number of diseases with a strong focus on cancer, the greatest potential gains to be realized from this area will be when these technologies can be leveraged for diseases that are much more common within the population, and those that require patient behavioral modifications to effectively treat and/or prevent (e.g., diabetes and heart disease, among others).

While not a traditional research field in healthcare, the area of ethics and regulatory frameworks for the use of AI and ML in healthcare is an area of critical importance. Multiple groups are taking a collaborative, consensus-based approach to drafting and publishing increasingly comprehensive guidelines, frameworks, and other documents that will eventually serve as mandated care standards for using these technologies in mandating how care must be provided. These publications will likely promote the much-needed revisions to the health data and privacy laws, too.

There are several areas within clinical practice more likely to already have implemented some form of AI and ML technology. These areas include:

Algorithms that analyze the data collected from radiology help trained physicians to identify abnormalities earlier in the process, with a greater degree of accuracy. They can expedite the workflow processes by triaging to appropriate review pools.

The use of AI and ML tools to support radiologists is surprisingly widespread among the largest HCOs and hospitals, which may seem surprising until one recognizes that radiologic imaging was one of the first specializations to leverage store-and-forward telemedicine methodology. This was made possible by the relatively very early standardization of how medical imaging data was encoded, creating an environment suitable for pooling large amounts of image data, even among distinct institutions.

The EHR contains an extraordinary amount of patient health data, and all the insights, patterns, trends, and more that are hidden within, just waiting to be unlocked. While using natural language processing to extract the most nuanced of data points from provider encounter notes is still nascent, many other components of the typical EHR contain discrete data, making standardizing and extracting information from it much more advanced compared to the free text. As such, NLP tools leveraging ML remain relegated to the research realm.

But grabbing all the other information from the health record? That’s already made its way into a good number of institutions. Existing EHR providers and stand-alone analytics companies are serving the demand within this market.

The market for drug discovery and development is one of the oldest and largest (by sheer dollar spend), with a massive amount of spending on pharmaceutical products R&D. The market for real-world data used to further drug discovery and development was, interestingly, borne out of claims data sold to them by health insurance companies.

Natural language processing in this research contributes to growth in this area as more complex health datasets can be gleaned from encounter notes written by physicians and stored in medical records.

AI applications leverage ML-powered algorithms, to tailor specific treatment plans customized down to the individual patient level. For rare and complex diseases, individualized patient treatment plans are a standard part of care, but such a process needs to be more scalable.

Incorporating AI and ML expands the scope of potential diseases and disorders that can be optimally treated for an individual patient based on all available data, making doing so scalable and financially feasible. Clinical deployment of these tools is still generally disease-limited or provided in a concierge-level program driven by a self-pay reimbursement structure.

AI-powered virtual assistants and chatbots are increasingly being used to provide basic medical advice and to triage the first level of patient contact. In environments with controlled oversight structures, they help answer patient questions and assess patient symptoms for the assignment of clinical follow-up pathways.

Remote physiological management (RPM) billing codes, and their close cousins chronic care management (CCM), are groups of billing codes that have existed for some time now.

When they first rolled out, these processes were overwhelmingly manual on the back end: think an Excel sheet and an assigned nurse making calls after calls. Now, there is dedicated software, often interfacing directly with the EMR, that helps determine how patient outreach for monitored patients is optimally structured and safeguards for medical escalation to the physician level.

The research in AI and ML applications within healthcare is continually pushing boundaries. Breakthrough findings have opened vast new frontiers for yet further exploration. As new concepts are demonstrated as safe, feasible, and potentially viable within the larger healthcare ecosystem, they are piloted within controlled clinical populations with careful monitoring and duplicative oversight mechanisms. Successful trials are then expanded in size and diversity of subjects for later phases. Eventually, new practice standards incorporating AI and ML standards will become the norm. The unknown is no longer if AI and ML will become the standard of clinical care—it is when, and likely sooner than many would have guessed.

There is no question that HCOs will be navigating an extremely complex environment in the macro scale and still need to be making good investment decisions about their organizations’ AI and ML investments on top of that. Physicians must consider that their outsized influence on AI and ML decisions may begin to wane in the coming years and consider how this should impact the way they consider supporting various implementation decisions.

One universal requirement that will hold regardless of the institution will be forming a board of stakeholders that covers the necessary areas of expertise needed to make thoughtful decisions about technology adoption. Decision-making must incorporate subject matter experts, statisticians, providers, hospital leadership, medical researchers, lawyers, and technology representatives. The processes must ensure all voices may equitably contribute, that transparency is ensured for sharing information, and that the inevitability of trade-offs affecting different stakeholders differently is not glossed over, but instead recognized centrally and clearly, by all parties.

Implementing AI and ML in healthcare will be challenging, and its impact will grow exponentially. Every single one of us plays a potential role in policy development and legislative decision-making and engaging in the democratic process. Appropriate safeguards will continue to be a key feature of systemic change within the healthcare sector as these new technologies continue redefining clinical care across the globe.