Practical Considerations of Artificial Intelligence in Oncology Care
Part I: Improving Safety through Post-Approval Analytics
December 23, 2016
Can HAL improve oncology care?
For almost all Baby Boomers, our first introduction to artificial intelligence (AI) was the computer HAL (Heuristically programmed Algorithmic computer) in 2001: A Space Odyssey. HAL was a sentient computer (or artificial general intelligence) that controlled the systems of the Discovery One spacecraft and interacts with the ship’s astronaut crew, and saw life as generally much more logical without the needs of the messy humans that created him.
From this fictional, and decidedly inauspicious start, AI has invaded our lives in ways great and small, mostly commonly associated with the amazing ability of Amazon and other merchants to anticipate our shopping desires. On a more tangible basis, we are seeing the development of self-driving cars and truly astonishingly good language translation services. AI has the ability to dramatically improve the care of oncology patients, but has to fit into the mind numbingly complex social, scientific and regulatory morass that is the delivery of oncology care.
Oncology care delivery is a very complex problem – a classic study in complexity. There are three major branches of oncology care – medical (drugs and chemo), surgery and radiation therapy. Most cancer patients receive at least two of these types of therapy, and almost all late stage and complex cases receive all three. Add into the equation, input from labs, imaging, payers, other medical specialties, caregivers, support providers and pharma; and the complexity of the problem fully emerges. The timing and selection of therapy is critical to outcome and patient safety – the two most important measures of success.
AI has the ability to impact patient care through; 1) Innovations and improvements in efficiency, 2) Patient safety in the care delivery process and 3) Efficacy of care delivery. In this blog, we’ll take a look at point 2), how to improve patient safety through AI and big data analytics.
AI is driven by the functions that computers do incredibly well – parse data, calculate and remember. I describe these functions in human terms rather than computer geek speak – sort, analyze, store and retrieve data. All of this functionality is without intelligence, until and unless we point it at a decision tree – an algorithm. Fortunately, all drug utilization is based on decision trees as well, a decision tree that takes in a number of diagnostic and situational factors into account, and comes up with a single best recommendation to use a medication to treat a disease – in this case cancer. In clinical care these recommended pathways of care are promulgated by scientific and medical bodies, most typically the National Comprehensive Cancer Network (NCCN).
Clinical Pathway for Pain in Cancer Patients
Treatment Options form the Basis for Clinical Pathways
Medical research and clinical trials form the basis for all clinical pathways, and pose a fundamental challenge for a particular class of cancer drugs called immune therapies. These drugs use the body’s immune system to fight cancer – in part by stripping away the disguises that cancer uses to confound the body’s natural ability to detect and destroy errant cells – like those that go on to prove cancerous. The human immune system is incredibly varied and always changing – immune therapies use powerful drugs to activate that system in a way that is specifically targeted to kill cancer cells.
These drugs offer a unique hope for cancer patients for whom prior therapies offered little. Consider the case of President Jimmy Carter (91), who was recently diagnosed with stage 4 melanoma, with metastatic disease that had spread to his brain and liver. President Carter was treated with a combination of radiation therapy and pembrolizumab (Keytruda) and subsequently declared himself cancer free – a remarkable turn of events for late stage melanoma.
The problem is that by “turning up the volume” on the immune system, a number of other very undesirable things can result, as pointed out in the recent New York Times article, Immune System, Unleashed by Cancer Therapies, Can Attack Organs (Dec. 3, 2106). As the article indicates, these therapies can be effective, with positive responses in up to 40% of certain advanced cancer patients, but can generate serious adverse events in up to 54% of cases – particularly when used in combination.
It is important to note that clinical trials used to clear valuable medications like pembrolizumab through the FDA are specifically designed to eliminate as many variables as possible – other than the specific drug therapy being tested in the trial. Identifying the many factors that might be factors in Adverse Events (AEs) is impossible in a clinical trial setting – trials are designed to eliminate rather than analyze variation in the trial population. The pivotal trial for Keytruda included a final total of 556 participants – a relatively tiny number when compared to the thousands of patients that will receive this medication every year, and certainly insufficient to predict AEs in individual patients.
AI, generated by big data offers a unique opportunity to potentially predict these side effects, and give medical teams the chance to mitigate the problem through prophylactic medical intervention. Specifically, by employing a “big data” approach to actual recipients of immunotherapy, and analyzing adverse events within the population, AI may be able to spot those factors which are clinically important but opaque to our current examination. Like the causative role of Helicobacter pylori in stomach ulcers, there are often causal factors that are hidden in plain sight, and the analytical power of AI can identify them through the analysis of millions of data points relating to thousands of patients.
Making this work does not require a monolithic data lake. The realities of differing IT systems in clinical practice and regulations – particularly HIPAA – limit the practicality of such an arrangement. It is much more pragmatic to have an AI engine embedded at the practice level, with the ability to access all of the data accessible by that practice – including labs, radiology, radiation therapy and operative notes. Queries and searches can be generated locally or through a shared clinical network such as a Patient Safety Organization (PSO).
It makes sense for Pharma to support this effort through funding, and participate at the PSO level in shared research findings. This effort is directed at patient safety through reduced AEs, but can also materially improve drug targeting and clinical pathway design. I have no doubt that this system will result in a better understanding of specific biomarkers and may produce unanticipated understanding of additional combination therapies.
The sequence for making this happen is pretty easy: 1) Do a proof of concept at an individual practice, 2) Organize as a regional registry, and 3) Roll-out over multiple regions as a PSO. Ultimately, this has the potential to morph into a more powerful tool than registries not based on AI. A significant problem with cross practice registries is that data definitions are not consistent between (and sometimes even within) a practice. AI has the ability to: 1) Identify the data that actually is worth analyzing, 2) Cleaning the data to make it consistent, and 3) Using meta-data to generate additional queries for analysis.
Ultimately the purpose of this is to improve patient safety by better understanding and defining causal relationships surrounding AEs. Ultimately, we would hope to facilitate the understanding of biological mechanisms and determinations of biomarkers indicating both efficacy and adverse events.