Practical Applications of Artificial Intelligence in Oncology Care
Part II: Private Payers Using the OCM Infrastructure and AI:
February 7, 2016
Occam’s razor: The Origin of KISS
(Keep It Simple Stupid)
In July of last year Medicare implemented a new program – The Oncology Care Model (OCM) in over 100 medical oncology groups in the US. This program uses the fundamentals of the Patient Centered Medical Home (PCMH) to simultaneously improve care and lower costs. This seemingly illogical feat has been repeatedly proven to be effective by using practical tools to improve communication and simplify care – largely through the systematic use of proven best practices, tailored to individual patient needs, called clinical pathways.
The essence of the OCM and PCMH is captured by the old adage, “Plan the work, work the plan” which is itself liberally borrowed from Proverb 21:5 (Good planning and hard work lead to prosperity, but hasty shortcuts lead to poverty). The simple brilliance of the OCM and most PCMH methods is to use pathways to create a plan of work, immerse the patient as an actor in the plan (rather than an object), and then work the plan to the patient’s benefit. The pathways are a tremendous simplifying tool, which would have made old William of Ockham proud, and makes the use of artificial intelligence (AI) in care optimization a practical reality.
Why Pathways in Cancer Care?
The movement to clinical pathways is, in part, a practical necessity based on the enormity of the complexity of oncology care – caused in part by the rapid increase in the number of cancer drugs, and the practical requirement to use them in combination to maximize patient benefit. The NCI lists 489 different drugs approved for cancer therapy. Assuming that at least 10-15% of these might be potentially relevant for any given patient in some combination of drug, timing and dose, in theory, this could produce 61! (pronounced 61 factorial) potential drug/therapy combinations (1*2*3*…61), which is a very big number – in fact, more than the number of atoms in the known universe.
Add to the requirements of medical oncology, the clinical choices associated with surgical oncology, radiation oncology and symptom triage; and clinical pathways become a practical necessity. Fortunately, clinical research has narrowed down the guidelines upon which pathways are based to a few hundred, which are practical and compiled by clinical oversight bodies – notably NCCN and ASCO.
Oncology care delivery is further complicated by the need to integrate medical oncology pathways with other clinical specialties. 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.
Oncology Pathways Enable AI
AI depends on the use of Bayesian networks to integrate probability with complexity and enable machine learning. Bayesian models use a priori probability assessments to predict the probability of an outcome, and then go back and adjust those probabilities based on actual experience. Bayesian models allow for the rapid testing of multiple predictive variables, and are always benefitted from large and growing data pools, which facilitate the refinements of the prediction. The cancer treatment pathways used in the OCM allow the rapid deployment of Bayesian models, based on the data contained in providers EMRs and in payer’s databases. This application of AI has the ability to improve outcomes, lower costs and reduce adverse events associated with advanced drug therapies.
Oncology Care is Complex
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.
In AI terms, pathway based algorithms are a Heuristic technique (a Heuristic) to create a short cut for decision making – in this case based on best practice clinical evidence. AI uses Heuristics as the basis for machine learning, through Bayesian Probabilities. As discussed earlier, Bayesian probabilities use a priori probabilities to predict outcomes, and then use actual data to adjust those probabilities based on actual outcomes.
As an example, consider the following example:
- The general prevalence of cancer in the general population is 1%
- The probability of being 65 years old is 2%, and
- The probability of someone diagnosed with cancer being 65 is 0.5%, then
The probability of having cancer as a 65 year old is ((.005*.01)/.002) = 2.5%.
Suppose for a minute, then, that our computer analysis of our actual population in our area shows that the incidence of cancer is much higher than the state average – say 3% (due to a heavy smoking population). Let’s also assume that the probability of being 65 is considerably lower (due to the same smoking) – say 1.5%. Then the probability of a 65 year old coming into our clinic with cancer is much higher – say: ((.005*.03)/.0015) = 10.0%. This actual knowledge of cancer incidence in our population, dramatically changed the probability, and would clearly impact the appropriate use of tests for our patient population.
Clinical Pathway for Pain in Cancer Patients
Payers Dilemma – Beneficiaries want lower costs and better outcomes
In addition to being extraordinarily complex, oncology care is very expensive, which puts a real burden on payers, trying to control costs, but ensure best care for beneficiaries. This complexity has historically compelled payers to respond with internally driven mechanisms to understand and control the care (for cost and quality) provided to their members through a process called Prior Authorization (PA). Historically, this has involved the payer (and selected outside groups in their employ) requiring PA before undertaking expensive and uncertain treatments – in effect helping the medical team weigh the probability of success against the cost and risk.
In certain circumstances, payers have used internally approved and controlled medical oncology pathways as part of their PA protocols. These have been terribly difficult for practices to deal with from an administrative perspective, and tread awfully close to making the payer a primary decision maker in patient care – placing them in a difficult position with patients, physicians and regulators.
PA has always been a frustrating and expensive undertaking for all parties involved, and frequently leaves the patient feeling confused and frustrated.
The value of AI to payers, is that it allows the use of all patient data from all patients to evaluate and improve the pathways used by the practice in areas including:
- Targeting – which patients respond best to a particular therapy, and which do not, and which factors are most important in determining the difference?
- Risk Prediction – which patients are most prone to adverse events (AEs) from a particular therapy, and how can we tell?
- Cost benefit – is a particular expensive therapy potentially more beneficial than a less expensive alternative, will an expensive test actually provide actionable information?
- Which data is most predictive?
Payers have a layer of patient data that is unique to them relative to most OCM practices – they have cost and outcomes data across the entire continuum of care. Payers can evaluate and help optimize physician selected pathways to maximize the potential benefit to the patient, while controlling costs. This data is uniquely available to payers, and valuable to OCM (and other pathway based) practices, helping them target care, diagnostic imaging and lab tests that are particularly cost effective – and not all provider groups are created equal. Ideally, AI can be used at the practice layer and at the payer to create an integrated solution to:
- Target the most effective and efficient test or therapy for an individual patient,
- Minimize the risk of adverse events for the patient,
- Ensure lowest cost site of care delivery.
Providers differ based on cost and outcomes
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 payer level, where the practice can upload their selected pathway and treatment decisions. The AI interface can interpret and evaluate these decisions against recommended best practices generated by the AI engine, based on the physician’s selected pathway. Providers and payers can mutually decide to stick with a real time PA model (fee-for-service), based on the provider selected pathway, or utilize an alternative value based payment model. In the value based model, the AI engine can generate predicted costs and outcomes for each practice based on its individual patients, and use these to determine the care plan, basing final payment on expected costs and outcomes, rather than fee-for-service determined charges.
Development of dynamic, intelligent pathways
In practice, today’s clinical pathways and related decision criteria are relatively static, and totally deterministic. They deliver a “best care” pathway in almost every case – with an implicit Bayesian probability of 1.0 – something that we know just isn’t right. Furthermore, because these pathways are static, they never learn – we never get probability adjustments based on actual outcomes – despite having mountains of data capable of doing just that. Through the application of AI, we can move to intelligent pathways, dynamically responding to new clinical data and real world market conditions to improve care and lower costs.
Oncology accounts for just under 1% of patients, but around 10% of total costs for most payers, and intelligent pathways offer a near-term and viable solution for controlling costs. More importantly, intelligent pathways offer a way to ensure that patients get timely and correct care, which is essential for optimal outcomes and minimum adverse events. The ideal environment to start this effort is in a state like Tennessee, where the majority of medical oncology providers are OCM participants, and CMS has already paid for the enabling infrastructure.