In the last four issues, we discussed a new software that can go through over 6 million possible treatments using anywhere from 1 to 5 drugs for an effective treatment for type 2 diabetes. Along with determining the best treatment for the patient, it includes the ability to put in what the patient can afford in their budget. Click here if you are interested in trying GlucosePATH software.
Dr. John Interlandi, one of our readers, has submitted a counterpoint to using this software to decide the best treatment options. Dr. Interlandi shares his concerns about the use of AI in medical decision making.
CounterPoint: View from the Diabetes Trenches: Slow It Down
By Dr. John Interlandi
The most recent four-part series of articles by Dr. Ellerman introduces us to the concept of decision support software for prescriptions that will help us overcome the human inability to calculate the 6 million potential combinations of diabetes drugs and to assist us with other issues involving complexity in diabetes treatment. Dr. Ellerman‘s first three articles are an erudite summary of the barriers confronting providers who treat complex chronic conditions like diabetes. I am in full agreement with his unhappiness in the current scenario. However, I disagree that adding another layer of software on top of a totally dysfunctional health care system will solve the problem(s). I would like to review each installment of his series briefly and offer some comments on alternatives that are available. My recommendations will be presented for the typical scenario we all see — type 2 diabetes patients who have had the disease for 1-10 years.
In article 1 of the series, complexity in diabetes decisions is said to require AI support like in other professions due to a large number of choices available for drug treatment and combinations thereof. I don‘t think AI is a requirement for type 2‘s. Indeed, there are thousands of drug combinations for diabetes, but diet and exercise are always the first two treatments. After all, the disease is caused by calorie overload, not by a drug deficiency. Pick one or pick both. They are both safe and work exceptionally well. If a patient is symptomatic from hyperglycemia, temporary sliding scale insulin can be added to relieve pancreatic exhaustion. Sliding scale will also require the patient to learn some essential skills he or she will need later, such as using prn postoperative sliding scale, prn sliding scale after corticosteroid injections, SMBG, recordkeeping, and treatment of hypoglycemia.
Once improved control is evident, changing to diet, exercise, and adding metformin will usually maintain control. If insulin was not required initially for symptom control, then metformin can be the initial pharmacologic treatment.
From a practical standpoint, I usually initiate one new drug at a time. Otherwise, one cannot know which drug is causing side effects. Arriving at the optimal multiple drug therapy is an “add-one-at-a-time“ process, at least until we find a combination that works. A viewpoint based on a heuristic approach suggests that we rarely, if ever, have to make 6 million decisions all at once, and the early treatment decisions can and should be focused on lifestyle change and relief of symptoms rather than trying to figure out how to pay for expensive medications. Also, using AI to pick drug combinations could take longer to titrate a patient who has a side effect from a fixed drug combination because now we have to exclude two from the list of usable agents.
Also, inputting Drug-Plan rules is not necessary, especially for initial treatment, since three different kinds of Relion insulin, glucose strips, Metformin, Glyburide, Pioglitazone, diet, and exercise are all available OUTSIDE of any patients‘ insurance plan, at low cost. We do not need an AI algorithm that compensates for the patients‘ drug plan rules, co-pay tiering system, discount cards, co-pay cards, or other interferences since we are outside the system. Even if we did, the rules would be changed shortly by the insurer or by the pharmaceutical company. The key is going outside our dysfunctional health care system whenever possible. Yes, drugs covered by plans can be used, but no drug combination, no matter how well-advertised, promoted, or how extensively covered by insurance, will work for more than a few years if the patient is not exercising regularly and following a diet plan. I do use the patient‘s drug plan for combinations of OHA‘s, but only once they have been tested individually, especially those where metformin is one component of the combination. However, hidden costs are always present in drug plans, and are usually passed on to the patient.
In installment 2 of Dr. Ellerman‘s series, additional complexities due to the disease state are described. All of these are true and lead us to arguably the most challenging problem in the long-term treatment of diabetes – staying on drug combinations for long enough that we can offer protection against both glycemia and extra–glycemic diabetes complications. This issue is lost on insurers and drug companies whose inside deals misled us into using drugs with hidden system costs that are passed on to patients. This results in periods of noncoverage to patients every time their drug formulary changes. Multiple unnecessary office visits required to “re-tune“ result in more cost in the current dysfunctional system due to repeated visits. Rules come down about which antihypertensives and anti–lipid drugs are to be used this season. Changing multiple meds simultaneously when required by the PBM, which is becoming the norm, is multiplied if we want to add renoprotection and/or macrovascular protection to our regime. AI will not help by putting another layer of software on top of this mess.
Again, the solution is to go outside the dysfunctional system whenever possible to avoid treatment gaps, or to use the dysfunctional system minimally. The small gains achieved in Cardiovascular Mortality for a diabetes drug, for example, are not worth the “bang for the buck“ when such effective drugs for hypertension and lipid disorders are already available for almost no cost to patients. The UKDPS showed cardiovascular benefits using drugs that are all available as generics.
The third installment in the series points to another level of complexity — the trade-offs we make with our patients when trying to maximize the benefit of a single drug. For example, one drug, like short-acting insulin, is very effective when taken three times a day and may not cost much more than if you took it once a day, but getting a patient to take it three times a day is exhausting. Another example is that short-acting insulin, even taken twice a day, might be less expensive than a GLP-3241 RA taken once a day or once a week, but it is inconvenient. Also, every trade-off is different and requires discussion with the patient. We do not need AI to negotiate for us.
These trade-offs are real-life scenarios, and I would caution against drawing all trade-offs in strictly pharmacologic terms. Patients who are willing to exercise can be convinced that perhaps no injections will be necessary if they make a few lifestyle changes. Many trade-offs can be negotiated in the strictly non-pharmacologic territory.
The fourth installment of the series discusses the specifics of what information would have to be entered into an AI system before it could yield answers. This would require having two sets of EMR entries for each patient. This amounts to another communication barrier that gives us less time to talk to our patients. I am a big fan of using the EMR for prescriptions, but I do not accept the reasoning behind the necessity for an additional layer of AI, as proposed in the first three installments. Such programs could fuel a digital “arms race“ between Providers vs. Insurers (and their ilk), a battle which we will surely lose. I also fear that such programs will be diverted to for-profit uses, or that unfunded mandates to use such programs will just add to the burden.
So do we need AI for patient care? I can think of a hundred ways it is useful already, but I do not recommend it to make choices for diabetes treatment. The process of prescribing treatment is a two-way discussion that allows the patient to have input. Our health care system is already paralyzed by too many nonhuman interactions. However, I believe AI can be taken in other directions that would be useful. It is already revolutionizing the diabetes space in many ways that I have not discussed here.
“Whoever wishes to foresee the future must consult the past; for human events ever resemble those of preceding times. This arises from the fact that they are produced by men who ever have been, and ever shall be, animated by the same passions, and thus they necessarily have the same results.“ –Machiavelli
RESPONSE TO Dr. Interlandi Comment from SERIES, from Dr. Ellerman
We respect the dialog provided by Dr. Interlandi regarding our article series. We share a common belief that lifestyle change provides a core level of benefit about both the risk and treatment of diabetes. As not all patients can achieve this goal, we believe that this approach often requires the addition of medication. Regarding medication, we also agree that diabetes regimens should have personalization. As type 2 diabetes represents the interaction of a baseline genetic risk with the environment, optimal results will require considerations of multiple approaches toward therapy.
We disagree with Dr. Interlandi regarding the complexity of decision making in the treatment of type 2 diabetes. In light of growing evidence of the differentiation of outcomes between and within treatment classes, both the AACE and ADA recognize the need to customize treatment selection based on individual needs. We propose that algorithmic navigation of these choices allows clinicians to present the best options to patients for shared decision making.
In his response, Dr. Interlandi suggests that complexity can be tackled by narrowing treatment choices to 2 to 4 inexpensive options (metformin, glyburide, pioglitazone, and relion insulin). Of the options presented, three drugs are associated with weight gain, and two are associated with the risk of hypoglycemia. Additionally, no consideration is made about recently described benefits in cardiovascular disease with the SGLT2 and GLP1 classes. Ultimately, this approach is not a rejection of computer-assisted decision support; it is the rejection of the treatment approach of two of the largest and most respected organizations for the treatment of patients with diabetes.
We would also challenge the respondent‘s calculation of cost. While initial data entry is sometimes time–consuming, accessibility of information can be integrated into the electronic medical record. As this information is combined with insurance information that is often difficult to obtain, the end result of decision support is often a decision that is simultaneously more effective, more tolerable, and more affordable. We ultimately believe that this will lead to a decrease in the need for interaction with providers and a decrease in overall cost. Again, principled rejection of the concept of prescription insurance can be used to solve this problem, but this unconventional approach limits choice to a small number of generic prescription drugs. As it is, the criticism of using software is a complaint about how the respondent gets paid – by workload instead of the outcome.
In the end, our work in decision support is based on a core set of values. Logical clinical decision making should have principals that can be defined. Interventions can be quantified about these clinical principals. Quantification leads to a ranking of interventional benefits. Quantification and ranking are inherently mathematical. Computers can perform mathematical processes on a much grander scale than a human mind. This will lead to a broadening of the scope of treatment options and a shift towards human to human communication in medicine.
If you are interested in trying the GlucosePATH software for educational, research, or other purposes at no cost, please email email@example.com with your purpose and background.
This article was updated on May 5, 2020; originally published Feb. 17, 2018.
Bradley Eilerman MD MHI obtained his medical degree from the University of Kentucky in 2001. He completed his internship at Vanderbilt University Medical Center in 2001 and his residency in Internal Medicine/Pediatrics at the University of Cincinnati in 2005. Dr. Eilerman went on to complete a fellowship in Diabetes, Endocrinology, and Metabolism in 2008. He finished his Master‘s in Health Informatics in 2016. Dr. Ellerman works as the lead physician, director of clinical research, and clinical endocrinologist for the St. Elizabeth Physician‘s group, where he cares for a patient base of approximately 2,500. He is licensed to practice medicine in the state of Kentucky.
Len Testa is a computer scientist at GlucosePATH.com, focusing on the combinatorial optimization engine and architecture. Len‘s other work includes optimization software that has helped millions of people minimize their waits in line at Disney World. (He‘s happy to talk about that, too.) That said, it‘s widely acknowledged that Len is the “pretty face“ of a much more talented group of people and that “pretty face“ is charitable, at best.
John Interlandi, MD is a practicing Endocrinologist in Hermitage, TN. Dr. Interlandi graduated from University of Oklahoma College of Medicine in 1976 and has been in practice for 42 years. He completed a residency at Medical College of Wisconsin. Dr. Interlandi also specializes in Internal Medicine. He currently practices at John Interlandi and is affiliated with TriStar Summit Medical Center, Saint Thomas Midtown Hospital and Sumner Regional Medical Center. Dr. Interlandi is board certified in Endocrinology, Diabetes and Metabolism.