Part 2: 'Artificial Intelligence' and Machine Learning are Overhyped Scams

In the previous post of this series, I wrote about how “AI” has been unable to provide usable solutions for very lucrative problems which have been shown to be partially amenable to computational techniques. The fact that we still do not have “AI” which can drive a car better than the median human driver under all road conditions and at normal speeds tells you a lot about its systemic limitations. Similarly the inability of “AI” to identify molecules capable of binding to sites within proteins at a rate better, or with more ease, than a human is another glaring display of the true limitations of what is being promoted as “AI”. It is also noteworthy that computational techniques which form basis of ML, DL etc have been used in these areas for over two decades, not to mention that both are grounded in engineering and hard science with much less space for subjective ambiguity than other fields.

Now let us talk about the potential use of “AI” in medicine, an area which is much more subjective than the previous examples. As a few of you might know, this is not the first time people have tried to use computers for medical diagnosis or suggest treatments. While IBM’s “Watson” is probably the most recent large-scale failed attempt at using “AI” for medical diagnosis, programs which used the text input of symptoms to provide assistance in narrowing down an diagnosis have been around since at least the late-1980s. And to be honest, the text input programs from 1990s did a pretty good job for common and some rare conditions. Then again, even any half-competent nurse or pharmacist can do an acceptable job when it comes to diagnosing and treating common illnesses. But as you will see, there is much more to medicine than treating sore throats and urinary tract infections.

Since we have started talking about common and fairly straight-forward illness, let us talk about upper respiratory infections such as sore throat, laryngitis etc and lower urinary tract infections in women. I am choosing these examples to represent the lowest level of diagnostic complexity as they are common and easy to treat. In both illness, a 5-10 day course of fairly well-known antibiotics is more than adequate to cure them- and even a three day course is enough for uncomplicated UTIs in young women. As long as one avoids drugs which the patient is allergic or otherwise intolerant to, prognosis is very good and recovery is quick and uneventful. But you don’t need AI to prescribe a course of Amoxicillin-clavulanic acid or Azithromycin or something similar for common infections. And yes, many upper respiratory infections are viral in origin and antibiotics are often prescribed to prevent secondary bacterial involvement or just keep the patient happy. Moving on to more complicated illnesses.

Let us talk about treating a newly diagnosed 50-something for hypertension. Once again, the type of drugs initially prescribed are fairly standard (Angiotensin Receptor Blockers, ACE inhibitors, Thiazides, Amlodipine etc) and have remarkably few side-effects compared to older drugs. But there is a lot of devil in the details. For example, simply losing some weight and increasing physical activity will often reduce systolic and diastolic BP to the same extent as a single drug. Then there is the issue of race- specifically that black patients will respond better to thiazides and amlodipine (or their combination) than drugs targeting the Angiotensin pathway. Initiating drug therapy with low doses of a combination is frequently superior and cause fewer side-effects than using one drug. It is also necessary to factor in other medical conditions before choosing a drug or two. We also cannot forget the gap between what is desirable and achievable (and many conflicts between them) for BP reduction.

My point is that even the most common, well understood and easily treated chronic medical condition is far more complicated to treat properly than you might think. By now, you might have also caught on to something else- in medicine, there is no perfect answer and it is often a choice between two, three or more less than perfect answers, usually made using less than optimal knowledge. Now let us move to examples of diseases which are hard to treat properly- for a variety of reasons. Here is a fun one- schizophrenia. While there are over three dozen compounds approved in this country to treat that disease or similar medical conditions, not a single one can do much beyond stop overt psychosis. With the partial exception of Clozapine (which has some problematic toxicity), no currently approved anti-psychotic drug has much effect on the ‘negative’ symptoms of schizophrenia. And this is not for lack of money or effort, as every single drug with promised efficacy on negative symptoms of that disease has proved to be a disappointment in clinical practice.

The treatment of depression is similar a disaster with most drugs having some effects on the more severe forms of that illness, and maybe some therapeutic effects on anxiety and OCD. This is a nice way of saying that treatment of the most common form of depression (milder and reactive) with drugs is probably no better than with a placebo. And we still haven’t entered the real funhouse of illnesses which includes such noteworthies as Congestive Heart Failure, Renal Failure, various pre-cancerous or barely-cancerous conditions and autoimmune illnesses. Regardless of the precise etiology, presentation or type of these conditions- treating them optimally can be a bitch, and not just because of their complexity and number of therapeutic options. Here are a couple of examples: CHF treatment and outcomes varies massively depending upon age, etiology and type. Same for renal failure. And one can write entire books on whether to actively treat prostate cancer above a certain age.

Finally let us not forget the last category of diseases- the terminal ones. While there have been some advances in treatment of some less common forms of cancers in past decade, the vast majority of anti-cancer drugs approved in past two decades and costing us tens of billions per year have no worthwhile effect on overall survival when compared to older and far less expensive older drugs, surgery and radiation therapy. It does not help that most drug trials used to approve and push new anti-cancer drugs are heavily rigged to show efficacy where none exists. Similarly, the diagnosis and treatment of Alzheimer’s and other senile dementias is still tricky and mostly useless, respectively. If you have been reading this post up to this point, you will have probably figured out that diagnosis and treatment of illnesses, beyond the simple ones, is much trickier and depressing than many people want to believe. There is also a significant component of subjectivity necessary to make decisions about diagnosis and treatment. Here is an example: is a somewhat agitated patient with depressive symptoms a good candidate for addition of anti-psychotics or not. Hint: in most cases, the answer is NO and such polypharmacy will cause more problems than it allegedly solves, but there are exceptions. Or is a frail 70-something person with ejection fraction less than 30% a good candidate for surgical intervention. There is no good and clear answer to this and many other questions in medicine.

In the next part of this series, I hope to go into why all those widely publicized “AI” breakthroughs are closer to clever parlor tricks or the next iteration of ‘autocomplete’ than anything useful, adaptive and reliable.

What do you think? Comments?