The computer will see you now

I recently saw a cartoonist’s view of where the clinical encounter is going. A new doctor was being shown an exam room with a doctor and a nurse standing by a computer and was told “The computer will diagnose the patient and prescribe treatment.” But, the new doctor asked, what does the doctor do? “Oh, the doctor reads what the computer said to the patient.” And the nurse? “The nurse slaps the doctor’s hand if they deviate from what the computer says.” With all the recent hype about Artificial Intelligence (AI) and Big Data, is this truly the future of medicine?

In the traditional medical encounter the patient describes their story and the doctor asks questions to expand and clarify the symptoms, followed by a physical exam: the classic H&P (History and Physical) that have been little changed for millennia. A paper in the British Medical Journal back in the 70’s found that about 75% of diagnoses were made on the basis of the history, with the physical exam taking this up over 80%. Less than 20% of diagnoses needed lab tests or X-rays, though these helped confirm the H&P.

The future, we are often told, involves computers crunching the numbers on millions of patients and finding previously unknown connections and making better diagnoses and better prognostication that doctors have been able to do. Most of the big tech companies, Apple, Google, Amazon, IBM, as well as numerous start-ups, are betting billions of dollars on the future role of AI in medicine. Successes to date have been limited but well-touted. The areas which seem most promising involve very specific well-defined tasks where there is good data for the computers to crunch and where the answer is clear-cut. Examples include evaluation of skin lesions to pick out melanomas, diagnosis of retinal lesions from retinal photographs and specific questions in radiology. When such AI systems look at less well-defined problems, the successes to date are not evident.

When looking at people with ill-defined complaints, the medical record contains a lot of “noise:” data that may be relevant or utterly irrelevant to their problem. To date, physicians seem much better at separating the wheat from the chaff than do algorithms. There is another factor at play: association is not the same as cause and effect. The classic example is the old finding that when an American League team won the world series, the stock market went up. While this is interesting, I would not want to bet my retirement on this investment advice. Similarly, if computers look at large data samples, there will invariably be associations found purely by chance if enough data is crunched. Humans are needed to ask if this is plausible, if there is some possible biologic basis for the association. Another monkey wrench has been thrown into the mix: a study compared what emergency medicine residents documented in the electronic record and what they were observed to do by researchers who shadowed them. A claim of AI evangelists is that now that so much data is available via the EMR that was previously locked away in doctors’ scribbled handwritten notes, new insights are ready for the analysis. The study found that when the observers’ records were compared to the data in the EMR, there was a match for only 38% of medical history items and 53% for the physical exam. It is so easy with EMRs to click and document something you never did that notes are invariably bloated – after all, a longer note justifies a higher bill. Crunching invalid data is unlikely to result in valid conclusions!

There is also the fact that physicians are to a degree confidants and counselors as well as technicians. Your genome may suggest what diseases you are prone to get, but a human who knows you can put that in context and suggest what things you as a unique individual should do.

With all our high-tech tools, the old-fashioned H&P was recently shown to still have value. A study in Circulation found that despite the availability of echocardiograms and bio-markers, physical exam findings in patients with congestive heart failure offered significant independent prognostic information.

Bottom line? A physician using current and developing AI tools will often do a better job, but AI is not yet ready to replace the human.

Prescription for Bankruptcy. Buy the book on Amazon

New but necessarily improved

One of the major arguments that Big Pharma and its apologists make against any attempt to lower the sky-high prices Americans pay for medication is that doing so would stifle innovation. While it is true that many new drugs introduced in recent years have helped extend life, it is also true that many have been of marginal benefit at best. One example is the recent flurry of “check-point inhibitors” that have been tested against a variety of cancers. There are now six checkpoint inhibitors on the market and more on the way. The Cancer Research Institute estimated recently that some 2,250 clinical trials are underway testing this class of drugs. Unfortunately, many of these trials duplicate each other, and three recent studies looked at the use of these drugs against multiple myeloma. All three showed no benefit. Did we really need to put so many patients through these trials? Could not the investigators have waited for preliminary results from one before starting another trial?
Many cancer drugs are approved by the FDA based on “statistically significant” improvement in life span even if the extension is measured in weeks and comes at the price of nasty side effects. Still others are approved based on so-called “surrogate end-points:” tumor shrinkage or lab test improvement, even if they have not shown any improvement in lifespan. We really need pharmaceutical companies to be looking for drugs that give meaningful extension of life without horrid side effects rather than developing yet another “me-too” drug that will probably be able to obtain FDA approval and then be marketed at high price.
Another way drugs can get FDA approval without offering much benefit is by showing “non-inferiority” compared to a treatment already approved. Do patients really want to be offered a drug that “is probably not much worse” than another? A recent study published in JAMA Network Open at the end of August 2019 looked at 74 such trials of cancer drugs. While 61% could justify their use by offering convenience (such as oral rather than IV use) along with similar survival, 39% offered no obvious reason to choose the new drug over an old one - but did come at higher cost. The majority of trials showing non-inferiority without any justification for caring if the new drug was probably as good were industry-sponsored.
To regain our trust, the pharmaceutical companies need to show they are concentrating on developing truly innovative drugs that make a difference in patients’ lives and not just pushing out high-priced drugs that do little to improve quality or quantity of life but do add to our already staggering financial burden.

Prescription for Bankruptcy. Buy the book on Amazon