This is a classic lecture that only spurs me to learn more about Feynman's work.
What struck me last time I watched this, was Feynman citing computer aided medical diagnosis as low-hanging-fruit - a problem just waiting to be tackled. It's sobering to consider this lecture is now almost 30 years old.
Here's a little discussion about some of the issues in computer-aided diagnosis:
Historical attempts at decision support include Mycin for diagnosing infectious blood diseases, which contained assertions and rules in the form if IF-THEN clauses:
IF
the site of the culture is blood, AND the organism gram +, AND
the original infectious site was the GI tract, AND
the abdomen is the locus of infection, OR
the pelvis is the locus of infection
THEN
therapy should cover Enterobacteriaceae
Rules structured in this manner are brittle, and don’t use unification. Such a system, for example, would not robustly provide answers to queries for all flora that would be likely found in a pelvic infection. The narrow domains and lack of some of what we might call “common sense” knowledge can be problematic: H.R. Ekbia [1] humorously notes that querying a medical inference engine for suggestions on what could be causing the reddish-brown spots on the chassis and body of your Jeep, you’d get “measles”. Another medical support program, asked to suggest treatment for bacterial infection in the kidney, suggested boiling the kidney in hot water.
Feynman citing computer aided medical diagnosis as low-hanging-fruit
I think the issue is mainly related to regulations. There are a lot of health issues that an application can detect faster and more correctly than a doctor. I am not talking about Dr. House like diseases where the human brain works better.
>There are a lot of health issues that an application can detect faster and more correctly than a doctor.
Indeed. Much of everyday diagnosis is of the form: these are the symptoms, what is the distribution of possible health issues related to these symptoms? This is the stuff that computers crush.
The problem is not so much identifying problems based on symptoms, but in getting the computer to understand the symptoms. You need humans to talk to patients, to assess the patient's physical appearance and behavior, to look into noses, throats, eyes, and ears and judge what they see, to feel joints and organs and muscles, to understand the hundreds of ways patients might describe their own symptoms, and to recognize symptoms the patients themselves may not.
The humans required to perform all of these tasks could then quantify all that data into something a computer can comprehend and let it make the diagnosis--and that's perhaps not necessarily a bad idea--but if the human is appropriately trained, they too will already likely know the problem.
The risk of bringing computers into the process is that you will inevitably teach doctors and nurses to ignore their own training and instincts and just do what the computer says--the insurance companies, the health care providers, the lawyers will all insist on it implicitly or explicitly. But when AIs get stuff wrong, they get it waaay wrong. And in the 5% of cases where the computer is waaay wrong, you will find outcomes get worse, not better.
What struck me last time I watched this, was Feynman citing computer aided medical diagnosis as low-hanging-fruit - a problem just waiting to be tackled. It's sobering to consider this lecture is now almost 30 years old.