medicine

What is medicine?

I did not really know before I started my third year of medical school.

The cognitive portion of medicine - ignoring surgery - is very much about an inference problem.  My purpose in speaking with a patient and taking a history is to elicit the relevant features of a disease that will allow diagnosis: pre-test probability of having the disease based on age, ethnicity, occupation; symptoms and their time course (acute, chronic, constant, intermittent); provoking factors (did you eat something before having the symptoms? wear a cheap watch?), past medical and surgical history, medications, allergies, social history (alcohol, smoking, drugs, living situation, occupation).

However, this all takes place in a conversation.  As more information emerges, one must then use pre-existing knowledge to tailor questions to the symptoms and try to hone in on a diagnosis.  With a limited amount of time, those questions should maximize the information that will distinguish different possible diagnoses.  For example, recent travel or long periods of immobility will increase the probability that the patient has a pulmonary embolism, or blood clot that traveled from the legs to the lungs, but chest pain radiating to the neck might increase the probability that the patient is having a heart attack.

One should have a list of potential diagnoses in mind, and as the conversation evolves and the physical exam is performed, one should be able to narrow down the potential list of problems and determine what additional information is needed to arrive at a particular diagnosis.  Then, with treatment (if any is available), one can confirm empirically whether the diagnosis was right or not; otherwise, backtracking might be important.

* * *

After some exposure to this, I find the process very interesting.  In machine learning, one needs to train the algorithm using "gold standard" cases in which the answer is known.  The more cases one sees, the better trained the algorithm can be.  However, the algorithm will be tailored to the population at hand.  The patients we see in the emergency room in one portion of the city have different problems than those in another.  We are in an area with a lot of immigrants, travelers, and poverty, but not as much crime.  A different hospital emergency room might see more stabbings and gunshot wounds.  So far, I have seen mostly chest pain, abdominal pain, vaginal bleeding, and minor lacerations.

My goal as a medical student is to increase my ability to elicit and present the relevant information to the senior physician, who is the expert at inference, to determine what the next steps are in testing and/or management of the disease.

What I wonder about is how much feedback physicians will have in the careers.  As populations evolve and change, will physicians (especially in the emergency room) know if their diagnoses were correct?  IF we are always looking at the most probable cause, we will inevitably see patients that do not have the most common or likely disease.  Medical diagnosis is very probabilistic because the tests are imperfect and so are the physicians.  What is an acceptable error rate?  How do we balance the resources needed to not miss "zebra" (rare) diagnoses with unnecessary screening and the risks of testing itself?

I would definitely be interested now in working on medical inference problems now that I have some domain knowledge.  At least in my current setting, the computer scientists and physicians working on medical inference often have no knowledge of medicine.  I think my medical training would definitely inform how I approach these questions.

I am definitely of the opinion that making computer diagnosis of disease really good must involve computers that can ask questions and take a history.  What a physician can do that at this time, a computer cannot, is observe non-verbal data (appearance of the patient, color, breathing, temperature, sound of the patient's voice, rigidity in moving the arms, palpation of the abdomen) and synthesize that into additional features.  I think if a computer had access to all the source of data that humans have (video, audio, ability to perform a physical exam), that eventually algorithms could be much better than humans at pure diagnosis.  What a computer cannot do a this point is have relationships with patients, hold their hands when they are in pain, and grieve with them.  I don't see computers displacing that aspect at all. 

The reason patients can't just type their symptoms into an automated diagnosis machine at this point is because they are not experienced enough to know what is "normal" or not.  The mysterious "clinical intuition" of a master clinician I think can eventually be learned if algorithms become good enough to learn audio and video data.

It won't be for a while, but after seeing how far computer vision has come in the past ten years, I do believe that in the next ten years that voice and video processing will become good enough in order for machine learning to incorporate audio and visual cues. 

But I don't see this as more than a tool.  Machines at this point cannot do CPR (cardiopulmonary resuscitation), blood draws, or bandage up a patient.  In the future?  Maybe so, but not at the moment.