Artificial Intelligence: The Doctor is In

When I hear that AI will be replacing doctors in the near future, images of Westworld cybernetics come to mind, with robots toting stethoscopes instead of rifles. The debate of the role of AI in medicine is raging, and with good reason. To understand the perspectives, you just have to ask these questions:

• What will AI be used for in medicine?
• If for diagnosis, does AI have the capability of understanding physiology in order to make a diagnosis?
• Will AI ever harm the patient?

To the first point, AI can be a significant player in areas such as gauging adverse events and outcomes for clinical trials and processing genomic data or immunological patterns. Image recognition in pathology and radiology is a flourishing field for AI, and there have even been gasp white papers proving so. The dangers start emerging when AI is used for new diagnoses or predictive analytics for treatment and patient outcomes. How a doctor navigates through the history and symptoms of a new patient to formulate a diagnosis is akin to the manner in which supervised learning occurs. We see a new patient, hear their history, do an exam, and come up with an idea of diagnosis. While that is going on, we have already wired into our brains, let’s say, a convolutional neural network. That CNN has already been created by medical school/residency/fellowship training with ongoing feature engineering every time we see a patient, read an article, or go to a medical conference. Wonderful. We have our own weights for each point found in the patient visit and voila! A differential diagnosis. Isn’t that how AI works?

Probably not. There is a gaping disconnect between the scenario described above and what actually goes on in a doctor’s mind. The problem is that machine learning can only learn from data that is fed into it, probably through an electronic medical record (EHR), a database also created by human users, with inherent bias. Without connecting the medical knowledge and physiology that physicians have, that the CNN does not have. If this is too abstract, consider this scenario – a new patient comes into your clinic with a referral for evaluation of chronic cough. Your clinic is located in the southwest US. Based on the patient’s history and symptoms, coupled with your knowledge of medicine, you diagnose her with histoplasmosis infection. However, your CNN is based on EHR data from the northeast coast, which has almost no cases of histoplasmosis. Instead, the CNN diagnoses the patient with asthma, a prevalent issue across the US and a disease which has a completely different treatment.

AI could harm the patient. After all, we do not have the luxury of missing one case like when we screen emails for spam. Testing models and reengineering features will come with risks that everyone – the medical staff and the patient – must understand and accept. But before we jump to conclusions of Dr. Robot, we must have much more discussion on the ethics as we improve healthcare with AI.

Leave a Reply