"Medicine is a science of uncertainty and an art of probability."
- Sir William Osler
Two people presented to my clinic on the same day with classic symptoms of head and neck cancer:
Each reported several weeks of unilateral throat discomfort, ear pain, and a neck mass. Each was having some trouble swallowing and had changed his diet to accommodate the soreness. When they opened their mouths, each had a mass with a tinge of blood visible in his throat. Red flags were everywhere.
Here are more things they had in common:
Each had sought help from a walk-in clinic near his home several weeks prior to meeting me. Each had been placed on antibiotics and told he had a throat infection. Each had been reassured that things should clear up in a couple of weeks but that he might want to call his primary doctor and set up an appointment if the sore throat persisted.
In my practice as a head and neck cancer surgeon, these stories are frustratingly common. Because throat cancer is relatively rare, patients who develop classic symptoms are often thought to have more common problems. I have seen cancer patients who arrive in my clinic after having had dental extractions, months of TMJ therapy, and repeated courses of antibiotics. Ultimately, each patient has experienced frustration, time away from work, unnecessary testing, and needless expense before coming to see our team. But because each person initially saw a different care provider, the scenario remained the same.
Of course, the people working in the clinics did the best they could. We are taught that “common things occur commonly,” and that “when you hear hoof beats, think horses, not zebras.” But how might we help our doctors and nurse practitioners remain open to unusual diagnoses? How could we share our experiences to make our systems “smarter?”
Imagine this scenario:
A provider in a local urgent care center sees a patient with a history and examination like one of the people described above. She types the symptoms and examination findings into the electronic medical record system. As she enters data, the system searches through a database that includes millions of patient stories from all around the world and hundreds of thousands of peer-reviewed guidelines and journal articles. Her computer prompts her to ask a few questions she had not considered. A screen pops up: “Based on the constellation of findings, this patient has an 81.4% probability of a squamous carcinoma of the left oropharynx.” The computer automatically checks the patient’s insurance coverage and home address, and arranges for an urgent CT scan and consultation with a nearby cancer team. If the provider orders unneeded tests or medications, the computer politely suggests that there are less expensive and more effective approaches. Before the patient is sent home, the system provides helpful, culturally appropriate information, maps, and contact information. The system generates calls and text messages the next day to check on the patient and makes certain that he follows up.
I can imagine how this would have helped me. I recall a patient who was eventually found to have a throat cancer but presented with very unusual symptoms. I was initially drawn to think she had symptoms more consistent with a non-malignant diagnosis. What if my medical record system had been able to search through the records of thousands of patients and found a few others who had presented the same way? Might I have immediately broadened my approach, thereby shortening the time until I was able to make the correct diagnosis?
In a New Yorker essay entitled, “A.I. v. M.D.: The Algorithm Will See You Now,” medical oncologist Siddhartha Mukherjee describes a future where we have learned to focus Big Data on clinical problems. As examples, he describes “deep learning” programs that are being developed now to assist radiologists and dermatologists to make accurate diagnoses. For example, computer scientist, Stanford professor, and Google X founder Sebastian Thrun has built a machine-learning system that includes thousands of images of skin lesions. Thrun has tested the diagnostic accuracy of the algorithm against seasoned dermatologists. Not surprisingly, the computer does very well, usually predicting melanoma more precisely than the human dermatologists. In the future, Thrun sees these types of algorithm-based learning systems as augmenting human decision making in many specialties.
Companies like IBM Watson Health and DeepMind are already developing commercial applications for these types of “learning systems.” There are immense hurdles to overcome including privacy and security concerns, data quality, and data throughput. Nevertheless, the potential to save money, resources, time, and – most importantly – lives, is nearly visible on our horizon.
The Norman Rockwell-style family physician who relied solely on his (usually his) intuition and anecdotal experience was fading from the scene when I was in training. Young physicians will barely remember paper-based books and journals. The next generations will have even more powerful tools at their fingertips.
So why, you might ask, will we need doctors at all? Dr. Mukherjee shares a story of a dermatologist he followed during her busy day in clinic. He notes that almost all of her patients felt better after their appointments. “They had been touched and scrutinized,” he writes, “a conversation took place.” The dermatologist did more, of course, than make diagnoses. “[She] spent the bulk of her time investigating causes. Why had the symptoms appeared? Was it stress? A new shampoo? … Why now?” Perhaps future generations of caregivers will focus less on making diagnoses and more on addressing the “why” of illnesses and the importance of remaining present for patients at critical moments.
Providing a diagnosis is not of much use unless we have conversations around “This is what it all means,” “So, what’s next?” and “I promise to share this journey with you.” A computer algorithm might be able to fulfill those goals someday but, I suspect, it will be a long, long time before it can.
The patients who presented that day in my clinic did well and I continue to see them. I listen to their stories and check to make certain that everything looks fine. We talk about treatment side effects but also about their lives then shake hands and make arrangements to see them again down the road. Let’s see a computer algorithm do that.
