Even if you’re not into tech development you have probably heard of machine learning. In short machine learning algorithms rewrite a hypothesis when new data is processed revolutionizing data analysis all over the world. A creepy testament is Target superstores who could work out if a female shopper was pregnant based on purchasing trends. They were so accurate, they sent targeted advertising before the woman shared the news with family. Machine learning has also been used to teach drones to fly themselves. Neural network mathematics was a break though inspired by neuroscience. The visual cortex could process sound in nerves were rewired and vice versa leading to the assumption that there was a generic learning algorithm which could process different types of information. Google translate used machine learning to compare UN documents written in multiple languages resulting in the most comprehensive translational archives the world has ever seen. Will this translate into medicine? Will machine learning replace or deskill you as a doctor or nurse? Yes…. if you are not prepared.
[my desk sporting basic neural network mathematics]
Trust me I know what you’re thinking. I’ve worked in accident and emergency for 5 years. We’ve all experienced protocols that force down a certain route that isn’t appropriate and you’ve had to bypass this. This will make you skeptical and rightly so. However, this is a false comparison. These protocols usually take into account a few variables. Many protocols I’ve seen usually say exclude so and so, once done if x do this. The protocol is usually designed on small datasets (compared to machine learning data). A machine learning algorithm can not only consider thousands of variables but in it’s training sessions it will also be logging what the doctor or nurse does in this situation. I’ve met many people who take comfort in the fact that artificial intelligence cannot compete with a dog. However, I do not see dogs flying drones, making stock exchanges, shopping suggestions or booking hotel rooms. Now before you write me off as losing all my clinical acumen the moment I went back to university to study physics, I will clarify that I am well aware of the nuances of engaging with patients. Sometimes patients lie, sometimes they exaggerate and good, experienced doctors and nurses pick up on what the patient missed out or the way they said something. Quantifying pain is a great example of artificial intelligence limitations. Considering this we have to look at machine learning in a different way considering the economic concept comparative advantage in mind.
To explain comparative advantage let’s make up a scenario. Bob is a successful and busy businessman. If doesn’t take action on emails quickly he can lose money. On his days off he has professional racecar driving lessons resulting in him becoming exceptionally good, and would most probably outperform most average drivers for hire. However, when travelling to work he hires a driver. He could drive himself but instead he is making calls, answering emails and setting up meetings. He also arrives at work without being flustered as a result of not having to maneuver through traffic or find a parking space. Comparative advantage also plays a part when employing artificial intelligence. Instead of googling what blood tests to do for a particular disease or calculating antibiotic dosing and looking up the guidelines you time could be better spent engaging with patients, developing yourself, seeing more patients or actually having a break and getting something to eat. Anyone who has spent more than a year working in healthcare has skipped a lunch break or two to keep on top.
Let’s look at radiology reports. Some patients suffer by waiting for reports. The delay isn’t the radiologist being lazy. At night, they might cover multiple sites, discuss with other specialities and prioritise reporting. Machine learning combined with facial recognition software offers the opportunity for instant reporting wherever you are through cloud computing. Now radiologists into account multiple things when reporting, they also draw from experience. Some areas of scan interpretation are open to debate revealing questionable results even amongst the most experienced radiologists. However, clinging onto this and sticking your head in the sand will only work against you in the future. Machine learning algorithms access millions of scans, they process thousands of variables with each scan such as obs, past med history, and patient demographics. The algorithm also logs individual pixel intensity. As computing power increases a human will find it harder and harder to compete.
So what should you do? Although a computer will not replace you completely how can you stop it from reducing your market value? First of all avoid hyper-specialising. We know there’s a range of motivators that shape career choice. If you became a doctor or nurse just to “save lives” then you should have been a water engineer as you’d save a lot more. Despite the political correctness we all know things such as pay, prestige, practical aspects, patient interaction, pressure, risk etc play a role in career choice. I’ve met ophthalmology surgeons pick the profession because of nice hours and earning potential. Despite your motives as long as you’re good at I who am I to judge? Hyper-specialisation is attractive in this sense. I’ve met my fair share of specialists who say “that’s not technically my specialty” in-turn bumping the patient back to GP, medics or A and E with a slightly smug/relieved look at 5pm on a Friday. However, none of us can predict the future and artificial intelligence can do specific tasks extremely accurately and efficiently. If a tech company makes a random breakthrough in your particular speciality this could knock your market value.
Use the tech available. IBM and Google are pumping millions into artificial intelligence in medicine. They are not going to throw it away because you don’t like the sound of it. In fact if you’ve googled something and used it you are already letting machine learning change/dictate your practice. IBM’s supercomputer Watson uses verbal and statistical reasoning when searching through medical journals. It knows when it knows enough to make a call which was demonstrated when it beat champions in the TV game show Jeopardy. It has accumulated so much data it is now teaching medical students in the USA approaches to diagnoses. You would have spend 160 hours a week reading to keep up with it. Instead of trying to compete with this your time would be better spent communicating with patients, evaluating the suggestions Watson gives and making a judgement based on the patient’s’ lifestyle, mental capability, and social situation when choosing a treatment plan. If you learn to code a little in a high-level language there are basic machine learning algorithms already prepackaged for python in the sci-kit learn module. Pull your own data, use it to see the bigger picture and improve your practice, notice the trends and focus on how make changes in your department. This is not as hard as it sounds. I taught a room full of doctors with no coding experience how to code a DVT calculator in Python in under 40 minutes when I was a guest speaker at a conference.
Know what you’re good at and focus on this. If you wanted to be part of the academic elite you would have done a pure math degree. They generally have the highest IQ on graduation, two-thirds of students drop out, you’re considered old at the age of 25 and some areas of pure math are so academically challenging questioning your grip on logic that they drive academics who study them insane. Infinity is famous for this. If you’re skeptical Google mathematical philosophy, some of the paradoxes raised and some of the logic that solves them is mind boggling. I’m 27 which is why I did physics (applied mathematical modelling of physical systems). My academic firepower was unlikely to be strong enough to compete with the pure mathematicians. A 66% dropout rate was too much of a gamble at 24. Gigerenzer’s research showed that three-quarters of consultants failed basic probability questions. We all know that pure academia isn’t useful in real life situations and not everything. Instead of understanding advanced logic you’ve spent your time learning how to empathise with patients, make calls with limited information, develop practical skills and improvise in time-pressured situations, all valuable skills. Focusing on the delivery of medical care with the aid of big data analytics and machine learning predictions is more fruitful than making a stand and claiming that you have a better understanding of the statistics and that you’ve read more sources.
You can view this piece in a few ways. Machine learning is an evil progression that will do more harm than good and that we should fight it, this post is the ramblings of a madman who doesn’t know what he’s talking about, or that machine learning is a gift enabling you to spend more time developing practical skills, communicating with patients and sorting the things you had to let slip by due to time constraints. Hopefully, all three are correct to an extent. I’m not a supercomputer there will be stuff in here I got wrong. Hopefully, some of what I’ve written has given you some fresh insight and any rational person with clinical experience will also worry about machine learning being inappropriately used and causing harm. But most importantly I hope readers want to harness this power appropriately to solve problems in healthcare.