If a 5 year old can use the system, you are getting close

In my previous blog post, I spoke a great deal about machine learning and why it is so important, and finally, how I am trying to master the field. I referred to an example of 13-year-old software that I had used to very quickly identify a problem with an amount of data that I had been provided with. The software I was talking about was Microsoft Access. For those of you who were in the field of database design at the time that Microsoft Access was released, the excitement and the freeing of the IT developers was tremendous. They finally had a tool that made it possible to construct databases and query the data in a very visual and easy to use format. To this day, I will hear IT people comment that present-day tools still do not match Microsoft Access’s simplicity of design and implementation.

Just before the first version of Microsoft Access came out, there was a growing appreciation of the power of databases to help collect, analyze and report on the collected data. If not for Microsoft Access, I would never have been able to design the electronic medical record system that is still in use, and still very effective, for TEREM. At the time that I was designing my system, other developers literally laughed at me, saying that my system would never really work because Microsoft Access was more so a toy rather than a serious database tool. When I informed the same group of people that I was using an open source and free backend [the software that actually stores and manipulates the data] called MySQL, they laughed even harder, saying that I was using totally underpowered tools for my endpoint.

The proof is in the pudding and I have managed to demonstrate countless times that specifically in a medical environment, the tools that I used were not only sufficient, but made it possible to create effective and reliable software that is still holding up 13 years later. This software has been modified countless times, with features constantly being added. The software is written in a clear and structured fashion, following standard rules for compartmentalization, and code reuse. VB 6 which is the software engine behind Microsoft Access,  does support the use of classes. But I grew up in a time when my first major project was in machine assembler. I never really got into using classes and inheritance. But, as I said, it works and it even works well.

MySQL, by the way, is another example of a beautifully capable system that is easy to install and maintain. There are a number of tools available to make MySQL very flexible, and very compatible  with Microsoft Access. For no extra money, due to its open-source nature, I was able to set up parallel real time copies of all of the database files. This allowed me a freedom such that even if a main server totally crashed, I could be up and running with up-to-the-minute data within tens of minutes if not less. While I can’t say for sure that MySQL could handle the contents of the Library of Congress, it definitely manages just fine with a few million records that are associated with the work that has been done in TEREM.

I am sharing all of this background because I personally believe that machine learning will take a huge step forward when a front-end is developed that simplifies its use to the extreme. Just like Microsoft Access made creating and using relational databases something that non-IT people could do, I believe that the same process will happen with machine learning.

For example, I want to be able to visually classify radiological images. I want to use a machine learning algorithm to classify such images into normal versus pathological. In an ideal world, I should be able to provide such a system with two folders, one of which contains normal films and the other which contains films with pathology. Then, by using a dashboard with virtual knobs and switches, the user could manipulate the parameters that train the machine learning system. Depending on how fast the system is,  the user could try many different groups of settings to find the best. Ideally, the system would have an automatic optimization option that would find the best match between algorithm and identifying normal versus pathological films.

Machine learning is a very complicated field. People are still writing their PhD’s on different machine learning algorithms. Countless books are being written about it and endless websites talk about various ways to implement it. All I personally know is that I will be very happy when I can write even the most basic machine learning implementation and have it actually spit out something intelligent. As I’ve mentioned in the past, I am in no rush, and I am more than comfortable to spend the necessary time to gain expertise in this field.

For the general public, even the most basic concepts incorporated into Microsoft Access, are daunting. I have tried on more than one occasion to explain the concept of a relational database,  and the idea of building queries that can extract otherwise hard-to-find information. Perhaps I am a poor teacher, but I tend to find myself face-to-face with a glassy eyed colleague, the moment I start to talk about primary indices and transactions. I expect that the situation will be even more extreme when trying to explain the various types of machine learning algorithms to the non-initiated. In theory, it doesn’t matter if the doctor really understands how machine learning systems work, as long as the interfaces are simple enough to be effectively used. But there is something to be said about an inherent problem in a doctor using a tool that he has absolutely no comprehension of.

When I was an undergraduate, I took a year-long course in pharmacology. It was an absolutely fascinating course and it was of great value when I was studying various medications in medical school. I don’t think it is too much to ask that a doctor take the time to at least understand  a superficial description of how a particular medication works. If the medication works by shutting down a particular enzyme that’s associated with cell damage, I would expect the physician to know at least this. In this way, when the patient asks how this medication is different than the medication he or she was taking earlier, the doctor can give at least a cursory answer.

If the interfaces for machine learning are simplified to the point that an average physician can become comfortable with manipulating clinical data, then the physician can at least say to the patient that he or she investigated the clinical data and found that the present treatment is the best option. But if the physician can do no more than shrug his shoulders when asked why surgery is preferable over medication, this will leave a pretty poor taste in the patient’s mouth.

Please also remember that we are working towards a new world where medicine will be personalized. As such, quoting generalized research will not be considered sufficient for the patient. The patient will say “so doctor, you just told me that in a study of 1000 patients, 80% of them responded to the treatment. How do you know that I will respond to the treatment and not suffer the serious side effects that 5% of the study patients endured?”. With personalized medicine becoming more and more of a standard, it will be necessary to review a patient’s personal medical history, the patient’s genomics, the patient’s response to previous medications and then to compare all of this information to massive databases collected from other patients, even from around the world. The machine learning system will select a group of patients that match the present patient in the hospital as closely as possible, and will then render its opinion as to whether the medication is likely to work or not.

While one day, there may be specialized technicians who do this type of comparative data analysis, the doctors will at least need to understand the fundamentals of what is being done. If the patients sense that the doctor really has no understanding of how the decision was made to give the medication or not, the patient will lose trust in the doctor and in the system.

I mentioned in my last blog post that medical school may very well need to be extended by a year or more, in order to train doctors in fields of knowledge that were previously never considered key to the management of patients. There is nothing magical in the number four years as being the exact amount of time necessary to prepare a doctor for actually functioning in the community or in a hospital. Perhaps it will take six years to prepare a doctor for actual clinical work. Perhaps all doctors will have to become expert in mathematics, computer simulations, probability and machine learning before ever hoping to be able to start studying anatomy towards a degree in clinical medicine.

I have to say that such a change in the medical curriculum will definitely produce doctors with a broader view of information, and most probably overall smarter doctors who are able to think outside the box. This type of training could end up making almost every doctor a parallel researcher who will constantly be looking for new ways to understand clinical information, for the benefit of the patients. Perhaps the concept of an isolated MD will fade away and every doctor will have to complete a combined MD, PhD program in order to move on into residency. I am fully aware of the fact that this will translate into a training period that could delay the beginning of clinical practice to well into the doctors’ 40s. But given that life expectancy is constantly being extended, a few more years of training will be nothing compared to the extra years that the doctor will be able to perform his or her duties. And by being further trained, there is a real hope that the doctors who complete the entire program will be far better clinicians and create a far better health care system.

On top of all of this, it is impossible to predict what medicine will be like in even 5 to 10 years from now. With recent advances in cancer research, the entire field of oncology could drastically change. Vascular surgery is becoming almost exclusively a field that is more interventional radiology than surgery. With targeted nanomachines, which are no longer the stuff of science fiction, there may literally be no reason to operate on a patient for almost any ailment. And none of this is taking into account advances like bionic organs and cloning of organs. If there is one thing that doctors will have to be prepared for is groundbreaking change, sometimes on a day-to-day basis.

It won’t be easy. It’s very possible that the number of graduates from future medical schools will drop. I personally believe that the remuneration that doctors will receive for their extra years of training and advanced skill set will more than compensate for the extra time dedicated to their preparation. But I can’t say that for sure. We are entering an unknown space.. For all the knowledge that we now have, all that we can say is that we have absolutely no idea what tomorrow will bring. But I remain an optimist. And I have faith in the medical practice. What a wonderful place this world will be when every patient can share that same faith.

Thanks for listening

If you have any questions, or are looking for an extra pair of eyes on your ideas in medical related technology, I can be contacted via one of the following:

My web site: http://mtc.expert           
Emails: nk@mtc.expert   
Twitter: @nahkov 

About the Author
Dr. Nahum Kovalski received his bachelor's of science in computer science and his medical degree in Canada. He came to Israel in 1991 and married his wife of 22 years in 1992. He has 3 amazing children and has lived in Jerusalem since making Aliyah. Dr. Kovalski was with TEREM Emergency Medical Services for 21 years until June of 2014, and is now a private consultant on medicine and technology.