Clinician’s perspective of artificial intelligence in healthcare – hype vs. hope

AI is a transformative technology that will touch almost all aspects of one’s life, especially as life becomes more digital.

The term artificial intelligence (AI) has been used frequently and cavalierly to tout new aspects of technologies to create cachet and make them sexier and more marketable to potential customers. No where is this more rampant than applications in the annual $3 trillion industry that is U.S. healthcare.

Health built from chips

Indeed, the imprimatur of IBM Watson in healthcare seems to lend legitimacy to a variety of applications based on the fame generated by its chess victory 20 years ago over former world champion Gary Kasparov and the famous success in the TV game Jeopardy. Factoring in the movie and television (mostly terrifying) representations of the applications of AI in scripted entertainment, the public has, in my estimation, a somewhat distorted view of what AI is and what its potential may be.

It is important to recognize these comments as only a preface to a topic that is being written and rewritten moment by moment. To those knowledgeable about AI in healthcare, this will clearly be an oversimplification, but for those not so familiar with AI, my goal is to shed some light on what AI in healthcare is, what it isn’t, what its limitations are, and what hope it brings.

What is artificial intelligence?

Let’s begin this conversation by defining artificial intelligence. As we are aware, there are many valid definitions of intelligence. For our purpose here, recognizing and stipulating that this definition is very basic, let us define human intelligence as the ability of someone to make sense of ostensibly disparate pieces of information. It is the ability to see order in what could be construed as chaos or at least seemingly random items.

“The public has, in my estimation, a somewhat distorted view of what AI is and what its potential may be.”

Artificial intelligence, by extension, is the ability for a computer, or actually complex algorithms, to determine relationships of these data sets and to find correlations that may yield testable hypotheses.

Machine learning is the ability for the computers to incorporate these new data for future assessments. Tweaking of programming by humans makes all of these processes more efficient. When the programming to make machine learning better and the algorithms for processing data no longer require human computer scientists, i.e., the constant improvement in the technology is driven by the technology itself, we will have arrived at the age of the master algorithm, autonomous and unsupervised machine learning, the so-called Singulari... and where the science fiction alarmists have been since the 1950’s, a place now occupied by the fringier elements of the internet and dark web.

Impact on healthcare

Artificial intelligence in healthcare potentially spans the entire clinical experience, but I see its most significant contributions in three large and intersected areas. These are:

  • Data mining – gathering meaningful information by surveying huge quantities of data, or analyzing so called Big Data.
  • Medical imaging – interpretation of medical images or diagnostic digitized information from large populations or single individuals, to sort out patterns otherwise obfuscated by seemingly indiscernible noise.
  • Clinical decision support – where data gathered from patients, electronic records, any evaluative testing can be sorted into probabilities that some possibilities and options are more likely to be correct and steer the patient and provider to the choices most likely to have optimal benefit.

As stated earlier, these general areas overlap in many artificial intelligence dimensions. The number of private and public entities engaged in AI healthcare applications run the gamut from giants like Google, Baidu, and IBM to tiny startups, from seed funded startups in incubator lofts, to the premier academic and governmental laboratories all over the world. Plus there are myriad collaborative efforts with these entities to hopefully and synergistically move the field forward.

Any names that appear here are for illustrative purposes only, and the interested reader can find many more companies in all of these areas to learn more about topics of interest.

Data mining

Artificial intelligence in healthcare has been deployed on a variety of fronts. Clearly the rapid searching of data to glean information is one of the areas of so-called low hanging fruit that AI can pick. The pace of medical knowledge is exponential, with thousands of manuscripts generated daily on all topics, great and small, major and minor, in publications of variable scientific integrity from top tier academic peer-reviewed journals to self-published newsletters. No individual can keep up with his or her own sub-specialty, much less those myriad articles that run the spectrum of all medical science endeavors.

But IBM Watson Healthcare, Google's DeepMind and others... no breaks, no vacation, no family leave. AI can review tens of thousands of documents in orders of magnitude shorter time, to look for patterns and correlations, and may find associations that would take humans hundreds of years to identify, much less interpret.

Medical data processing Medical data processing

Think of the process of building enormous word clouds that “naturally” become associated with others in manners that would otherwise not be recognized. Potentially meaningful relationships not otherwise known could float to the top and be tested for significance. Tumor markers, new targets for therapy, new pharmaceutical or biopharmaceutical mechanisms of actions, relationships to one’s genome, etc., may be now correlated and provide new avenues of research. Translating those data into predictive algorithms to prevent preventable illnesses, such as cardiovascular, endocrinologic, neurologic, psychologic, and oncologic diseases, are promises that are already being realized.

“Potentially meaningful relationships not otherwise known could float to the top and be tested for significance.”

As long as medical data are mined in anonymized or population formats, privacy can be protected. Indeed, data mining with artificial intelligence may provide the backbone of what is now faddishly termed precision medicine. However, when tailored to an individual, the ethics and medico-legal implications of patient privacy will quickly rise. Science is almost always ahead of ethics and controlling law.

Medical Imaging

Another application with huge potential impact of AI, where neural networks and deep learning play a huge role, is in the arena of medical imaging. The most successful AI algorithm in the last few years, surpassing human performance in everything from vision to playing Go, is deep learning. This refers to a type of mathematical model called a neural network that contains many layers (over 1,000) of increasingly complex pattern recognizers. Details of machine learning with neural networks and deep learning are beyond the overview of this article, but I encourage you to explore those topics if these applications pique your interest.

Generative Adversial Network Generative Adversial Network (GAN)

If you’ve ever looked at a chest X-ray, or any imaging modality at all, you realize that looking for meaning requires a knowledge of the anatomy, biophysics, technical limitations, disease state and pathophysiology of the subject being imaged. You also need to know if the image is unreadable. And any radiologist will tell you that they have systematic way of looking at any image to ensure that they don’t miss anything.

Now imagine you are in an environment where there are more films to be read than you can do accurately, e.g., looking for tuberculosis in a third world country and you are the only radiologist, even with tele-radiology sending you the images to wherever you are. Imagine that you are tired, or the radiologist relieving you has just finished training without your depth of experience. What are the odds that due to the volume of images, or the experience of the reader, or the fatigue of interpreting images for hours at a time, that a mistake or error of omission could be made?

Now imagine an AI system using deep learning that can “look” and “process” images and patterns that it has originally been programmed to interpret, but now has also “learned” by experience what is, is not, or may be, normal or abnormal. What if it could also instantly correlate the images and patterns with any prior images or patterns of that same individual, even with disparate techniques affecting image quality, to identify possibly significant changes? What if it could do these interpretations in a second, actually milliseconds, per image? It doesn’t get tired, need a coffee break, need a bio-break, and it flags the possible abnormals with possible diagnoses for human over-read and clinical follow-up. Many companies are in this space, to whom I refer the interested reader for more information, e.g., Enlitic, NVIDIA, and AliveCor.

Clinical Decision Support

An application for artificial intelligence where my attention is focused as Infermedica’s chief medical officer and cofounder is in the arena of clinical decision support. That is, the ability of AI to assist the patient or the healthcare provider to arrive at a correct assessment of a patient’s status. This may be as detailed as a possible differential diagnosis, i.e., a range of possible explanations for the patient’s signs and symptoms, or as general as determining if the condition may be safely followed in the ambulatory setting with judicious follow up, or if the condition requires a professional healthcare provider to assess the patient, but can be done so electively, or if the condition warrants urgent evaluation by a healthcare professional or emergency department.

Imagine a patient without ready access to a primary care health provider who has developed symptoms of unknown cause and severity, but reaches the threshold of patient or friend/family concern. If they have internet access, they consult Doctor Google and are immediately terrified by the list of possible diagnoses that may lead to death or permanent disability. They may consult with Dr. AI from Healthtap and get more reasonable recommendations culled from the many millions of professional responses generated by Healthtap’s pool of physicians (including yours truly, I might add).

Doctor Google 1 in 3 U.S. adults used the Internet to self-diagnose

But now imagine the patient interfacing by text or chat with an AI algorithm that interprets the symptom and then asks the next question based on the probability that the answer will be the most helpful in sorting out the diagnosis and seriousness of the underlying condition. Every response is sorted by probability of arriving at the most logical diagnosis without having to ask unrelated or redundant questions as forced by fixed decision-tree algorithms.

For those of you who are not doctors, let explain how we physicians know what we know. We spend a great deal of time learning about diseases and how they present themselves subjectively and objectively. We learn how to survey a patient for whatever may be wrong. So, the very first time we take a history from and examine a real patient, it takes a few hours because we ask every question we’ve been taught to ask (called the review of systems) and do every examination of every system we know, and dutifully record our findings. And we still don’t know what the patient has.

“AI learning algorithm acts like a senior doctor who, like its human counterpart, gets smarter with more experience.”

You know that when you see an experienced doctor for the first time with a specific issue, it takes not much more than a few minutes for your doctor to hone in on what may be the problem. That’s because as we get more experience, we know what your symptoms likely (i.e., a probability estimate) mean and what to ask next to help sort it out. You can also imagine a setting where your healthcare provider, say in a screening clinic in a pharmacy or urgent care or even a primary care office, may be a crusty battle proven nurse, or a new graduate. The variability in experience may dictate what they ask, do, and think.

Infermedica’s AI strategy utilizes a proprietary inference engine that has been built by our computer and data scientists and programmed by human physicians. Age, sex, demographics, risk factors, geographic location, contemporary disease outbreaks and patterns are factored into the algorithm. What Infermedica’s AI inference engine then does is take the cumulative experience of physicians gleaned from the medical literature, textbooks, case reports, and actual patient encounters to pool them utilizing our proprietary algorithm into a core database that is constantly improving based on the probability that the answer provided leads us to the correct assessment, be it a differential diagnosis or level of acuity. In other words, our AI learning algorithm acts like a senior doctor who, like its human counterpart, gets smarter with more experience.

Medical interview – Healthform by Infermedica

Moreover, our Bayesian network engine is a more transparent mechanism than the opaque “black box” strategies implicit by deep learning where the processes between input and output are difficult to discern. Bayesian networks allow human medical experts to verify and validate the model at anytime. This permits humans to actually interrogate the machine suggested assessment and management results to ensure that they “make sense” (implicitly improved if judged appropriate and correct) in the context and relationships to real clinical scenarios. This logically includes the ability to weigh inevitable ethical issues by incorporating humane and philosophical perspectives.

Conclusion

You can use your own imagination to envision how and where these technologies would be useful to improve healthcare especially in remote locations optimizing telemedicine, or at the point-of-care. If you want to learn more and possibly explore artificial intelligence and make it your own, I refer you to fast.ai, run by Jeremy Howard who has now apparently made it his life project to train the next generation of AI coders and scientists.

Clearly, the potential future applications of AI in 21st century healthcare are myriad. I have outlined only the broadest areas, and not its applications for devices, wearables, clinical research, disease predictions, population health management, epidemiology, prevention, etc. The projected shortfalls in healthcare providers can be significantly mitigated by each professional practicing at the top of their license, and helped by the AI technologies I’ve briefly outlined. The commensurate reduction in healthcare costs because of reduced top tier human resource requirements can be translated to using those savings to provide greater access to extant and yet-to-be-developed diagnostic and therapeutic tools that we can scarcely imagine today.

“Artificial intelligence will not supplant human physicians. These AI tools will assist healthcare providers to be more efficient and accurate.”

I have borne witness to Silicon Valley venture capitalists and evangelists who have stated that there will come a time that we won’t need doctors anymore since the computers will be able to do everything we doctors do without error. I believe that those people do not understand healthcare, nor do they understand human beings.

Artificial intelligence will not supplant human physicians. These AI tools will assist healthcare providers to be more efficient and accurate. AI goes to IA, or intelligence applications. This form of medical Multiplicity (cloud collaboration between men, wisdom of the crowds, and machine learning) will help us uncover more true positives and fewer false negatives, and move us towards near perfect predictive accuracy in assessment and near optimal management recommendations.

But the healthcare of another person at its core is one of the most human endeavors to which a physician can strive. Nothing touches another person like shepherding questions of life and death, and all the steps in between. Although natural language processing and understanding algorithms, chatbots and avatars can simulate human emotion and behavior, there is absolutely no substitute for the empathy that makes the bond between the patient and the best doctors so... human.

Never miss a post from the blog, sign up to the Infermedica Newsletter.