Any time you’ve ever communicated with your doctor via email or had medical information shared with you or your other care providers online, you were participating in telemedicine. Even though this practice has only become commonplace in recent years, telemedicine in the U.S. has a longer history than you may think. Beginning in the 1960s, NASA began funding telemedicine projects in order to provide healthcare to astronauts in space1. Since 1999, the Center for Medicare/Medicaid Services (CMS) has allowed remote access to doctors for qualifying patients in rural areas. For non-qualifying patients, however, telemedicine services are only reimbursed at half the amount of in-person doctor visits. Payers are concerned about the increased costs of patients accessing healthcare more frequently than they would otherwise. Only during the midst of the coronavirus outbreak, in which we have all been restricted to our homes, has the CMS opened up telemedicine as a truly equal option for all patients2.
When we speak of telemedicine nowadays, we are often referring to mhealth, the practice of medicine using mobile and wireless technologies. There are a significant number of mobile apps that provide a variety of services, including mental and medical healthcare, reminders and health tips, and tracking of vital signs remotely using personal devices. Naturally, many of these services are looking to machine learning (ML) and artificial intelligence (AI) to provide solutions to issues that have plagued healthcare since its inception.
Today we will discuss some of the ways in which our healthcare system has historically struggled to provide the highest quality of care, and how ML can be used to improve the ways in which healthcare is delivered. There are many ethical considerations when discussing the use of AI in healthcare applications, but we will only touch on these briefly as such a large topic is beyond our ability to fully cover here.
To put it succinctly, medical professionals these days are stretched thin. First of all, there are widespread shortages of both nurses and doctors. On top of that, doctors simply do not have enough time to spend with patients. They are not able to collect adequate patient histories, which can be critical to proper diagnostics and treatment. Additionally, doctors are unable to commit the necessary amount of their free time to keep up with the latest discoveries in medicine. In fact, one study showed that it would take a primary care physician 29 hours per weekday to review all relevant and up-to-date literature3. This in part leads to patients receiving incorrect diagnoses 10-15% of the time, as well as suboptimal treatment choices4.
From a patient perspective, it can be difficult to find a doctor in their region that can provide the highest level of care for their specific condition. For patients in rural areas, finding a doctor at all can be a great challenge. Furthermore, many patients may find the cost of in-person doctor visits to be cost prohibitive due to the large overhead of maintaining these facilities.
The healthcare industry has been one of the last major industries to adopt machine learning. Data itself has long been valued and utilized by healthcare providers to expand their knowledge beyond their own clinical experience. Rule-based guidelines for care have been commonplace for decades. However, there has been very little implementation of ML algorithms into clinical workflows. Rule-based guidelines are restricted to only a small number of patient features at a time, whereas AI could aid clinical decisions based on a much more complex feature set, including the massive amount of data coming out of the genomic, proteomic, metabolic, and other ‘omics’ fields5.
The use of machine-based diagnostics or treatment recommendations would require a shift in thinking from experience- or rule-based care guidelines to evidence- or probability-based care. It would also require a massive overhaul in the way EHR systems are currently implemented6.
Currently, the telemedicine industry consists largely of communication tools that allow patients to connect with doctors. However, usage of AI and machine learning is constantly expanding as the industry evolves. The primary use cases are chatbots, diagnostics, and treatment recommendations. There are also some companies using machine learning algorithms to determine scheduling of virtual appointments, and even to match patients with doctors who have had the best outcomes for other patients with similar symptoms. Let’s take some time to review these technologies and speculate on how they might be implemented, while recognizing that much of the models being used in real life are proprietary and shrouded in mystery.
In general, chatbots are well-suited to replace or augment humans to collect information, provide reminders and motivational messages, and provide a sense of security when a doctor cannot immediately be reached. The simplest and most common chatbot is rule-based, meaning that the possible responses that a user can provide are fixed. With carefully planned rules, these chatbots can be powerful, but they lack the natural feel of a conversation.
The most sophisticated chatbots use natural language processing (NLP) to both interpret input and produce responses. For each message sent, the bot must determine the intent of the human and what entity they are referring to, as shown in the example below. Behind the scenes, the algorithm will make the proper query, take appropriate action, and transform the relevant data into a natural sounding response.
One of the great challenges in building these chatbots is making them specific to the healthcare domain. To accomplish this, it can be helpful to train the bot on a body of relevant text instead of using an out-of-the-box chatbot solution, but great care needs to be taken here. For example, the developer must make sure the bot does not try to diagnose if it is not supposed to. They must also consider the personality of the bot, and whether its voice should be friendly and comforting, or more straightforward and professional.
As mentioned, there are several mhealth apps that collect information about a patient’s medical history, lifestyle, and symptoms via a chatbot, and use this data to propose potential diagnoses. It is unclear exactly how each diagnostic app applies machine learning techniques, but a simple application would likely contain the following basic steps:
- utilize a massive dataset of patients with their symptoms and doctor-informed or doctor-verified diagnoses
- train a model to predict diagnoses
- use the model to generate potential diagnoses for the patient of interest.
Currently, the most common use cases of diagnostics in telemedicine are in the medical fields that make extensive use of imaging ー radiology, pathology, ophthalmology, and dermatology. Unlike the example using patient history and symptoms, where a variety of machine learning algorithms could be applied, there is really only one go-to algorithm for analyzing images: convolutional neural networks (CNNs). However, there are many potential tasks that can be accomplished under the umbrella of computer vision using CNNs. These include object detection to identify instances or counts of objects, image segmentation to partition an image, and simple classification of a whole image.
There are several real-world implementations of computer vision problems in telemedicine. One program was implemented in England for regular screening of diabetics for diabetic retinopathy, a condition which can lead to blindness. With remote imaging, there is no need for patients to see an ophthalmologist with any more frequency than low-risk patients7. Computer vision models were also trained to identify melanoma and have been shown to perform as well or better than dermatologists8.
Evaluating the accuracy of diagnoses is arguably a more difficult problem than evaluating treatment effectiveness. Because of this, and the importance of implementing optimal treatments the first time around, treatment recommendation AI has potential to make a huge difference in improving quality of care. IBM Watson Health is one of the most well-known and established examples of this, particularly for cancer treatment recommendations.
Some of the most cutting edge research in this area involves using reinforcement learning (RL), a type of machine learning that learns from experience and tries to make decisions that optimize a given reward function. The goal in using RL for treatment decisions is to develop a policy for treating patients that would be very similar to the policy of a knowledgeable doctor.
RL cannot be used in its traditional form for several reasons. Training RL models online and potentially making poor decisions on real patients would be highly unethical. We must also consider that healthcare is based on only some aspects of a person’s health (partial observability) and whose treatment goals evolve over time (non-stationarity). Both of these situations require special handling in an RL model9. All that being said, there are a few known applications of RL in healthcare that take great care to deal with these modeling and ethical challenges. One group has used deep RL with the goal of reducing organ failure among sepsis patients, by determining the dosage of vasopressors and amount of IV fluid to administer10. RL has also been used for clinical trial dosing in a simulation setting and experimenting with restricting the possible dose cycles to be within medically appropriate limits11.
While AI in telemedicine is still in its infancy, its applications are already widespread. It would not be surprising to see it in use in all aspects of both traditional and digital medicine in the not-so-distant future.
Despite the potential for telemedicine and AI to improve patient outcomes, there is still a great deal of hesitation to adopt telemedicine solutions more widely. For one, many people would never give up the personal touch that in-person doctor visits can provide. Additionally, both doctors and patients will have valid concerns around trusting AI to make diagnostic and treatment decisions. Questions remain about how liability would change when machines are contributing to healthcare decisions.
It is also worth considering the ethical implications of providing care virtually, as this may increase inequity between those with and those without internet access. Additionally, healthcare should not simply be moved to virtual formats because it is easier and cheaper ー it would ideally also be providing better healthcare than more traditional methods. Machine learning may be the key to accomplishing this, however, competition between platforms has led to a lack of transparency on methods and accuracy. This leads to a situation where even machine learning experts will be skeptical of the quality of the machine-powered diagnoses and treatment recommendations they may be receiving. That being said, AI-driven telemedicine is not going anywhere, and we should all brace for a future of healthcare that looks very different from the one we are accustomed to.
1 A Brief History of NASA’s Contributions to Telemedicine https://www.nasa.gov/content/a-brief-history-of-nasa-s-contributions-to-telemedicine
2 Telemedicine Surges, Fueled By Coronavirus Fears And Shift In Payment Rules
3 How much effort is needed to keep up with the literature relevant for primary care? https://www.ncbi.nlm.nih.gov/pmc/articles/PMC521514/
4 The incidence of diagnostic error in medicine https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3786666/
5 Rise of Deep Learning for Genomic, Proteomic, and Metabolomic Data Integration in Precision Medicine https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6207407/
6 The potential for artificial intelligence in healthcare https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6616181/
7 Diabetic Retinopathy Screening Using Computer Vision, https://doi.org/10.3182/20090812-3-DK-2006.0086
8 Computer Aided Melanoma Skin Cancer Detection Using Image Processing, https://doi.org/10.1016/j.procs.2015.04.209
9 A review of recent reinforcement learning applications to healthcare, https://towardsdatascience.com/A-review-of-recent-reinforcment-learning-applications-to-healthcare-1f8357600407
10 Deep Reinforcement Learning for Sepsis Treatment. arXiv 2017
11 Zhao Y, Kosorok MR, Zeng D. Reinforcement learning design for cancer clinical trials. Stat Med. 2009;28(26):3294–3315. doi:10.1002/sim.3720