AI: Transforming Healthcare for a Better Tomorrow

Artificial Intelligence (AI)

The field of healthcare is no stranger to technological innovation. Recently it has seen the rise of 3D-printing within medical device and prosthesis manufacturing, and telehealth options providing more convenient, accessible care to a larger population of consumers. AI, the main subject of this discussion, has recently taken over our collective mindshare, providing innovations previously seen only in fiction and inspiration for ideas yet to come. But what exactly is AI? Is it something to fear, in which a HAL 9000-esque intelligence will eventually rule the world? Or will it be the key to solving all problems, present and future?

AI is Born

AI (Artificial Intelligence) is commonly defined as the ability of a computer to perform tasks normally associated with people and other life forms. You can see how we all interact with AI multiple times a day. For example, asking Alexa and Siri for the weather forecast or to add an appointment to your calendar is an AI interaction. Tailored recommendations from Amazon are the result of AI “learning” your preferences. Within healthcare, Machine Learning and Deep Learning provide the tools augmenting traditional care delivery models. Machine Learning is a collection of algorithms that take input data for learning and make predictions on new data based on what they learned. Deep Learning is also composed of algorithms but allows software to train itself to perform tasks similar to human cognition through multi-layered neural networks. These neural networks are loosely modeled after biological neural networks that form in brains.

AI in Practice

We covered the birth of AI and some examples used today. Let’s dig deeper and look at interesting implementations of it in healthcare:

Machine Learning

Your average healthcare consumer probably doesn’t know that machine learning is already an integral part of the ecosystem. Within the healthcare payer space, there are employees whose sole job is to calculate risk, forecast claim payouts, etc. You might be familiar with this role — actuarial science. Actuaries perform statistical modeling to drive their analysis and create predictions for future outcomes, using regression models and other machine learning algorithms. Here are some notable, interesting machine learning uses within healthcare as of late:

· Clinical Risk — A large player within risk adjustment, Optum uses machine learning models trained on both clinical and payer data to forecast clinical and financial risk for their provider clients. Other players have also entered the space, like KensSci, who provide predictions for sepsis and end of life.

· Physician Recommendations — A California-based health technology company has a platform that produces data-driven provider search and recommendation capabilities for health systems and large payers. They use machine learning to identify key physician performance metrics in the areas of experience and quality.

· Data Platform — BioSymetrics provides an end-to-end platform for machine learning as applied to pharmaceutical research. This means not only the application of algorithms, but also data collection, pre-processing, and feature engineering (features are the independent variables used to train a model). The goal of their platform is to provide a shorter time to value for drug development.

Deep Learning

The brain is a complex network of interconnected neurons which work in conjunction with each other to create a human’s cognitive abilities, such as the ability to see, control motor functions, or decide what to eat for lunch. Similarly, deep learning artificially mimics intelligence through algorithms that learn how to drive vehicles, identify objects in images, or make decisions, all using artificial neural networks. Deep learning is a powerful machine learning technique but is notoriously difficult to interpret. With deep learning, the algorithm consists of hidden layers that make it difficult to identify how the algorithm is producing decisions, or the data features that are driving a specific output. While not as familiar as typical machine learning methods, deep learning provides novel capabilities for which we’ve yet to scratch the surface because we are still in the relative infancy stage of widespread adoption. AI research is being done by your typical technology companies, but also large pharmaceutical manufacturers and provider organizations. It has tremendous potential in healthcare because of the profound impacts it can have in providing and receiving care. The AI in healthcare market, led by deep learning, was estimated at $3.9 billion in 2019 and is expected to hit $30 billion by 2025, according to Grand View Research. Deep learning has great potential:

· Cancer Detection — You might have heard about this development, where researchers at Google built the LYNA algorithm (LYmph Node Assistant) to detect metastasized breast cancer with 99% accuracy. In this case, Google used deep learning as an image recognition tool, allowing LYNA to further pinpoint cancer regions and suspicious regions on a given slide. Google tested breast cancer detection with 6 pathologists, both with and without the use of LYNA, and they all noted the task of detecting metastases was easier with LYNA.

· Voice Transcription — Think back to your days in college, where you jotted (or typed) down notes based on the lectures given by your professors. Now, think about how AI can do the same within a healthcare setting. Nuance, known for their voice-to-text services, provides solutions that greatly reduce the administrative burden of documenting in-person and telehealth visits, allowing the physician to focus on the patient. They provide this capability with all major EHR platforms. In addition to the administrative benefits, what is even more important is Nuance preserves the human on human interaction between doctors and their patients.

· Image Creation — We previously mentioned how deep learning is used for image detection. To extend that thought, researchers have even gone so far as to use neural networks for generating images! A team of researchers from NVIDIA (a computer GPU company) and Mass General Brigham trained a Generative Adversarial Network (GAN) on publicly available MRI scans of brains with Alzheimer’s disease and brain tumors. The GAN was able to generate infinitely diverse synthetic MRI images. You might be wondering, what’s the point of creating synthetic MRI images? As it turns out, these images were diverse enough for use in training models. And because they are synthetic images, there are no concerns about patient privacy or sharing these images between research institutions. These generated images can ultimately be used to train models that will be used to detect brain tumors in real-life MRIs.

Benefits

Thinking about the examples of AI above and the companies working on these solutions, you can start to see how AI could be applied to the broader challenges facing the industry today. Beyond risk scoring, AI will be used by payers and government to further justify alternative payment models and their cost efficiency. While adopting electronic medical records was a major benefit for recordkeeping, we’ve yet to solve for automating the data abstraction process. AI will one day automate the process of translating physicians’ notes, problem lists, procedure history, immunizations, etc. into an EMR — think of it like a smart version of auto-filling fields in a text box. A drug’s therapeutic efficacy is primarily understood through randomized clinical trials, but drug makers are increasingly using insights from real world evidence (RWE) to increase their understanding. Additionally, RWE can help accelerate the development and approval process. What if we went a step further and generated synthetic real-world data? Regardless of the implementation, the goal of using AI is in achieving greater outcomes and delivering better care to patients. How can you use AI today to help the healthcare ecosystem of tomorrow?

Verdict

In many ways, what’s old is new. We’ve used AI in healthcare before it really came to prominence the past decade. Going back to our original question about whether AI will rule the world or solve all problems — the answer is somewhere in the middle. While AI probably won’t take over the world a la Skynet, there are myriad considerations involved in responsibly using it to prevent negative consequences, such as those that can arise from model bias. Model bias can take many forms, from the case of omitting appropriate features for training, to bias in the features themselves. With our breast cancer detection example, if LYNA was trained only on images of Caucasian patients, there would be an inherent bias that may pose decreased accuracy in detecting breast cancer for black, Hispanic, and other patients.

In our everyday lives, we typically think of AI as computers that automate repetitive tasks or make difficult tasks easier. Does this mean AI will eventually get smart enough to replace doctors, causing a cratering of the medical school pipeline? Not anytime soon. The advancements seen from AI now and in the future can be likened to the revolution brought on by the printing press or the cotton gin — they made work easier and possibly even reduced the number of workers required — but these technologies never removed humans from the equation. Doctors are currently, and rightfully, using AI to augment care delivery instead of replacing their roles. Even with the progress we’ve made, and will continue to make, there will always be a human component to AI in healthcare.

I don’t think I would ever be comfortable with Alexa giving me a breast cancer diagnosis.

If this article resonated with you and you’re looking for help in putting AI to work in solving your organization’s most pressing problems, please reach out to me. Let’s make healthcare better for everyone.

Many thanks to Courtney McKay, Matthew Trisic, Danny Sidani, Shannon Montanez, Shane Krieger, and Cabul Mehta, whose valuable input made this article possible.

Vanna Trieu is on Slalom Boston’s Healthcare Community of Practice and has assumed multiple roles as a consultant/engineer within healthcare and life sciences. His experience centers on building data platforms that enable organizations to harness the value of their data throughout the analytics continuum, from descriptive and diagnostic to predictive and prescriptive analytics.