There has been a lot of buzz in the past few years about Artificial Intelligence (AI) and its subfield Machine Learning (ML) - especially in healthcare. Certainly, wrapping one’s brain around AI, its myriad applications, and its superhuman capabilities can seem daunting. So, let’s take a step back for a moment and discuss; what even are AI and ML anyway?
According to IBM, “artificial intelligence [...] combines computer science and robust datasets, to enable problem-solving.” Or simply, AI is about replicating human-level problem-solving with computers.
The subfield of ML is basically pattern matching. You present examples and the computer identifies the patterns.
Input -> Output
- Cute fuzzy creature with a wet nose and a wagging tail -> dog 🐶
- Animal that’s bigger than a breadbox with black and white spots and says “moo” -> cow 🐮
- Cherry red with four wheels, a steering wheel, zipping around windy roads -> sports car 🚘
The great thing about ML is that you don’t have to sit down and write all the rules for which details to present and how to balance the presence or absence of overlapping attributes. For example, there are a lot of vehicles that have four wheels, but not all of them are sports cars 🚙 ! Instead, the focus is on getting together a dataset (a big list of examples) where you have each input (a picture, a sound, a sentence, etc.) paired with the label you would want the computer to learn to predict. Then, when you present a similar input in the future it would predict the correct label.
The principle is simple, but in practice it can be quite challenging. For image classification, you need to present enough perspectives and samples that your model can robustly identify the desired pattern on real-world examples. You wouldn’t want an animal-identifying model to get confused when a different breed of dog is presented or a cat with black and white spots (“meow” or “moo”?).
ML -> AI
Learning patterns is what our brains are designed to do. The words we know, the voices and images we recognize are all patterns we map to each other and then synthesize into a broader understanding of our world. The fundamental learning algorithm that enables us to do all of this seems to be the same. Even though most human abilities are associated with a particular area in the brain (i.e. vision), many other parts of the brain can step in to replace functionality (watch this short video). What actual neural networks or, in the case of ML, what “synthetic” neural networks are doing, is taking and processing inputs to classify each pattern.
When you combine enough of these together you can get some very impressive behavior. A great example is how PredictionHealth’s Documentation service utilizes ML combined with human review to produce high quality clinical documentation. The system uses ML to listen to and record the patient encounter, and summarize the transcript yielding the history, assessment, and plan.
Why AI in Healthcare?
Not every problem is suitable for machine learning, but in medicine many are. When a patient presents to your office, there are hundreds of questions that all make perfect machine learning problems. What diagnoses does the patient have? Is each likely to improve or worsen? What treatments or medications will they need? Do we need more data to decide how to get the best outcome? What labs or studies should we order? Is this patient at risk for any complications?
For ML in medicine, the tricky part comes with getting together a great and accurate dataset - the list of inputs representing the patient’s information at the relevant time matched to the correct label. If you can get the data and algorithms working together you unlock high quality care that can be so much more accessible and affordable! This is where we have to take AI/ML in healthcare. Check out some of our favorite real word examples below.
A group of doctors in the US and India developed a dataset of 128,000 images used to train a deep neural network to detect referable diabetic retinopathy (DR). DR is the fastest growing cause of blindness, but if caught early, the disease can be treated and sight saved. Researchers wondered, could ML help predict DR in parts of the world where medical specialists are unavailable and many populations lack resources for adequate healthcare? Turns out, the answer is yes!
Much like the early detection of DR, early detection of mental health struggles can lead to dramatically improved outcomes. Another not-so-surprising parallel is the inaccessibility to timely intervention and quality health professionals. This inspired the idea that voice intonations that could be analyzed by ML might be an excellent predictor of depression onset. Researchers were right - “with just 20 seconds of an audio clip, [the] ML solution detects mental health issues with over 80% accuracy.” (Forbes, 2021)
While these early ML models have some room for improvement, ML has the ability to drastically enhance healthcare. Just imagine a world in which clinicians can take time with their patients, while those without easy access to clinicians can still receive top notch care. Incorporating AI into routine clinical care can greatly improve the patient and provider experience. Processing millions of data points and matching patterns ensures patients receive the best care every time. This is a great challenge but our team is working hard to help advance incorporation of ML into day-to-day care. It’s another small step towards better and more affordable care for all.
Pedro Teixeira, MD, PhD