-Sonia Wang, MS1-
Penn HealthX welcomed Dr. Ravi Parikh for our first event of the semester to talk about what all current and future clinicians should be aware of when utilizing a machine learning model in healthcare.
Dr. Parikh is a practicing oncologist and instructor in Medical Ethics and Health Policy at the University of Pennsylvania.
Machine learning is everywhere, like it or not. With applications extending from recommending products based on previous Amazon orders to recommending the next binge-worthy Netflix show, machine learning has become the poster child for technology applications. It’s also making its way into healthcare. Computer models are being developed to make diagnoses based on pathology slides and imaging tests, as well as to process electronic health records to make use of the massive repository of unstructured data.
But as machine learning models enter the healthcare market, Dr. Ravi Parikh cautions us on our ability to apply models to the clinic. Here are the takeaways from his talk.
1. What is machine learning and how does it differ from statistics?
While statistics focuses on trying to figure out a relationship or an inference from data, machine learning is used to make predictions and find patterns. Models fall into two categories: supervised models, where certain outcomes are specified in the data and we rely on the machine to predict the pre-specified outcome, or unsupervised models, where data is fed into the machine and we allow the machine to find patterns on its own. For areas like healthcare, where relationships are complex and not always clear, machine learning presents a unique advantage.
2. We don’t quite know how to use machine learning as of yet.
Although it’s great to have a machine learning model as a useful tool in the clinic, few machine learning algorithms have actually changed the decisions doctors make, according to Dr. Parikh. Machine learning models present a lot of promise, performing better than standard models of care in predicting 10-year cardiovascular complications, for instance. However, translating that to the clinic presents unique problems. For example, a user interface within electronic health records showing model predictions of hospital re-admission must be easy to interpret, or physicians may simply turn off the model display.
Dr. Parikh’s checklist for moving machine learning into the clinic includes:
- Using meaningful clinical endpoints.
- Benchmarking against meaningful standards (such as physician intuition).
- Ensuring that algorithms are generalizable and interoperable.
- Clarifying how clinicians should act on predictions.
- Auditing performances after training and adjusting the models accordingly.
3. We need to understand and address the problems with AI (i.e. bias).
Algorithms can automate bias because of biased data generation. Machine learning models can perpetuate current biases in data, especially when there is an underlying inequity in healthcare delivery and an intrinsic bias in the data. Incomplete data can lead to inaccurate predictions, giving faulty models that give biased predictions. To address this, it may be better to use unbiased data sources before clinician decision-making or to track outputs continuously and “stress-test” models using simulated datasets to check bias in real-time. Machine learning presents so much potential, but we need to understand the data before we proceed.