Digital

AI models are reshaping medicine. Here’s how they learn.

. 4 MIN READ
By
Timothy M. Smith , Contributing News Writer

There are lots of ways that augmented intelligence (AI)—often called artificial intelligence—can help physicians transform health care for the better, from achieving more precise diagnoses to developing better individualized treatment plans to enhancing overall patient care. Physicians know this too: An AMA survey of more than 1,000 doctors found that nearly two-thirds see potential benefits.

Half the dues, all the AMA benefits!

  • Free access to JAMA Network™ and CME
  • Save hundreds on insurance
  • Fight for physicians and patient rights

An AMA Ed Hub™ CME series introduces learners to foundational principles in AI and machine learning, a subdomain of AI that enables computers to learn patterns and relationships from data without being explicitly programmed by people. Developed by the AMA ChangeMedEd initiative and the University of Michigan DATA-MD team and geared toward medical students, it is also suitable for residents, fellows, practicing physicians and other health professionals who can earn CME by completing these modules.

The second module in the series, “AI in Health Care: Methodologies,” explores the three fundamental machine learning methods: supervised, unsupervised and reinforcement learning. It also describes deep learning and demonstrates how natural language processing is being used in health care.

To help AI meet its full potential to advance clinical care and improve clinician well-being, the AMA has developed new advocacy principles that build on current AI policy. These new principles (PDF) address the development, deployment and use of health care AI.

Learn more with the AMA about the emerging landscape of augmented intelligence in health care (PDF).

3 ways machine learning happens

The module explores these three main methodologies for training machine learning in health care.

Supervised learning. This involves computers learning from examples of correct predictions of outcomes of interest, with the goal of generating accurate predictions for new examples. Supervised learning requires labels and inputs: Labels describe what is being predicted, such as the presence or absence of a diagnosis, whereas inputs are made up of electronic health record data, omics, medical images and medical text.

Unsupervised learning. This is a type of machine learning in which the algorithm learns from unlabeled data without a predefined outcome or target variable. With unsupervised learning, the goal is to uncover common patterns, structures, or relationships within the data, such as distinct clusters of patients, disease subtypes or outliers. An unsupervised learning model can be used to, say, cluster patients with autism spectrum disorder to help discover typical utilization and disease progression trajectories.

Reinforcement learning. This uses data about sequences of interventions and their consequences or rewards to identify the best sequences of interventions to maximize a reward. For example, when treating sepsis, physicians can observe the patient's health status, including vital signs, lab results and comorbidities. They then decide how much fluid and vasopressors to give to the patient. Following the treatment, they see if the patient fully recovered or if the patient died; the reward would be one for the former or zero for the latter. Using this data, the model outputs the best sequence of actions to follow depending on the patient state.

Common challenges are highlighted for each of the methodologies.

The module also provides an example of how to apply a supervised learning method to a dataset to predict the presence or absence of cardiovascular disease. In addition, it explains how deep learning—which uses artificial neural networks with multiple layers to process data—is applied to each of the three methodologies, as well as how natural language processing enables computers to analyze and generate human language from text data.

Periodic knowledge checks and review sections test the user’s vocabulary and understanding of how concepts are applied.

The CME module “AI in Health Care: Methodologies” is enduring material and designated by the AMA for a maximum of 0.50 AMA PRA Category 1 Credit™.

It is part of the AMA Ed Hub, an online platform with high-quality CME and education that supports the professional development needs of physicians and other health professionals. With topics relevant to you, it also offers an easy, streamlined way to find, take, track and report educational activities.

Learn more about AMA CME accreditation.

FEATURED STORIES