Machine learning models can improve workflow and accuracy, but there are still challenges being faced.
Artificial intelligence (AI) could drastically reduce systemic glitches and errors in the decision-making of clinicians — if done properly — according to an article written by scientists at Harvard Medical School and Google.
Published in The New England Journal of Medicine, the article provides a blueprint for integrating machine learning into the practice of medicine and outlines the promises and pitfalls of the technology.
There are a number of ways that AI can be used to make clinicians the best possible — or better than they were before using the technology.
Machine learning models learn the patterns of health trajectories of a vast number of patients. With the technology, physicians can better anticipate the likeliness of a future event occurring by using information beyond the physician’s experience.
For example, a machine learning model can help a physician know how quickly a patient’s disease will progress or how soon the patient can return to work.
Simple machine learning models have already been deployed in large integrated health systems, and studies show that more complex and accurate prognostic models can be built with raw data from the electronic health system and medical imaging.
According to The Institute of Medicine, nearly every patient will experience a diagnostic error in his or her lifetime.
Machine learning could be used to identify likely diagnosis during an office visit and raise awareness of diagnoses that are likely to occur later in life.
But some clinicians might not be able to use the data properly for the model to assist them in a meaningful way and the diagnoses that the models are built from might be provisional, incorrect, or simply not recorded.
If used properly, the models could suggest questions or tests to physicians based on real-time data, which could assist where misdiagnoses are common or when clinicians are uncertain.
A straightforward application of a machine learning model is to compare what is prescribed at the point-of-care to what a model predicts would be prescribed.
But a model trained on historical data would only learn the prescribing habits of physicians, not the ideal practices.
Machine learning can automatically select patients eligible for randomized controlled trials based on clinical documentation and can identify high-risk patients or subpopulations who are more likely to benefit from novel therapies.
Machine learning technologies can be used to make clinicians more efficient. The technology can help expose relevant information in a patient’s health chart for a clinician more easily and without multiple clicks.
Machine learning models can also replace prior authorization.
“The motivation behind adopting these abilities is not just the convenience to physicians — making it frictionless to view and enter the most clinically useful data is essential to capturing and recording health data that will enable machine learning to help give the best possible care to each and every patient.” — Alvin Rajkomar, M.D., Jeffrey Dean Ph.D., Isaac Kohane, M.D., Ph.D.
The increase in efficiency and automated clinical workflow could allow clinicians to spend more time with their patients.
“Understanding the limitations of machine learning is vital,” said Kohane, from the department of biomedical informatics at Harvard Medical School. “This includes understanding what the model is designed and, more importantly, what it’s not designed to do.”
To build a machine learning model, it is critical to assemble a representative, diverse dataset. The models ideally should be trained on similar data to what will be used during deployment of the technology.
These models perform better when given access to larger amounts of training data. With more data comes an increased need to balance privacy and regulatory requirements.
Machine learning systems need to consider how biases affect the data being used to train a model and adopt practices to address and monitor them.
A model’s ability to recognize patterns in historical data is both a strength and weakness of the technology.
Historical data could be missing data about patients who should have received care but did not.
Machine learning models require regulatory oversight, legal frameworks and local practices to ensure the systems are developed, deployed and maintained safely.
Relying too heavily on the models during decision support or image analysis could lead to automation bias and decreased physician awareness of errors.
The authors believe that the patient-doctor relationship will be enriched by additional insights from machine learning.
“We look forward to the hopefully not-too-distant future when all medically-relevant data used to make decisions by millions of clinicians in caring for billions of patients are analyzed by machine learning models to assist with the delivery of the best possible care to all patients,” the authors wrote.
Get the best insights in healthcare analytics directly to your inbox.