Beyond the hype: Opportunities and challenges with AI, machine learning for clinical decision support
Artificial intelligence (AI) for healthcare, while set to change the future of direct patient care, faces a path that is full of complexities. No longer a science fiction trope, the technology can empower medicine in numerous ways. One of the most obvious applications is clinical decision support (CDS), with CDS systems providing clinicians and staff timely information that assists with decisions at the point of care. This information is available in various forms, including actionable alerts, reminders, and diagnostic support.
By Anatoly Postilnik, Head of Healthcare IT practice
But while CDS’are built into every modern EHR today and play an essential role in a provider’s daily workflow, these systems suffer from numerous deficiencies. They generate meaningless alerts contributing to provider burnout. The information they supply is frequently too general or not useful within the unique context of a given patient. Important factors – such as comorbidities, mental states and behavioral indicators– may not be taken in consideration, rendering CDS recommendations ineffective or even dangerous.
It is therefore useful to take a look at the underlying technology that powers CDS systems today to understand how and why some of these obstacles occur, and how machine learning algorithms stand to improve and expand their capabilities.
Utilizing clinical decision support
Most CDS systems are based on clinical guidelines, evidence-based research, and best practices, which are translated into rules that guide them in how to respond to different circumstances and scenarios. Various activities can trigger these rules, from a clinician opening a patient chart in an EHR system to a nurse administering medication.
The library of rules coded into a CDS system represents the clinical knowledge that defines alerts and recommendations produced by this system. This library is maintained by teams of trained medical informaticians who continuously change and update the rules in response to new research, changing guidelines, and best practices.
But even this process comes with its share of complexities and deficiencies. For instance, translating guidelines, best practices, and information extracted from research publications into rules is a massive manual undertaking. Team members must stay informed about the latest changes in relevant information and update the rules accordingly. The number of rules inevitably grows over time as additional research is done and greater knowledge is gained. So, the logic in these systems becomes increasingly complicated, leading to potential conflicts of logic and more errors.
Translating information into rules is also subjective. Two different teams are likely to express and code the same information into rules differently. As a result, two CDS systems will behave differently even if the information that was used to define each one’s rules is the same.
ML algorithms can alleviate the effort for rules development and maintenance, and reduce the number of potential errors and conflicts. By using a rules-based system and recommendations that they generate as a reference baseline, they can learn on a training data set and take patient-specific characteristics into consideration, making these recommendations more relevant, targeted, and precise.
They also can perform numerous decision support tasks that conventional rules-based systems are simply unable to perform, including extracting and presenting diagnostic information for clinicians from radiology images, such as the presence of cancer cells, or types of skin abnormalities from photographs supplied by the patient. AI-powered Natural Language Processing (NLP) simplifies interactions with EHR systems, and provides new ways to deliver decision support to clinicians.
Challenges with ML algorithms in CDS systems
The quality of recommendations from ML algorithms is entirely dependent on the data sets these algorithms are trained on. The size of the data set may be too small or may contain “bias”, for example if the distribution of patient characteristics doesn’t realistically reflect the general population of patients.
Scientists recently found evidence of racial bias in a widely used algorithm that was meant to predict which patients will benefit from extra medical care. However, the algorithm was based on using healthcare costs to establish illness severity.
“Less money is spent on black patients who have the same level of need, and the algorithm thus falsely concludes that black patients are healthier than equally sick white patients,” said the scientists. “Reformulating the algorithm so that it no longer uses cost as a proxy for needs eliminates the racial bias in predicting who needs extra care.”
Data can also be of poor quality or insufficiently structured, which can be challenging for ensuring a seamless use of ML algorithms. Additionally, ML algorithms are essentially “black boxes” – it is not very easy to understand why a given recommendation was generated, while decisions produced by the rules-based systems can usually be traversed and justified.
There are also ethical implications in using AI and ML technologies on decision support systems. Several leading radiology, medical physics and imaging informatics groups released a statement earlier this year on the ethical use of AI in radiology, saying that while AI has the potential to improve patient care and physician workflows, an ethical framework will “help steer technological development, influence how different stakeholders respond to and use AI, and implement these tools to make best decisions and actions for patients.”
The future of CDS
CDS systems are an integral part of modern EHR systems. But major EHR vendors are not known for being great innovators and have been slow to incorporate AI/ML. These innovations generally come from new, dynamic technology companies or even from other industries.
Some EHR vendors, however,have begun to recognize the opportunities that AI brings to the table and are starting to collaborate with leading AI players to bring these opportunities to clinicians and organizations. While the practical adoption of AI in CDS (as in medicine in general) is not free of barriers, and the immediate benefits are often exaggerated, there continues to be rapid advances in its capabilities and ease of use. New open source tools are available and even some cloud platforms provide AI/ML modules for use. Innovative providers and health systems are starting to play with these tools. The development of custom solutions, unique to given situations, may be the fastest and most efficient way to move forward while waiting for these tools to be embedded into an EHR solution.