CHICAGO--Decision support technology can help radiologists improve accuracy of diagnosis, lead to more personalized medicine and provide an alternative to radiology benefits managers, according to a presentation on Nov. 30 at the 97th Scientific Assembly and Annual Meeting of the Radiological Society of North America (RSNA).
There is a key difference between decision support tools and traditional methods of education, said Charles E. Kahn, Jr., MD, of the Medical College of Wisconsin in Milwaukee. Decision support programs should be considered “active” methods of aiding radiologists, as opposed to “passive” educational methods such as medical literature or textbooks. Even information retrieval programs put out by organizations such as the International Society of Radiology are not intended to be used at the point-of-order.
“Education is…knowledge when you can’t use it, but decision support is information that you get when you are really too busy to learn anything,” said Kahn.
Active decision support systems are applications that help radiologists provide patient-specific care and aid in interpretation. Different programs use different methods, with some being able to guide procedure selection and others using computer-aided detection. For instance, Kahn demonstrated how one program could take American College of Radiology appropriateness guidelines showing low utility for lumbar spine MRI and offer alternate procedures to consider. A radiologist can follow the guidelines or proceed while offering an explanation of why he or she is going against the recommendations.
Kahn described several different processes these programs can use, including:
- Rule-based reasoning: The model for the earliest systems, rule-based programs make recommendations based on ‘if-then’ statements. An advantage to this method is that it’s easy to generate explanations for a recommendation, although it can be difficult to maintain a large system of rules as there may be “collisions” where two rules contradict each other.
- Case-based reasoning: Uses similar previously encountered cases to generate recommendations. This method feels more natural as it mirrors the way people actually think and has the advantage of being able to function with incomplete knowledge or without a total causal model, said Kahn.
- Artificial neural networks: Based on a series of input nodes (such as a patient information) and outputs (such as conditions), artificial neural networks are able to connect the dots between inputs and actually learn through training. The challenge is that the networks can’t explain reasoning, and systems can be over or undertrained.
- Bayesian networks: Similar to artificial neural networks, Bayesian networks connect the dots between diagnoses and contributing factors, but are based heavily on probability values.
Additionally, computer-aided detection programs can assist radiologists in tasks such as lesion detection. Kahn pointed to a study published by the American Journal of Roentgenology that showed these systems improved diagnostic performance compared with radiologists making unaided interpretations.
While the idea of having a computer program offering instructions may feel unsatisfying to some radiologists, Kahn stressed that these systems are not designed to usurp the role of a trained professional, but rather to provide an additional tool. They may be more palatable when compared with authorization from a radiology benefits manager, he said.