Accurate, non-hallucinating medical references with AI?
🥁 drum roll please...
In early April, we began discovering and experimenting with the idea of references and citations based upon appeal context.
Early on in denyify's development, we prompted for direct references but the AI hallucinations were too severe that there was no good production use case... until now!
Clinical Citations (a beta program)
Featuring the cutting-edge of applied AI, Clinical Citations uses contextual claims data to search and return evidence-based reasoning with appropriate and relevant references.
For example, take a patient who has been suffering with vascular dementia and as a result has been experiencing behavioral disturbances (such as falls). The prescription of a durable mobility aid (such as a walker or cane) would appear reasonable, appropriate, cost effective, and medically necessary.
Let's take those treatment/condition inputs and see what our clinical citation AI can find.
Treatment: walker or cane as a mobility aid
Condition: vascular dementia with behavioral disturbance
Clinical Citation Output: The recommended treatment for mobility assistance is to use either a walker or a cane.The individual has vascular dementia, which is a type of dementia caused by reduced blood flow to the brain. They also exhibit behavioral disturbances.The article "Walking and Dementia in Physically Capable Elderly Men" published in JAMA suggests that physical activity, such as walking, may reduce the risk of developing dementia in elderly men who are physically capable [1]. The study found that men who walked less than a quarter mile per day had a higher risk of developing dementia compared to those who walked more than two miles per day [1]. Therefore, prescribing regular walking as a form of physical activity may be medically necessary for physically capable elderly men to reduce their risk of developing dementia. [1] (https://jamanetwork.com/journals/jama/fullarticle/199484)
As you can see, we're
almost
there. The treatment/condition were found paired in research found in a JAMA journal, however the nuance of "fall risk" or otherwise was excluded and the use case was positioned as benefit to the patient (e.g. how walking can help with their condition).Let's try it again, this time with a specific ICD10: Z91. 81 (History of falling).
Treatment: walker or cane as a mobility aid
Condition: vascular dementia with behavioral disturbance and a history of falling (Z91.81)
Clinical Citation Output: The treatment involves the use of a walker, which is a mobility aid.The person has dementia and a history of falls.Regular exercise has been shown to reduce the likelihood of falling in older adults, even if the exercise was performed a quarter of a century earlier [1]. Additionally, gait speed has been identified as a predictor of survival in older adults, with slower gait speeds being associated with increased mortality rates [2]. These findings suggest that exercise interventions aimed at improving gait speed and balance may be medically necessary for older adults to reduce the risk of falls and improve overall survival. [1] (https://pubmed.ncbi.nlm.nih.gov/28369185/), [2] (https://jamanetwork.com/journals/jama/fullarticle/644554)
While the argument being made makes sense, we are still left with whether or not the patient history supports this reasoning, whether the appeal itself would benefit from these citations, and overall if a medical director would feel sufficiently persuaded.
Source [1]
Source [2]
Now, all hope is not lost, as we have covered an enormous ground in being able to take patient treatment/conditions data, generate an appeal, and then use that same data to generate citations which support the argument in-appeal.
It's
absolutely
not perfect, and sharing this is no intention of marketing as it will not be ready for some time. The goal of this update is to share with our community of current users, prospective users, partners, and potential partners, the power of denyify and what we're building here.We're currently green-lighting this for access to Beyond plan users exclusively, and look forward to rolling this out sometime in late july or early August 2023.