Kaiser to advance use of AI to personalize care interventions
Kaiser Permanente is teaming with machine learning vendor Medial EarlySign on an initiative to use artificial intelligence to identify which Kaiser patients are most likely to respond to care interventions.
Kaiser has previously worked with the vendor in a project to identify patients with the highest risk of having undiagnosed colorectal cancer while in the early stages of the disease.
Now, Kaiser will use Medial EarlySign’s AlgoAnalyzer platform to stratify populations. It will enable the use of multiple sets of different types of algorithms that can help clinicians, care managers and population health managers predict risk in a population or an individual.
The algorithms also can support analyses to help clinicians understand risks associated with a change in treatment and enable the clinician to decide if the change should be adopted.
The difficult part of using such analyses is learning to trust the analytics, says Micah Thorp, DO, a Kaiser nephrologist in Portland, Ore., specializing in kidney care. “Clinicians need to ask the question, is your predictive model any better than me looking at the patient and potentially changing what I do clinically? That’s the hard part of trusting.”
For physicians just starting to use analytics and unsure where to begin, their first step should be to ask themselves which decisions they most often struggle with, Thorp explains. The second step is to say, “Here’s a clinical treatment decision, but do we have the data to actually predict that one patient with a specific condition will have a problem but another patient with the same condition won’t have a problem?”
Thorp believes that Medial has put together a system to determine if data is available, appropriate and clinicians can assess it quickly and efficiently, but he counsels others to be wary about predictive analytics because it is still an emerging technology that needs to mature.
“This is a big huge area in medicine right now,” he contends. “There’s been a lot of smoke and not a lot of fire about predictive models being used appropriately.”