By Chad Konchak, AVP, Health IT, NorthShore University HealthSystem
Healthcare to date has seen a tremendous advance in the use of predictive modeling. However, while we have advanced with these abilities, we have also encountered a new problem: with each new predictive model deployed into clinical workflows, we risk creating confusion as clinicians are presented with different and competing scores without a clear “prescription” on what to do about it. The vision should be to integrate all those disparate models into a prescriptive engine that would produce a unified lexicon to identify and communicate patients at high risk of multiple adverse outcomes across the continuum of care; segment or cluster like-patients into sub-populations that share risk profiles and intervention amenability; and to cascade that into a coordinated set of interventions and care pathways to the appropriate care teams. Furthermore, advanced and intelligent algorithms should be applied in order for this engine to learn about the impact of those interventions.
Ever since the Institute of Medicine (IoM) released their vision for a learning health system in 2006, a lot of discussion has evolved on ways we can capture the vast knowledge we are gaining and integrate it into our increasingly sophisticated informatics systems to continuously improve at a more rapid pace. In their article, the IoM dreams of “the development of a continuously learning health system in which science, informatics, incentives, and culture are aligned for continuous improvement and innovation, with best practices seamlessly embedded in the delivery process and new knowledge captured as an integral by-product of the delivery experience”. Moving beyond the ability to simply identify patients at risk of an adverse event to prescribing the most effective intervention that would prevent that outcome should be the ultimate panacea of data science for healthcare, and we should all strive to achieve that.
Obviously, this is hard. In a previous article, I talked about how the predictive modeling itself is the easy part. The real challenge is how we integrate that knowledge into workflows so our clinicians on the front lines can be empowered with better information to do what they do best. A prescriptive engine takes this challenge a step further and presents new ones. Firstly, we want to be extremely careful not to go too far and side-step the judgment of a clinician, which should stand as the final decision point for optimal care. Second, in order to map interventions based on multiple outcomes and population segments of risk, this needs to be carefully understood and planned by teams of clinicians, administrators, and data scientists. Consider, for a moment, readmission risk. In today’s EMR, many systems have a risk score for readmission risk and we have even mapped that to specific workflows. For example, a list might be provided to case managers of high risk readmission patients to help coordinate follow-up appointments to prevent a readmission. However, this treats all high risk patients the same. They are not. That list likely contains a heterogeneous group of patients with a complicated mix of medical and social barriers that are driving that readmission. Now imagine another model that can predict high risk of post discharge mortality. We can integrate those two risk models and create an interesting segmentation of high risk readmission patients: Those that are also at high risk of post discharge mortality and those that are not. The patient who is at high risk of readmission and mortality is likely older and has many comorbidities requiring a more medical-focused decision process. The patient who is at high risk of readmission and low risk of mortality is likely younger and may have more social barriers requiring a different decision process. Again, this analytics engine should not decide what the process should be, but imagine sending the list of high-readmission/high-mortality patients to a case manager with medical expertise (advanced practice nurse) and high-readmission/low-mortality list to a social worker. An engine like this could help us better triage the work effort to the care team best positioned to serve interventions to high risk patients. Consider further that this engine can track the interventions taken and learn about which steps led to better outcomes. What if we were able to track the time and frequency of the follow-up appointments and then use statistical models to align those variables with other patient characteristics and then further sub-segment populations into what follow-up appointment “package” would likely work best for them. The beauty of this engine would be that as more cases are fed into this engine, it can learn and get better. We can then provide additional information in this list for our front line staff to inform how soon the intervention should occur or even help identify additional barriers that need to be addressed in order to protect against that putative readmission.
As our ability to use machine learning and artificial intelligence grows in the world around us, we, as healthcare leaders, need to keep up and take advantage. However, we also need to “first do no harm” by providing the right kind of decision processes that allow our clinical experts to be empowered by better tools and not side-stepped by them. Additionally, we need to make sure we have the right verification processes in place to ensure our algorithms are working as intended. The balance between harnessing the great power of technology while avoiding the dangers will always be a difficult equilibrium to achieve. I am hopeful, though, that our mission to serve our patients will always triumph and that our quest for better will drive us to learn and improve. Analytics lets us do both. I think that’s pretty cool.