Outcome-Action Pairing (OAP) is a term heard quite frequently when discussing AI solutions, so let’s break it down:
Outcome = An output of the AI solution model that serves its purpose (predicting incidents, assigning risk, diagnosing diseases)
Action = A response or step taken based on the output
Outcome + Action = Improved Care
Generally, the aim of pairing these components is to substantially improve care. More specifically, it helps point towards likely, impending, or inevitable patient outcomes, and then offers the opportunity to take action to mitigate and/or change the real-life outcome.
As a simple example, an AI solution could predict the likelihood of myocardial infarction (heart attack). That solution could then prompt care providers to take steps with their patients for medical intervention or lifestyle change. When considering this outcome-action pairing framework, there are few aspects to consider that affect clinical feasibility or utility.
Solution developers need to carefully understand lead-time on their solutions for this framework. Does the outcome need immediate attention? Or is a longer-term action acceptable or more suitable? If the output and timing is mismatched (the algorithm cannot process fast enough for an acute action), then the solution may not be feasible. Generally, the longer the lead-time the better as studies show that earlier indications of medical conditions offer more opportunities to mitigate or intervene in health outcomes.
Type of Actions
Actions can fall into two general categories: operational or medical. Does it involve administrative work (e.g., transferring patients or scheduling appointments) or is a procedure (e.g., changing prescriptions, conducting surgery)? While the first refers to the operational and the second to the medical, there could be cases where an action will have overlap (e.g., the outcome predicts hospital readmission which involves both administrative and medical efforts).
One particular level of interest right now is population-level predictions and actions. This concerns populations (i.e., groups of people based on geography, conditions, etc.) and widespread public health incidents (e.g., COVID-19 pandemic). As a population-based example, an AI solution predicts risk for hospital readmission two weeks after discharge: the solution provides an output of patients who will likely need readmission; it then takes an action of generating a list of the highest risk patients for the hospital to use (known as a “chase list”). Quick note: Be sure to involve regulatory bodies as well as typical stakeholders when developing solutions for public health events and relevant population-level predictions.