A Healthcare AI Use Case

In my previous blogpost, Andy Slavitt gets practical – no one cares about AI’s capabilities unless we have applications for them. With AI being generally seen as an umbrella term that encompasses discovery of patterns in data, and making predictions based on them continuously, fast, and ever improving toward its reward function, let’s look at a use case I have personally worked on.

A client with 1,000+ monthly new patients going through their specialty-care clinics needs to deliver at least sufficient care to each patient.  Their workflow involves scheduling the new patient for an initial consult, then as appropriate, proceeding with testing, followed by treatment and follow-ups. Simple enough, one may say. BUT as the workflow spans a timeframe of several days and in many cases weeks, and every patient may have a series of interactions with different care team members over the course of his or her specific care pathway, tracking patients and coordinating the next actions start to bog down the team. Worse yet, any misses may materialize as lost revenues, dissatisfied patients, or even patient safety issues.

My job is to tell who are the dropouts (from their respective pathways), estimate the likely impacts in terms of lost revenues and/or at-risk patients, and coordinate the team to reengage the dropouts to improve revenues and patient outcomes. I can do that, but I just can’t do it all the time and instantly for any one client.  

I would leave out how I capture what I would do for them in an AI-powered, workflow-aware care collaboration tool for now to focus instead on how AI helps to resolve the 3 challenges in their workflow:

1.   Every patient is different – it is not about pushing a number of widgets down an assembly line. Every interaction, clinical or not, potentially can change the course of the pathway. On top of that, care team members are not robots in the assembly line either, and may interpret and react differently to end up with a different course of action.

2.   The default solutions are the EMR and office applications that are not aware of the workflow – the care team would write voluminous notes into the EMR, and create Excel spreadsheets to track each patient through his or her pathway. Tricks such as setting up pending future encounters may work in some cases, but they remain as tricks with limited applicability in the workflow. The notes contribute to the 80% of the data “IDC FutureScape: Worldwide Healthcare 2015 Predictions” dubbed as unstructured and posing hindrance for even the EMR itself to analyze.

3.   The sheer volume of patients – every patient deserves the appropriate care attentions. A trained care team can only provide that to its capacity associated with their current default solutions.

Every patient is different, but there are general goals that the care team and the patient share. Let’s say the care team simply wants to move as many patients into testing and treatment stages as long as it is clinically appropriate, and the patient is willing. One setup is to have AI learn how to make predictions from existing data, and provide it with rewards on the quality of its predictions. The rewards may be based on the clinical conditions and patient situations that determine whether s/he should proceed to the next stage to help the AI optimize its predictions. Then on an ongoing basis the AI can bring to the care team’s attentions patients who should have moved to the next stage, but haven’t.

For the AI to include signals from notes, I resorted to Natural Language Processing (NLP) to extract both clinical conditions and patient situations.  For example:

•  “… lvm x3 …”
• “… pt told me he is moving to Spokane to care for her ailing mother …”
• “… pt’s wife called to cancel the testing appointment, and mentioned pt was laid off. They cannot pay for the procedure… ”
• “... pt has central apnea ...”

may be represented to the AI as “difficult to get in touch”, “relocated”, “not able to pay”, and “central apnea”. Using these as part of the reward function in its training with the existing data, the AI may predict that the first 3 patients would not move to the next stage, and the last patient would move to the treatment stage for Adaptive Servo-Ventilation (ASV) therapy.

The deployed AI would be able to detect dropouts from a patient volume orders of magnitude larger than any care team can handle in their patient tracking efforts. Then in the daily huddles, the care team may focus on patient-care issues instead of the multiple Excel spreadsheets that serve roughly the same patient tracking functions, and come with a disclaimer: “… while this is not the most up-to-date, and certainly not comprehensive, I have the statuses of most of the patients I have worked with in the last week here.” 

Imagine the efforts saved, additional time available to patients, and a happier care team!