Let’s Talk Practical AI with Discharge Engagement

Once in a while, you get shown the light
In the strangest of places if you look at it right
R. Hunter

I began my healthcare marketing career in 1994 with Krames Patient Education, now part of WebMD Ignite. As I type this, I realize I’ve been at this for 30 years. But I digress. Around that time, I believe 1995, Krames published it’s CHF study which showed the efficacy of patient education for chronic disease self-management including the post-discharge period. Fast forward to today and amongst our clients, the Discharge Planning CareFlow is our most popular. In between, CMS introduced penalties related to avoidable readmissions for multiple conditions including heart failure. Bottom line, effective discharge planning for smooth transitions of care and readmission reduction has been a persistent goal for health care organizations for at least the three decades I’ve been working in healthcare.

Now take a minute to think about AI. It’s actually been around longer. There has been a cultural boom in the wake of OpenAI releasing the chat-based interface to ChatGPT 3.5 in November 2022. The power of the new generative, large language models was immediately apparent. Exponentially powerful releases have followed over the past 18 months. There has also been an exponential boom in the hype. As a marketeer, the amount of content coming at me about the latest, greatest in AI – products, cheat sheets, including AI-generated soulless content – has been dizzying. I haven’t seen the Gartner Hype Cycle of Generative AI (assuming there is one) but I would guess we may be somewhere between the Peak of Inflated Expectations and The Trough of Disillusionment. It is so easy to bet blinded by the shiny object instead of focusing on problem solving. Which brings me back to discharge planning.

Old Challenges

Despite improvements over time, the process is still plagued with disjointed communication, missed details, and an incomplete picture of the person beyond the patient. Additionally, a lack of post-discharge support can lead to frustrations, readmission, and ultimately, a suboptimal patient experience. Here is a situation ripe for the targeted application of AI – not to replace the replace the human element but to empower it – to help with solving the problem. Let’s break this down to see how it could work. (I don’t claim that I am the first to think about how the following could work. In fact, I’d be surprised if this type of solution is not currently being developed or tested.)

A wealth of data is available. Look beyond the medical record to bring together the right elements for developing prescriptive models for preparing a patient for discharge. Here is the highest level of recipe (As you’ll soon figure out, I’m no data scientist):

• Take the core data from an EHR (age, gender, medical history, race, occupation, address)
• Add intervention data (medications, treatment, education pre- and peri-admission, consultation notes from doctors, nurses, therapists, coordinator) that would help define the current discharge process,
• Layer in social determinants data from compiled geographic data (county, city, zip code, census track)
• Analyze available outcomes data (readmissions, PROs, post-discharge provider consultation data)

The output could suggest an optimal discharge process that would take into consideration both the attributes of the patient, the family, and the facility’s resources. I would imagine that depending on the amount of data modeled, an AI enhanced discharge process could be suggested at the cohort level if not for an individual.  It could identify potential red flags:

• A frequent flier in the ED
• Living alone and difficulty managing stairs
• Limited access to fresh food
• Gaps in medication refills
• No interaction with education resources or completed learning verification questions
• No scheduled post-discharge follow-up appointment.

These elements could be used to create a discharge plan to help nurses and care coordinators proactively collaborate with social workers to arrange home care services that address specific needs.

New Solutions

This is easier said than done thus presenting an opportunity for machine learning and generative AI to provide an assist for all stakeholders. With resources spread thin, it’s a big ask to expect nursing teams juggling immediate care issues to wade through social determinants of health data, home care logistics, and medication adherence concerns to come up with the ideal discharge plan. This is a use case well suited for generative AI. These models are driven by probabilities. A custom prescriptive model using de-identified patient data inputs creates best case solutions based on what has been done and worked in the past. Then generative AI kicks in to help create the discharge plan best suitable for the patient in a format easily accessible for the care team (a cross functional checklist) and understood by the patient via application of health literacy content principles.

This new AI tool could also use robotic process automation to fulfill basic tasks. For example, if the patient has a history of medication non-adherence identified by refill gaps in claims data, a chat bot could text the patient with a medication reminder AND collect authorization to automatically submit the refill request with home delivery to the pharmacy of record. The objective is to offload the repetitive tasks associated with medication adherence unless the patient is non-responsive to the automated outreach.

The point of all this is to identify the persistent problem(s) creating barriers to smooth care transitions. Then creatively apply new processes and tools to solve the problem. This is what makes it Practical AI. There are no silver bullets. And these models in their current state do not think – the plan described above is an execution of probabilities not thought.  Human creativity needs to provide the oversight, make the decisions, and manage what will be rounds of trial and error as solutions are iterated.  But the good news is that rapidly advancing AI may provide the “better mousetrap” to help us finally solve some vexing problems – in our healthcare system and beyond. The challenge for us it to remember to keep focused on useful and targeted use of these tools and not the tools themselves.