GenAI Applications

These are continuously-evolving notes and observations on what healthcare-specific use cases make sense for Generative AI. At times I try to connect them with the companies I see out there, with the full realization that I'm only interpreting what these companies claim to be doing based on their online presence.

Provider Org Operations

There are a lot of workflows in a healthcare organization that depend on manual back and forth of papers, faxes, phone calls. Multi-modal agents focused on logistical use cases can definitely reduce the workforce needed and even make overall efficiency gains (since they can parallel-process calls, navigate IVR, score calls by hold times, understand bad responses, deduce policies, escalate if needed, create call summaries, etc). Yes, they need to be integrated with the incumbent EHRs in large organizations. But for smaller, independent deployments I can see these offerings can actually be a wedge that ends up creating a master system that EHRs are subservient to.

1. Scheduling: No-shows, cancellations, reminders, changes. Multimodal agents can help everywhere. Esp. in outpatient settings. I used to keep a list of startups in this space.
2. Billing
   1. Coding: Clinical documentation improvement (CDI). Mapping to medical coding systems (like ICD, CPT, HCPCS etc) so the most viable claims can be submitted for payment. Learning the shorthands, variations of this crosswalk is a big issue.
   2. Payer Compliance: Payers have a lot of policies, regulations and they change frequently. Not adhering to them in can cause payment denials and delayed/lost revenue. Evaluating the adherence and consistently navigating all this bureaucratic maze is a big problem.
3. Banal communications: Multi-modal agents can start dealing with phone calls or fax/email conversations around frequently occurring topics like:
   1. Collecting insurance information from patients: medical/pharmacy insurance plans have bureaucratic details like plan number, group ID, effective date, PBM name, etc. Patient's don't know this and there is back/forth in discovering this structured information.
   2. Managing requests for letters of medical necessity
4. Special Communications:
   1. Verifying benefits with payers: Getting some basic eligibility information is possible through EDI (270/71) transactions but high-stakes treatments (like with $$$ drugs, infusions) require making phone calls, exchanging documents to figure out details (like what is the cost share, will a prior-auth be required, status updates, etc.). Without those confirmations, treatment can't be started. And once it does get started, there is a need to do milestone-based reconfirmations to check if the patient is still eligible.
   2. Prior Authorizations: The friction of PAs costs the average physician practice nearly two business days a week. Figuring out if PA is needed, requirements to submit a PA, status/updates on a submitted one, etc. is all very burdensome. This back and forth is a poster child for LLM-based automation. Which is why multiple startups are in it already.
   3. Formulary inquiries and submitting exceptions: Health plans have a list (formulary) of covered medications and typically for the expensive ones needed for a patient, it's worth checking ahead of starting the prescription. If it's outside the formulary, submitting an exception request requires specific paperwork and forms.

Care Delivery Support

Gen AI agents have a never-before fluent ability to have conversations and that can be use to support care delivery (for now, mostly in non-clinical contexts). No company is claiming anything that can be construed to replace a clinician, of course. But these use cases need awareness of care pathway - and this context makes them different from the operationally-focused use cases.

1. Clinician's Scribe: Creating valid clinical documentation from clinician-patient conversation is the poster child for this category. AI assistant was a hot category in the recent past (list) and post-LLM gold rush is happening with new players like Abridge.
2. Medical Record Summarization: Visit summaries for patients, TLDR Case Summaries for clinicians, discharge summaries. Big models are already tuned for summarization and applying it to specific clinical domains will yield good results, IMO.
3. Prioritized Inbox: In March 2020, CMS introduced CPT codes 99421, 99422, 99423 giving providers the opportunity to bill for time spent on patient communication. With patient panels of >2000, physicians can use help in figuring out which emails to answer and how to answer them.
4. Referrals: They are the highways of healthcare, controlling revenue-generating traffic (mostly from PCPs to specialists). Lots of context-specific formal content, inter-facility communication, scheduling, confirmation, reminder, followups, etc. happen - everything can be automated to a least some extent.
5. Banal Communication
   1. Prescription-related: follow-up, renewals, changes
   2. Visit related: Referral follow-up, upcoming appointment reminders, FAQs
6. Special Communication
   1. Conversational guidance: A co-pilot for having more successful phone calls with patients. Understanding their personal health contexts and coaching the care navigator/specialists real time. Note LagunaHealth. This becomes especially powerful if the business logic gets embedded in creating the guidance output. Eg. steering the conversation towards higher revenue-generation offerings.
   2. Post-op check-ins: I personally think that this starts getting complicated, fast. But bold PR from companies like Hippocratic AI keeps intriguing me.

Patient Tools

1. Charity Care: Most hospitals offer discounts or bill forgiveness based on income with a formal policy called “charity care" that often goes under utilized because it's hard to decipher and requires paper and phone coordination. The work that non-profits like DollarFor do can be scaled up with agents.
2. Personal Sage: Individuals exploring a health context can easily end up accumulating an overwhelming amount of information, even if they ignore the scientific, social/consumer generated sources. RAG tools can help become a customized navigator through that maze of information. I did this for my own exploration of Mitral Valve Replacement topic using NoteBookLM and was pleasantly surprised (esp. at being able to find highly relevant references back to original document sections). It's brilliant because the more information you shovel into it, the better your conversation outcomes become. Which is the opposite of what status quo used to be.