Some VoiceAI in Healthcare thoughts

This episode of Out-Of-Pocket was brought to you by…

The Healthcare Compliance Checklist

Building a healthcare compliance program requires not only meeting today's regulatory demands, but also preparing for tomorrow's challenges. 

Trust is everything in the healthcare industry. Patients, partners, and regulators all expect strong safeguards around sensitive health data.

This checklist will guide you through actionable steps for building and maturing a healthcare compliance program, whether you're an early-stage startup preparing to become HIPAA compliant or a scaling company ready for HITRUST.

→Download it here. 

–

Why voice AI is useful in healthcare

At our AI “buildathon” last month, I saw a ton of companies build with the Voice AI tools. Companies using bots to automate calls, turning voices into text data, or giving a voice interface for humans to converse with.

[BTW our roundup of the AI projects went to a lot of spam folders because the h-word triggers spam. They’re extremely cool, check them out here because they’ll probably inspire something for you to build at your job.]

Voice is particularly interesting in healthcare for a few reasons.

• A ton of important information is transmitted in conversations and phone calls. Your visit with a doctor, a reminder call about an upcoming appointment, your call to an insurance company, talking to the pharmacist, etc.
• In areas where Application Programming Interfaces (APIs) don’t exist to easily pull data, you can “hack around” this by calling someone regularly and asking for the information (e.g. your local pharmacy doesn’t have an API that shows their current inventory, but you can call to ask). 
• It gives bots real personality. If you try Sesame’s voice AI model for example, you can really see how we’re wired to feel connection and emotion via talking in a way typing into a chatbox doesn’t hit

A few disparate vignettes about Voice AI in healthcare. 

But first…

{{interlude 4}}

Also as a learning experience - we’re making our own companion apps for students! We worked with MindK to launch this chrome extension that explains jargon on all the major healthcare sites + social media.

This has been a fun experience for us to learn about what you can do with AI today. And also chrome extensions, which I think are still very underutilized in healthcare as a sidecar for your browser.

‍

Is Voice AI a feature or a new process?

All AI companies look up to the sky on a star filled night and think to themselves "Did God create me as a feature or a company".

In the last month we saw Zocdoc, Cedar, Epic, Solv, and more rolling out their own genAI voice/conversational AI for things like scheduling, medical billing, etc. We're now seeing a real test in where defensibility lies in genAI applications. I’m sure every voiceAI founder selling into healthcare got the very helpful “have you seen this?” email from their investors.

Companies with distribution are rolling out their own solutions with off the shelf models, but:

1) Do they perform well vs. a company that focuses specifically on this problem? Where do you need best-in-class solutions vs. "generally pretty good"? A scribe needs to get nearly every single word right otherwise it doesn't actually save time. A scheduling module probably gets a bit more leeway.

2) When do you need to do things that can be contained within a platform vs. need to deal with external workflows? 

For example, scheduling an appointment can be contained within a Zocdoc or Epic. But an end-to-end revenue cycle company might take everything from appointment scheduling + eligibility check all the way through submitting a claim (which requires doing a bunch of other things that don't just live in the scheduling module).

3) Do these existing healthcare companies have legacy constraints that work against them? A patient intake form that needs to go through the EHR is constrained to a lot of the available fields. If you made a totally new patient intake form that was flexible, had new modalities like video, etc. then it might make sense to rebuild what that flow looks like from the ground up and how it changes scheduling, triaging, etc.

Scribes capture raw data, a new paradigm

There’s a lot of discourse about how “scribes are tools meant to help systems optimize for billing”. That’s probably not wrong but I think it misses a more interesting point.

Currently, most healthcare data is extremely lossy. When a patient talks to a doctor, the doctor then filters what they think is important and puts it in the record. They say everything else is just you being stressed, and have you considered being less stressed? 

Today, a big portion of the documentation doctors put in are for things that are relevant for claims, billing, etc. This might not necessarily be all the CLINICALLY useful things though. If something a patient brought up won’t change how you bill, or how you treat the patient, then all that additional documentation could just bloat the note.  

IMO what’s cool about ambient dictation is that we are moving to a world with much lower loss rates because the raw encounter data with the patient is captured. One transformation of that raw data is for billing, but capturing that raw audio data means you can transform it for more things!

For example, I think ambient dictation companies will be able to find disease trends way before published research. Right now there's an enormous lag (decades) between docs noticing a trend > research getting done to prove it > it becoming normalized in clinical practice. During that in-between time only $50K+/year concierge docs, gym rats, and the three DGAF docs academic medical centers (all about to retire) will be incorporating that trend into their practice.  

Ambient dictation creates a brain that lives across physicians - if one patient says that they're getting intense migraines, it's tough for a doc to know if it's from the new drug or if the patient is just stressed. But if a dictation company hears 100 patients in the same time span talk about it with their doctor, they can surface to the doctor that other people are hearing this. 

The great part is with the raw data, you can query retrospectively too. The doctor might not note every single side effect you mention to them on the drug, and just keep the ones in the note that seem most relevant. But if you later discover that actually the lack of sleep the patient mentioned WAS related to the drug, scribes now have the raw audio data to search through vs. the note which may not have thought it was relevant to write down at the time.

I think this will shorten the time between anecdotes and changes to clinical practice because it will surface trends much faster. We already have analogous examples of this in infectious disease outbreaks, where the time lag makes a huge difference. Here’s an example of a paper that mined free-text notes across EHRs to detect COVID symptoms in populations in advance of the outbreaks. Here’s a preprint that looked across athenahealth EHRs from 72K providers to identify influenza outbreaks.

It’s clear that having a global brain that lives across these physicians can bring evidence to the forefront much more quickly. But I have no idea if ambient scribes get permission to use data from individual hospitals in a global brain, I just think it would be cool and useful.

[Interesting side question: do you think AI scribes increase or decrease total billing? If we recorded every visit would providers capture more codes and make it easier to get the documentation needed for claims? Or would we discover that a lot of the claims are actually upcoded relative to what ACTUALLY happened during the visit? Looking at you 99285s.]

Open questions about Voice: Voice to Voice? How do you price Voice?

As voice AI matures, it’s clear that phone lines are going to get more bombarded. Some things I’m watching

• As voice agents start talking to other voice agents, what’s the middle layer that makes it more efficient to transmit communication between them? GibberLink is one example where the bots create a new information dense language to speak to each other. But maybe there are others.
• How does pricing change? If a company charges per resolution, it can’t make sense if hundreds of thousands of AI inbound calls are then being resolved by other AI agents.

• We need a new caller ID. The inability to triage importance of an inbound call is a massive hindrance to anyone. There needs to be some new way to triage the bot calls to figure out if they’re worth allowing through.
• Where is the line around impersonation in voice? There’s automated emails that can come from your doctor to remind you of something (schedule an appointment, take your pills, etc.). But should an automated voice call that SOUNDS like your doctor be allowed? 
• If we assume everyone is recording at all times for their job, what does privacy look like? There are some conversations that are not meant for listening ears, and sometimes people are more honest if they know they’re not being recorded. How can we still have those conversations in the future? Even zoom calls ain't safe.

Thinkboi out,

Nikhil aka. “My old roommate trained a model on my podcast clips so it sounded exactly like me and would text me passive aggressive voice notes in my own voice, horrifying stuff.”

Twitter: ​@nikillinit​

IG: ​@outofpockethealth​

Other posts: ​outofpocket.health/posts​

--

{{sub-form}}

‎If you’re enjoying the newsletter, do me a solid and shoot this over to a friend or healthcare slack channel and tell them to sign up. The line between unemployment and founder of a startup is traction and whether your parents believe you have a job.