AI Revolution in Healthcare

By /

Vijay Pande, PhD and Julie Yoo , from a16z Bio + Health, spoke on their podcast about Grand Challenges for AI in Healthcare. It was a nice 360 degree view of the field.

Here is my summary of the main ideas – no substitute for the actual podcast, but a way to orient the listener and underline some of the things brought up.

AI and Revenue Cycle Management (RCM)

Vijay Pande and Julie Yoo brought up RCM first. The feeling I have is – RCM could see the quickest efficiency improvement with language models.

However, not much detail was provided on the podcast about RCM. I would have liked to see nuances – where can RCM applications be more successful, more quickly?

Would it be RCM using workflow tool management? Or RCM as payment platforms? With small clinics as customers? Or large hospital networks?

… Or, maybe, Durable Medical Equipment provider customers?

  • You could – Vijay Pande, PhD and Julia Yoo say – fully automate the entire RCM process in healthcare with AI.
  • The holy grail, here, would be – according to them – getting to real time payments for medical services, or for durable medical equipment.
  • This could lead to a 30% savings in healthcare revenue cycle management costs

Shameless plug: I specialize in full-stack AI design for revenue cycle management.

Challenges, here, are entrenched processes, and the difficulty of digitizing unstructured data. However, a number of companies are making progress in this space.

One notable such company is Turquoise Heath, a Series B startup in the a16z portfolio.

Leveraging AI for Contract Analysis

Turquoise Health uses AI to analyze complex payer/provider contracts.

  • Average payer/provider contract, covering hundreds of millions of dollars, is several hundred pages long, completely monolithic
  • Any one line in that contract can have huge implications for the revenue for that provider, as well as cost structure for payer
  • Yet those things don’t get litigated but for every 2 years, when they come up for renegotiation
  • No one looking at that line – they look at whole thing
  • What if you were to digitize structured data in that contract, and be able to run scenarios on it
  • Turquoise analytics tell you – what if the price for these 10 services was this, versus that. What would the implication be for those payment flows
  • This service could facilitate redoing contracts quicker than every 2 years

But, what if we could have an always-on clinical trial infrastructure in our country?

Continuous Clinical Trials

Imagine a continuous, always-on infrastructure for clinical trials in the country, where AI could enable retrospective and prospective data analysis. This could vastly improve the speed and efficiency of drug trials and overall healthcare outcomes.

  • You could, on demand, slice & dice the structure of the population
  • And produce data analysis retrospectively, or prospectively
  • And build causal relations, between which person followed what treatment or took which drug

Bayesian statistics would be well suited to understand causality.

  • Once you built this infrastructure, new drugs could be tested,
  • You could jointly optimize for health, and for cost

This is a complex data and logistics problem…. What other grand ideas can be envisioned by Vijay Pande and Julie Yoo?

Grand Ideas

  • Create AI-based Patient Journals, day to day analyzing what the patient does, how that impacts patient health. Also, with Doctor Journals, analyze what the doctor does, and debug the clinical approach.
  • With all the data available, you could do A/B testing in healthcare, just as it is done in engineering
  • Or, create a spot market for procedure price in healthcare
  • Or, machine learning could do analytics – look at what line items are useless, bring no revenue – and what line items are effective.

What problems does Julie Yoo like to solve?

Julie Yoo’s preferred investments

Julie likes problems in healthcare where there is a mismatch between supply and demand. The choice of companies in her portfolio reflects that.

  • For example, these can be in the scheduling space. A lot of capacity in the hospital system gets wasted.
  • In absence of reliable scheduling, doctors like to over-reserve time slots for shared resources, like surgery rooms, leading to overall inefficiency.
  • There is a need for a more modern systems of record, as a prerequisite. Workflow of doctors must change.

These are companies like Devoted Health:

A New Era for Digital Data

5 years ago, most doctors did not use digital data. We just entered a new era – and data is still not yet fully unleashed. This explosion of data in healthcare could take years to unfold:

  • Hospitals area struggling financially. They will want to monetize data – make it available to third parties, give others access to train models.
  • Hospital staffing crisis is a tailwind for AI, as was COVID (better called ‘forcing function’ than ‘tailwind’…)
  • At the last JP Morgan conference, providers talked not about what they want to do with AI in theory, but what they do in practice. That is a change since last year.

Another tailwind is the business model change – moving from fee-for-service to value care.

The shift from fee-for-service to value-based care models presents a unique opportunity for AI to play a catalytic role. AI can help redefine efficiency and effectiveness in healthcare delivery, supporting the transition to models that prioritize patient outcomes over service volume.

Risks of Using AI

A bunch of national payers are right now sued.

  • They took rules that humans wrote, and that humans were executing slowly – and are doing it with AI faster.
  • Denial rate is going up – and tech is blamed, instead of blaming the rules.
  • This will happen a lot going forward – especially when tech is build on top of an existing broken system underneath it.

AI and Electronic Health Records (EHR)

The future of EHR could be shaped by AI, transforming it from a static repository of information to an interactive, responsive system that respects HIPAA boundaries but provides actionable insights.

  • You can think of the LLM not as an oracle, but as a user interface.
  • With LLMs, data in the EHR can be queried, synthesized.
  • Data is must still be partitioned, though, to not cross HIPAA boundaries

But, the use of LLMs needs to be more than just connecting to GPT4:

  • There’s a need specialist, medical model. This is a  big area of development
  • Companies already summarize medical records. They look at how to tell the patient story, not just look at numbers, including the social determinants.
  • Other companies do medical scribe application – written as a story.

Use Cases of LLMs

Today’s healthcare info is very sporadic. If you’re healthy, you rarely see the doctor.

  • One possibility is to use wearables, agents, or monitors, and have continuous updates.
  • Another idea is – build a copilot for each team member
  • Yet another use case for LLMs is for triage – for example, suggesting to patients the site of care. Patients ask – should I stay home, go to the Emergency Department? That is a critical role. This is a role LLMs can play.
  • LLMs could be a better alternative to just searching on Google, or to curated web sites like WebMD.

Regulatory Challenges and the Future of AI in Healthcare

As AI continues to permeate healthcare, the regulatory landscape will need to evolve. Understanding the balance between automation and human oversight will be crucial as we explore the boundaries of what AI can achieve in healthcare.

The regulatory aspect, at this point, is still somewhat uncharted territory.

How can you embody full stack of AI doctor implementation?

One good approach, says Vijay Pande, is to understand which decisions are complex, which are simple. And which answers are robust to mistakes, or are not robust to mistakes.

  • Things that are simple, and robust to mistakes, can already be done by ML
  • Things that are simple, and not robust to mistakes – are hard. Self driving, for example, is still unsolved.

Where Vijay Pande sees opportunity is with things that are complex, and that are not robust to mistakes:

  • Start with AI for Nursing, for example. We have seen this approach with Hippocratic AI (https://www.hippocraticai.com/). Then, you’re not doing diagnosing. And are not causing as much potential harm.
  • Maybe, then, if that is solved – work your way to Physician Assistant. From there, to General Practitioner, which often sends the patient to a specialist.

Conclusion

This podcast sheds light on the significant impacts AI could have on various aspects of healthcare, from administration to direct patient care. The potential for AI to transform the healthcare system is immense, provided that innovations continue to navigate the complex regulatory and practical challenges of the industry.

Listen to the Full Conversation

For deeper insights into these transformative ideas, listen to the full episode of the “Grand Challenges in Healthcare” podcast here.

Andrei Radulescu-Banu is Founder of Analytiq Hub. We develop data and AI workflows for healthcare, revenue cycle management, and robotics.

Scroll to Top