Generative AI in Healthcare: "The biggest challenge is in PR"
HLTH 2023: The noise surrounding generative AI in healthcare is intensifying. The key questions for startups are: what is your hallucination rate and what's your unique dataset?
This is a monthly newsletter of Faces of Digital Health - a podcast that explores the diversity of healthcare systems and healthcare innovation worldwide. Interviews with policymakers, entrepreneurs, and clinicians provide the listeners with insights into market specifics, go-to-market strategies, barriers to success, characteristics of different healthcare systems, challenges of healthcare systems, and access to healthcare. Find out more on the website, tune in on Spotify or iTunes.
Generative AI is here to stay. It’s not just hype. Who can do it properly, is.
In healthcare, generative AI - a subset of artificial intelligence technologies that employ advanced machine learning algorithms to generate content, solutions, or outcomes that weren't explicitly programmed into them - has many potential applications (see the list at the bottom of the page). Currently, healthcare providers are widely adopting solutions that can replace arduous tasks such as medical note-taking. But it can be hard to pick the right solution.
What does it take to create a good generative AI product? Years.
“The biggest challenge is PR,” said Manny Krakaris CEO at Augmedix at HLTH 2023 in Las Vegas, when I asked him about the hardest thing his company is facing when incorporating generative AI in its products. After the introduction of ChatGPT, new companies are popping up every month, claiming they can solve various healthcare operational problems with generative AI. “During a startup pitch competition, 60 percent of the startups are doing an ambient documentation solution or part of that,” said Justin Norden, VC and Partner at GSR Ventures, in our discussion about what has changed since we last spoke a few months ago.
The tricky thing with generative AI is that creating a demo representing a seemingly impressive product is easy. It’s also easy to make claims that can be hard to verify.
“The only weapon that a startup has is the press. That's to be expected, no matter what industry you're in. We've got customers, we've got products, we've got revenue, we're public. It's going to take a lot of time, effort, and money for any startup to get to where we are,” says Manny Krakaris. Augmedix has been working in the space for 11 years, developing generative AI in combination with other technologies for the last two years. But LLMs have made machine learning a lot easier, a lot faster, said Manny Krakaris.
AI-supported Medical Scribes
The healthcare market has had several established players using natural language processing in ambient technology for medical note-taking well before 2023, Nuance being probably the most recognized player. The company was acquired by Microsoft in 2018. Abridge is a Pittsburg-based company founded in 2018 and backed by UPMC. It was first a telehealth-focused startup that used natural language processing to create after-visit summaries – enabling patients to revisit their physicians' instructions after remote consults.
Comparison of AI scribes in July 2023. Disclaimer: This assessment was prepared by Nabla, see the full post about the analysis methodology.
Permanente Medical Group, which provides care to Kaiser Permanente members and patients, and is the largest physician-led medical group in the US, is currently rolling out Nabla Copilot for its 10,000 physicians. Nabla is a French company with a growing team in the US, providing clinicians globally with their AI-enabled medical notes scribe. To build their large language model, they opened a digital clinic three years ago, ran it for a year and a half, and while providing care, gathered data for a large medical language model. Patients were aware of the data gathering. Today, Nabla’s Copilot is readily available as a browser extension or a mobile app. As a European company, operating under strict GDPR rules, Nabla’s Copilot doesn’t gather data. “This is an important differentiator from competitors and a feature liked by many US medical professionals, who don’t have to worry about patient privacy,” said Delphine Groll, COO at Nabla.
Compared to early 2023, conversations about generative AI have swiftly evolved. “We're starting to build interesting use cases. The questions now are - what guardrails need to be put in place so this works? How can we test and verify that this new solution we built delivers its intended purpose and is safe? What are the sustainable business models? What are the regulations that are going to come around generative AI,” said Justin Norden about his observations.
AI for Data Management
AI has been making data organization and presentation a lot easier in the last few years, especially in the US, because the amount of documentation that a given patient generates is increasing. To decrease the documentation burden, health systems in the US are starting to charge patients for asking for their physicians' advice through online patient portals. “The documentation burden keeps increasing. This means that you need more and more time from the clinicians and the nurses and everybody to deal with data. Which is unfortunate because that takes a lot of time and that takes time away from patient care,” said Matt Hollingsworth, CEO of Carta Healthcare, which streamlines clinical documentation and enables data-driven improvement through analysis of patient data.
Technology is key in data management, but finding a good system, implementing it successfully, and training the staff properly can be anything but easy.
The battle for customers in the US is harsh. “If you want to replace an existing solution, healthcare providers expect a 10-fold ROI,” said Jay Ackermann, CEO and president of Reveleer, a healthcare technology workflow, data, and analytics company, supporting payers and risk-bearing providers in their value-based care programs, right before we sat down to talk about data analytics and data gathering in the US healthcare.
Is Interoperability Still a Problem?
Absolutely, but companies in the US have long found ways around it by consolidating data and offering new products either for easing the view of a patient for individuals, researchers or clinicians for clinical care.
In case you are not familiar with the American data silo problem: the US healthcare system operates as a free market. Patients are insured by their employers, or by Medicare and Medicaid. Insurers have their own networks of healthcare providers. So when patients change jobs, they change insurers and healthcare providers. Healthcare providers use different healthcare systems. Patient data, therefore gets generated and siloed all over the place.
As mentioned, many companies have been working on creating consolidated patient views. Reveleer works directly with 45 EMR providers to connect to data in larger health systems; they gather data through 70 health information exchanges (HIE) using trading and a sharing protocol around how information is shared to make it more publicly available. To create a consolidated patient view, they sweep a 20-mile radius around the patient to gather her data. “When you can capture so much data on a member or patient, the bigger challenge is, how do you then consolidate it? Distill it down to the most important issues that might be going on? We capture 1000 pages of medical record data, and need to surface the three or four most important things,” explained Jay Ackermann. The company has been doing the consolidation using machine learning and AI for over five years.
Komodo Health is gathering data for research, and has an overview of 300 million people (not all data points though). Similarly, Epic Cosmos built an environment to enable easier clinical research for contributing Epic customers. In October 2023, Epic Cosmos contained over 220 million patient records from over 9.3 billion encounters, representing patients in all 50 states.
Patient Risk Scoring
Companies are also developing risk-scoring algorithms to ease care management on an institutional level. At HLTH, Health Data Analytics Institute (HDAI) was presenting their work at Houston Methodist and Cleveland Clinic.
HDAI takes a statistical summary of all the historical CPT codes that are part of the patient's history up to that point in time. Based on this, they create predictions for patients. “I think a lot of people try to think of AI as magic. Our approach is to think of it as a statistical operational tool. We do not believe that any of the models out there are of diagnostic quality or should be used as diagnostic. They should be used as tools that summarize, and synthesize data and operationally get all the clinicians on the same page relative to where the patient's status is at. But ultimately, the decision of what test to run or to verify all the details is going to come down to clinical decision making,” explained Nassib Chamoun, Founder, President & CEO, Health Data Analytics Institute in our brief discussion on how the HDAI AI works.
Regulation
Nobody denies the need for regulation of AI. A different question entirely is how to do it effectively to ensure safety without killing the market. “We are not equipped today to collect the data out in the wild across systems such as US healthcare system. How to do post-market surveillance effectively we're gonna have to figure out together,” said Jesse M. Ehrenfeld, President of AMA at a panel about responsible AI.
The US Department of Health and Human Services (HHS) is still working on a guidance framework for AI, said Karl Mathias, CIO for the US Department of Health and Human Services.
John Halamka, President of the Mayo Clinic Platform, outlined several possibilities:
The FDA to have a predetermined change control plan, ensuring that use cases and data shifts are adeptly managed and subsequently revalidated on a specified quarterly or annual basis.
We should look for follow-up costs: how did an algorithm improve care?
The federal government should give guidance on what to measure, and there should be a nationwide network of assurance labs that take each commercial product through those specified subgroup analyses. “All that information is publicly disclosed in a nationwide registry. And this has to be updated at some interval basis because have shifted ecosystems and date,” said John Halamka, adding that this potential framework refers to predictive AI. “Generative AI is gonna be a little bit different animal, given that every output of every prompt can be different,” he added.
How are You Using Generative AI already?
“Not every startup is going to be a generative AI-focused company, but I would say almost every startup should be looking to incorporate generative AI, at least to a certain extent, within their business or their operations,” says Justin Norden. Every leader should be thinking about generative AI, even if it’s outside their business in the initial phase. “It's not enough to just hope someone on your team will figure out everything and deliver you the golden strategy. You need to start to incorporate this into your daily life, whether your personal life at home, cooking, searching information, you know, booking a flight to how can I start to use this in my, in my business?”
How to start?
Understand the Technology: Be curious and comprehend the capabilities and limitations of the current technology in your domain.
Address Customer Pain Points: Evaluate where customers are experiencing difficulties and explore how technological capabilities, such as summarization, audio transcription, and knowledge retrieval, can be integrated into existing solutions for both external and internal stakeholders.
Stay Updated: Ensure that either you or a designated team member consistently stays informed about the ongoing developments and trends in the relevant technological space.
Our as ChatGPT would answer the prompt: “If you had to give advice to a healthcare leader about generative AI, what would it be? Make it two sentences long.”:
Embrace generative AI with a well-informed and strategic approach, ensuring that its implementation aligns with ethical guidelines, regulatory compliance, and enhances patient-centered care. Invest in continuous learning, collaboration with AI experts, and prioritize scalable pilot projects to explore, validate, and safely integrate generative AI applications into healthcare practices, always keeping patient safety and data privacy at the forefront.
Piece of cake.
What Can Generative AI be Used for Apart From Generating Medical Notes?
June 2023 Deep dive into the landscape by teams at GSR Ventures and Maverick Capital.
In healthcare, generative AI has a plethora of potential applications:
1. Drug Discovery and Development: Molecule Design for potential new drugs, Predicting Drug Interactions.
2. Medical Imaging: generating synthetic medical images for training diagnostic models without compromising patient privacy identifying anomalies in medical images such as X-rays, MRIs, and CT scans.
3. Personalized Medicine: Generating personalized treatment plans based on a patient’s genetic makeup, lifestyle, and environment, creating synthetic genomic data to study rare diseases without accessing sensitive data.
4. Prosthetics and Bioprinting: Designing Prosthetics and organ bioprinting by generating 3D structures of organs.
5. Mental Health: Developing chatbots that can provide initial therapeutic conversations and support, predict patient’s mental state and providing timely interventions.
6. Health Management: Generating predictive models to anticipate healthcare trends and patient admissions, optimizing the allocation of resources such as hospital beds, staff, and equipment.
7. Synthetic Data Generation: Creating synthetic data to augment existing datasets, especially when real-world data is scarce or sensitive, generating synthetic datasets that preserve privacy while enabling research and development.
8. Medical Research: generate new hypotheses for scientific exploration based on existing data, assisting in designing research studies and clinical trials.
9. Virtual Health Assistants: Health Monitoring through generating health alerts and reminders for patients, creating personalized diet and exercise plans for individuals.
10. Rehabilitation: Custom exercise routines, using AI to generate reports on a patient’s recovery progress and adapting strategies accordingly.
11. Education and Training: Creating realistic simulations for training, generating synthetic case studies for educational purposes without utilizing real patient data.
12. Bioinformatics: Assisting in generating models for predicting the outcomes of gene-editing techniques like CRISPR, creating synthetic genetic sequences for research purposes.