From: redpointai
Oscar Health, a $3 billion public health insurance company, has been at the forefront of innovating in health technology for the past decade, actively experimenting with large language models (LLMs) like GPT-4 in various healthcare applications [00:00:20]. Mario Schaer, CTO and co-founder of Oscar Health, shared insights into the practical implementation of these models for real-world use cases, as well as a broader outlook on the future of healthcare and AI [00:00:44].
AI’s Impact in Healthcare
LLMs are particularly effective at transitioning between informal and formal language [00:01:40]. Healthcare, with its highly formalized codes (e.g., ICD-10, CPT, utilization management guidelines) and abundant human language (patient-provider conversations, electronic medical record notes), is uniquely positioned to benefit from this capability [00:02:02]. While healthcare has seen less algorithmic coverage compared to other industries, LLMs can now bridge this gap, enabling greater transparency and efficiency [00:02:26].
Administrative Use Cases
Initially, many AI use cases in healthcare will focus on administrative tasks [00:03:11]. The goal is to make processes like claims, authorizations, and cost transparency real-time and bidirectional [00:03:55].
> [!example] Examples > > * **Claims Explanation**: LLMs can translate complex claim rule traces into understandable language for laypersons, explaining why a claim was paid or denied [00:03:30]. > * **Call Summarization**: LLMs can summarize customer service calls, potentially phasing out manual note-taking by care guides [00:17:22]. > * **Lab Test Summarization**: Launched in Oscar's medical group [00:17:41]. > * **Secure Messaging Medical Records Generation**: Implemented in Oscar's Medical Group [00:17:46]. > * **Improving Growth and Retention**: Running outbound campaigns to remind members of Oscar's benefits and personalized care (e.g., reminding Asian-Americans about colorectal cancer screenings) [00:13:39]. Messaging can be tailored based on member personas, such as focusing on empathy for chronically ill members and convenience for generally healthy ones [00:14:52]. > * **Ethnicity Information Extraction**: LLMs can infer ethnicity from names or conversation language, which aids in matching members to appropriate doctors and conversations [00:16:33].
Oscar’s top four use cases include three administrative ones and one clinical [00:05:11]. The administrative applications are expected to yield tangible results more quickly [00:05:22], with savings estimated in “cents PMPM” (per member per month) [00:18:01].
A challenge in administrative tasks like claims processing is the need for the LLM to process information at a specific level of hierarchy, requiring careful prompting to avoid context window issues and to focus on relevant details [00:38:25].
Clinical Use Cases
The long-term goal is to replace caregivers and clinical intelligence with machine intelligence, ultimately reducing the cost of doctor visits and potentially replacing specialists with AI [00:04:30].
> [!example] Examples > > * **Medical Record Generation**: LLMs can summarize conversations between Oscar providers and members in [[challenges_in_virtual_healthcare_and_ai_doctors | virtual care settings]] to generate medical records [00:06:50]. > * **"Talking to the Medical Records"**: Enabling doctors, customer service agents, or anyone within Oscar to query medical records using natural language [00:18:52].
> [!warning] Challenges in Clinical AI > > * **Contextual Knowledge Gaps**: LLMs often lack subtle contextual knowledge that human providers possess (e.g., remembering a previous conversation not in the medical record), leading to less accurate summaries [00:07:11]. Improving LLM performance in clinical settings requires not just better models but also expanding their "horizon of knowledge" through comprehensive input data [00:08:02]. > * **False Positives**: For complex concepts like "post-traumatic injury" in utilization management, LLMs can generate a high rate of false positives because their training data includes broader, layperson associations of the term, differing from its precise medical definition [00:31:49]. Strategies like self-consistency questionnaires, providing short examples, and pre-processing steps where the LLM generates its own knowledge about how the condition might appear in records can improve accuracy [00:33:41]. > * **Safety**: Ensuring patient safety is paramount, making it difficult for LLMs to directly interact with end-users in clinical contexts due to the risk of hallucinations or insulting outputs [00:50:50]. Current solutions often involve a human in the loop [00:51:00]. > * **Physical Interaction**: A significant portion of healthcare (around one-third of claims data) still requires in-person interaction, leading to "leakage" from [[challenges_in_virtual_healthcare_and_ai_doctors | virtual care]] to in-person visits and reducing patient loyalty to a single primary care provider [00:57:52]. > * **Business Model Incentives**: Large health systems currently have little financial incentive to switch to lower-cost [[cost_efficiency_and_accessibility_in_ai_technologies | channels of care delivery]] like automated virtual primary care, as it could lead to reduced reimbursement from insurers and the government [00:59:50].
Regulation and Implementation
Implementing AI in healthcare systems is subject to strict regulations, primarily HIPAA, which mandates business associate agreements (BAAs) to ensure patient data privacy [00:20:56]. Oscar Health was notably the first organization to sign a BAA directly with OpenAI [00:21:18].
New LLM models from providers like Google (Gemini Ultra) or OpenAI typically require a waiting period (e.g., 3-4 months) before they are officially covered under existing HIPAA agreements [00:23:04]. During this period, testing is done with synthetic or anonymized data [00:22:50].
For companies looking to sell AI solutions to hospitals, formal requirements include security and policy reviews, often involving long checklists, and certifications like HiTrust [00:23:57]. However, gaining trust from hospitals is more critical than merely passing certifications [00:24:42]. The healthcare industry is generally slow at rapid prototyping and emphasizes enterprise sales processes over product perfection [00:25:01].
Mario Schaer participated in a consortium of health systems and insurance companies that drafted principles for AI in healthcare, advocating for the democratization of analytics through AI [00:26:03].
Structuring AI Teams
Oscar Health has a “Pod” structure for its AI team, comprising seven people, including product managers, data scientists, and engineers [00:47:10]. This team holds weekly office hours and hacking sessions where anyone in the company can bring their AI ideas and discuss prompting strategies [00:47:22]. The Pod also has its own three priorities to ensure project completion and avoid endless research [00:47:34]. This structure balances centralized guidance with decentralized experimentation, fostering an environment where trying new things is encouraged [00:50:10].
Overhyped, Underhyped, and Future Opportunities
- Overhyped: Clinical chatbots, generally speaking [01:00:49].
- Underhyped: Voice outputs [01:00:56].
Mario Schaer is most excited about OpenAI for its continued model development and Hugging Face for centralizing models [01:01:13].
Regarding commercial opportunities, he suggests focusing on very obscure niches in healthcare, such as regulatory filings composition [00:53:22]. This includes generating regulatory documentation for health regulators, SEC compliance (Sox), NCQA clinical reports, and state health department reports [00:54:02]. He views prior authorization companies as risky, as it’s a core competency of insurance companies [00:55:59]. Another promising area is fraud, waste, and abuse detection, an industry currently dominated by outdated players [00:56:16].
> [!info] Learn More > > To learn more about Oscar Health's AI work and insights, visit [Hi.Oscar.Health](https://hi.oscar.health) [01:01:50]. Mario Schaer also shares his explorations on Twitter: @MarioTS [01:02:07]. > > Mario also shared some fun, non-healthcare related AI ideas, including building an RPG game based on a company's internal documents and an Oregon Trail-like game where an LLM dynamically writes game mechanics as it's played [01:02:32].