From: redpointai
AI, particularly Large Language Models (LLMs), is poised to significantly impact healthcare by enhancing communication between patients and providers [01:36:19]. This advancement is crucial given healthcare’s unique blend of highly formal and deeply human language [02:14:02].
Core Capabilities of LLMs
LLMs excel at transforming informal language into formal language and vice versa [01:40:24]. This capability is uniquely suited for healthcare, where highly structured codes (e.g., ICD-10, CPT) and regulations coexist with complex human conversations between patients and providers [02:02:10].
Administrative Use Cases
Initially, many AI applications in healthcare will focus on administrative tasks [03:11:04]. Oscar Health, a public health insurance company, has prioritized several administrative use cases to improve efficiency and transparency [03:09:07]:
- Claims Explainers AI can translate the complex, formal “trace” of how a claim was processed—detailing which rules were applied and why—into understandable, informal language for a layperson [03:37:37]. This aims to make processes like claim denials much clearer to members [03:47:04].
- Call Summarization LLMs are increasingly used to summarize customer service calls, phasing out manual note-taking by care guides [07:19:57].
- Medical Record Generation AI can generate medical records from secure messaging conversations [07:46:04].
- Lab Test Summarization LLMs summarize lab test results for medical group staff [07:41:43].
These administrative uses contribute to making the healthcare system more transparent, allowing patients to understand costs and alternatives in real-time [04:08:24].
Clinical Use Cases and Challenges
While administrative applications are more straightforward, AI also aims to enhance communication in clinical settings.
Summarizing Clinical Conversations
LLMs can summarize conversations between providers and patients (e.g., in virtual primary care settings) and generate medical record notes from them [07:02:05]. This allows for adapting communication based on the audience, whether it’s doctor-to-doctor or doctor-to-patient, by transforming the same data into different information levels [06:05:05].
Limitations and Nuance
A key challenge for LLMs in clinical contexts is capturing subtle contextual knowledge that human providers possess but may not explicitly record [07:17:21]. This includes remembering previous unrecorded conversations with a patient or understanding local geographical context (e.g., weather conditions impacting travel to a physician) [09:06:08]. This creates an “unfair playing field” for LLMs when inputs and outputs are less structured [07:38:15].
For instance, an LLM might struggle with highly specific medical definitions compared to common layman’s understanding, leading to false positives (e.g., incorrectly identifying “post-traumatic injury” in a specific clinical context) [03:49:50]. To address this, strategies like “self-consistency questionnaires” or “chain of thought” prompting are used, where the LLM is guided to generate and evaluate multiple perspectives or break down complex tasks into independent steps [03:52:13].
Regulatory and Trust Requirements
Implementing AI in healthcare systems is subject to strict regulations, most notably HIPAA, which prohibits sharing patient-specific information without proper agreements [02:08:59]. AI providers must sign Business Associate Agreements (BAAs) to handle protected health information [02:05:32]. New models from providers like Google or OpenAI are not immediately covered by these agreements, requiring the use of synthetic or anonymized test data for initial evaluations [02:51:30].
Beyond compliance, gaining the trust of hospitals and healthcare systems is paramount [02:44:03]. This often involves lengthy security and policy reviews, and building relationships through collaboration rather than just product development [02:56:16].
The Future of AI in Human Communication
AI Doctors and Clinical Automation
The long-term goal is to replace caregivers and clinical intelligence with machine intelligence, potentially reducing the cost of doctor visits significantly [04:30:11]. Medicine is highly algorithmic, making it theoretically suitable for AI to map existing knowledge and infer based on data points [00:57:15].
Challenges in Virtual Care
Despite the potential for AI in clinical settings, several challenges remain for virtual healthcare and AI doctors [00:57:36]:
- Safety: Ensuring AI outputs are safe and don’t hallucinate or contain biases is critical for direct patient interaction [04:50:09]. Currently, human oversight (human-in-the-loop) is necessary for sensitive use cases [05:01:02].
- Physical Interaction: Many medical needs still require in-person physical interaction, such as lab tests or foot exams for diabetics [05:57:52]. This “leakage” to in-person care can disrupt the continuity of virtual AI-driven care [05:59:36].
- Business Model: Current healthcare system incentives do not always align with adopting lower-cost virtual care channels, as it can lead to reduced reimbursement and capacity [05:59:50].
Overhyped vs. Underhyped
- Overhyped: Clinical chatbots generally are considered overhyped at present [01:00:49].
- Underhyped: Voice outputs are seen as underhyped, offering significant potential for communication, as long as they are not used for clinical advice [01:00:56].
Ultimately, general-purpose LLMs like GPT-4 are often preferred over specialized healthcare models because they maintain better alignment and instruction-following capabilities, even if specialized models might seem contextually more appropriate [00:43:57]. Combining approaches like RAG (Retrieval Augmented Generation) and fine-tuning can independently improve performance [00:45:11].