You have 4 summaries left

The Gradient: Perspectives on AI

Vivek Natarajan: Towards Biomedical AI

Thu Jun 06 2024
AI in MedicineHealthcare AccessibilityEvaluation MetricsMedical Decision-MakingGenerative AILLMs for Genetic DiscoveryLanguage Models in BiomedicineAI Systems for Patient Care JourneysAssistive Effect of LLMs in Complex CasesRefining Medical AI Systems


AI systems in medicine have the potential to expand clinical knowledge and democratize medical expertise. However, challenges such as healthcare accessibility, evaluation metrics, and refining AI systems need to be addressed. This podcast explores the advancements, challenges, and future vision of AI in healthcare.


Democratizing Medical Expertise

AI systems can expand clinical knowledge and help those with limited access to doctors. The concept of an AI doctor aims to address the lack of healthcare accessibility. Advancements in technology pave the way for democratizing medical expertise.

Enhancing Doctors' Abilities with AI

AI can enhance doctors' abilities in providing care, making healthcare more efficient. Specialization in AI models and collaboration between clinicians and AI experts are crucial.

Evaluating Medical AI Systems

Metrics for evaluating medical AI systems should encompass clinical expertise, technology research, and human aspects. Choosing appropriate evaluation metrics is essential for accurate assessment.

Challenges in Medical Decision-Making

Medical decision-making faces challenges such as resource optimization and contextual knowledge. Clear communication and considering liability are important when using AI-influenced systems.

Advancements in Generative AI

Generative AI like CHAD GPT is transforming healthcare, but proper verification and validation are crucial. Balancing optimism with rigor is necessary for the future of healthcare.

LLMs for Genetic Discovery

LLMs are used for genetic discovery and hypothesis generation. Human expertise is crucial in verifying AI-assisted discoveries.

Tweaking Language Models for Biomedicine

Careful tweaking and injection of domain knowledge enhance language models in biomedicine. Text-based medical consultations using AI have the potential to democratize access to care.

AI Systems for Patient Care Journeys

AI systems show promise in acquiring information under uncertainty for patient care journeys. Simulation environments play a crucial role in training AI models.

Assistive Effect of LLMs in Complex Cases

LLMs have a significant assistive effect in complex medical cases, providing broad possibilities for diagnoses. They complement human experts by offering a more unified view of differentials.

Refining Medical AI Systems

Refining AI systems involves specialization, feedback from experts, and increasing sample size. Building trust with clinicians and accurately conveying uncertainty are important.


  1. Democratizing Medical Expertise
  2. Challenges in Healthcare Accessibility
  3. Enhancing Doctors' Abilities with AI
  4. Evaluating Medical AI Systems
  5. Diverse Reactions and Perceptions
  6. Challenges in Medical Decision-Making
  7. Palm Free Training Corpus
  8. Advancements in Generative AI
  9. LLMs for Genetic Discovery
  10. Tweaking Language Models for Biomedicine
  11. AI Systems for Patient Care Journeys
  12. Simulating Doctor-Patient Dialogues
  13. Assistive Effect of LLMs in Complex Cases
  14. Refining Medical AI Systems
  15. Communication and Collaboration with AI Systems
  16. Advancing AI Systems in Healthcare
  17. The Vision of Advanced AI Systems in Healthcare

Democratizing Medical Expertise

00:00 - 07:32

  • AI systems in medicine can expand clinical knowledge and help those with limited access to doctors.
  • The concept of an AI doctor is no longer science fiction and aims to address the lack of healthcare accessibility in developing parts of the world.
  • Advancements in technology, such as smartphones and AI models like transformers, have paved the way for democratizing medical expertise and other fields.
  • Personalized medicine is crucial as each individual's biology is unique, highlighting the need to move away from generalized treatments.
  • Democratizing medical expertise is still in its early stages but holds immense potential for optimizing healthcare delivery.

Challenges in Healthcare Accessibility

07:04 - 14:30

  • Democratizing legal expertise parallels the need for democratizing medical expertise, especially for those without access to doctors.
  • Access to medical expertise is not uniform even in advanced countries like the United States, with long wait times for appointments being a common issue.
  • The medical industry has shifted focus towards financialization rather than patient care, straying from the original goal of medicine which was caring for others.
  • AI in medicine aims to amplify doctors' expertise and experience, making healthcare more efficient rather than replacing physicians entirely.
  • AI advancements can improve diagnosis and disease trajectory prediction, allowing doctors to focus on aspects of care that AI cannot replace, such as establishing human connections with patients.

Enhancing Doctors' Abilities with AI

14:03 - 21:14

  • AI can enhance doctors' abilities in providing care, leading to a more customer-centric approach to patient care.
  • Specialization in AI models for healthcare is crucial to ensure safety and reliability in clinical workflows.
  • Rigorous verification and validation studies are essential before deploying AI systems in medicine, following well-established processes similar to drug discovery.
  • Collaboration between clinicians and AI experts is important to align AI models with medical values and practices.

Evaluating Medical AI Systems

20:54 - 28:06

  • Metrics for evaluating medical AI systems should encompass clinical expertise, technology research, and human aspects like empathy and communication.
  • Baseline models are essential to assess system performance before optimization.
  • Metrics used in research may differ from those indicating real-world clinical applicability.
  • Choosing evaluation metrics depends on the intended use of the AI system, whether for laypersons or clinicians.
  • Evaluators of AI systems should match the expertise required for accurate assessment.

Diverse Reactions and Perceptions

27:36 - 34:42

  • The importance of evaluating AI systems based on diverse populations' reactions and perceptions
  • Using established evaluation rubrics from medical education boards to assess AI systems in the clinical domain
  • Considering alignment with scientific consensus when developing AI systems for long-term success
  • Incorporating safety measures, scalable oversight mechanisms, and expert judgment into AI systems to ensure proper use and patient care

Challenges in Medical Decision-Making

34:13 - 41:29

  • In healthcare, it may be necessary to deviate from guidelines to provide the best care for some patients, but measuring outcomes accurately is challenging.
  • Building hardened simulation environments can help in understanding real-world clinical settings and human body interactions.
  • Clinical data often lacks diversity in demographic backgrounds, leading to uncertainties in treatment effectiveness for different individuals.
  • Resource optimization is a crucial factor in medical decision-making that AI systems need to consider.
  • Contextual knowledge and situational awareness are essential for AI systems operating in healthcare settings.
  • Clear communication of information and liability considerations are important when using AI-influenced systems for health optimization.

Palm Free Training Corpus

41:01 - 48:15

  • The team behind the Palm Free Training Corpus wanted to ensure that people understand that not all medical knowledge in their model comes from a specific phase, but rather from large-scale internet pre-tuning on various sources.
  • The focus of the team's work was on accurately quantifying the overlap between evaluation sets and pre-training data, rather than identifying specific medical tokens.
  • The scientist emphasizes the importance of making grounded claims in healthcare AI to build trust with clinicians and the broader medical community.
  • Rigorous science and verification studies are crucial for advancing AI systems in healthcare and gaining trust from stakeholders.
  • There is a need to bring everyone along in the journey towards advancing AI systems in healthcare, including clinicians, policymakers, social scientists, and patients.

Advancements in Generative AI

47:46 - 55:05

  • Healthcare is facing challenges related to accessibility, availability, quality consistency, and price.
  • Advancements in generative AI like CHAD GPT are transforming healthcare by integrating AI into clinical workflows.
  • Rapid integration of technology in healthcare without proper verification and validation could lead to negative outcomes.
  • Balancing optimism with rigor in applying AI systems is crucial for the future of healthcare.
  • LLMs are being used for genetic discovery and hypothesis generation in healthcare research.

LLMs for Genetic Discovery

54:37 - 1:01:55

  • LLMs are trained on broad internet corpora and paired with retrieval augmented systems to generate interesting hypotheses in genomics.
  • The focus is on working at the gene token level rather than the base pair level for genetic discovery.
  • Human expertise is crucial in verifying hypotheses generated by AI models efficiently.
  • AI-assisted discoveries in genetics are being verified, potentially leading to novel findings.
  • The approach of using LLMs with human-in-the-loop AI shows promise in various areas beyond genomics, such as neurodegenerative diseases.

Tweaking Language Models for Biomedicine

1:01:37 - 1:09:07

  • Incorrectly assigning gene symbols can lead to false associations in research.
  • Careful tweaking and injection of domain knowledge can enhance the current generation of language models in accelerating progress in biomedicine and science.
  • Conversational diagnostic AI involves engaging in multi-turn conversations with patients to gather information under uncertainty, a different skill set compared to traditional evaluations.
  • Objective Structured Clinical Examination (OSCE) framework is used for evaluating conversational diagnostic AI, but there are limitations due to the absence of visual cues in text-based consultations.
  • Text-based medical consultations using AI could be a new modality that democratizes access to care, although human doctors may not be trained for this type of interaction.
  • AI systems can maintain consistent tone and detail in communication, unlike human doctors who may face limitations due to fatigue.

AI Systems for Patient Care Journeys

1:08:42 - 1:16:01

  • AI systems are showing promising results in acquiring information under uncertainty to assist in patient care journeys
  • Challenges in medical AI include data availability and sample sizes
  • A multi-agent framework was used to simulate doctor-patient dialogues for model refinement
  • Simulation environments play a crucial role in training AI models for language-based diagnostics
  • Analyzing conversational dispositions is important for handling various patient behaviors in diagnostic agent interactions

Simulating Doctor-Patient Dialogues

1:15:34 - 1:23:34

  • Simulation is crucial for evaluating different behaviors in models, ensuring they stick to guidelines and guardrails.
  • Chaining sequential model calls in online inference may lead to a loss of information, but efforts were made to retain context and prevent hallucinations.
  • Primary care physicians and AI systems performed worse in obstetrics/gynecology and internal medicine scenarios due to the need for in-person consultations for accurate diagnosis.
  • Neurology was a more specialized area where the model did well, while respiratory cases were common and handled effectively by both PCPs and AI systems.

Assistive Effect of LLMs in Complex Cases

1:23:04 - 1:30:25

  • The study aimed to test the hypothesis that AI systems like LLMs could be beneficial in assisting with complex medical cases.
  • LLMs were found to have a significant assistive effect compared to clinicians and tools like search engines in solving complex cases.
  • LLMs are able to integrate information better and come up with broad possibilities for diagnoses, especially in rare scenarios.
  • Doctors tend to be anchored to common priors, while LLMs can provide a more unified view of differentials for diagnoses.
  • The lack of narrow priors in LLMs may contribute to their complementarity benefits when working alongside human experts.

Refining Medical AI Systems

1:29:56 - 1:37:35

  • Specialization in AI models can be reinforced through data curation, instruction fine-tuning, post-training, and retrieval augmented generation.
  • Feedback from experts is crucial for refining AI systems, especially in cases where expert opinions may differ.
  • Increasing sample size and using adjudication panels help address discrepancies in expert opinions when training medical AI systems.
  • Ground truth and outcome data are essential challenges in developing reliable medical AI systems.
  • Models like LLMs should accurately convey uncertainty and provide sources to expand the range of considerations for doctors.
  • Building trust with clinicians regarding AI outputs is crucial to avoid blind reliance on model predictions.

Communication and Collaboration with AI Systems

1:37:05 - 1:44:18

  • Accurate communication of AI model uncertainties is crucial for user trust and understanding.
  • Designing optimal user interfaces for AI-human collaborations is a key area of innovation.
  • Clinicians need to be trained to not blindly trust AI model outputs and understand their limitations.
  • Pilot phases are important for onboarding clinicians onto AI tools to measure trust and understanding.
  • Incorporating expert feedback and qualitative information is essential for refining and improving AI systems.
  • Balancing the UI design between being helpful without causing blind trust in AI recommendations is a significant challenge.

Advancing AI Systems in Healthcare

1:43:51 - 1:51:17

  • Models in AI tend to draw conclusions from isolated systems, but there is a push to view complex cases more holistically for better outcomes.
  • Base properties of models can lead to limitations and harmful consequences if not addressed with technical and algorithmic solutions.
  • Compound AI systems with built-in uncertainty mechanisms and self-refinement can help improve model performance.
  • Crafting prompts for AI systems may enhance performance, but algorithmic approaches like directed graph evolution are seen as more scalable and promising.
  • The goal is to have AI systems amplify and augment doctors' knowledge and expertise in healthcare settings rather than replacing them.
  • In the future, AI systems are envisioned to be people-facing, providing continuous interaction over months and years to optimize health and well-being.

The Vision of Advanced AI Systems in Healthcare

1:50:47 - 1:55:03

  • The vision is to democratize medical expertise and make it instantly accessible to everyone through AI technology.
  • Continuous interaction with healthcare, health coaches, and fitness courses is envisioned over long time horizons.
  • Efforts are being made to ensure that the benefits of advanced AI systems in healthcare are available to everyone, not just a select few.
  • Advancements in AI systems could lead to eradicating human diseases and designing individualized therapies.
  • Hybrid physician scientist systems combining AI technology with human expertise aim to accelerate biomedical knowledge and push frontiers in medicine.