How UTMB improved diagnostic for thousands of patients using AI

Dr. Peter McCaffrey and James Weatherhead from the University of Texas Medical Branch discuss leveraging AI to enhance healthcare operations, focusing on three key projects. They highlight their use of AI in pathology for efficient diagnosis, pharmacogenetics for personalized medicine, and a legal document analysis tool to identify contradictions. Their approach integrates tools like llama index and retool, emphasizing rapid prototyping and human-in-the-loop systems to ensure accuracy and compliance. They stress the importance of AI in transforming unstructured data into actionable insights, aiming to streamline healthcare processes and improve patient outcomes.

• UTMB is a large academic medical center with multiple campuses, providing a wide range of healthcare services.
• The center aims to improve healthcare efficiency and operations through AI integration.
• AI is used to enhance patient diagnosis, streamline billing workflows, and improve operational efficiency.

"We are kind of your classic large academic Medical Center... we have a lot of operations of a lot of complexity we think AI could help us broadly."

• UTMB's complexity and scale necessitate innovative AI solutions to handle diverse healthcare operations efficiently.

• The Pathology Department at UTMB is a major hub for AI application, focusing on diagnostics.
• AI serves as a co-pilot in pathology, significantly reducing the time required for interpretations.

"We started applying in production AI as a co-pilot to do interpretation in the domain of pathology."

• AI's role as a co-pilot enhances efficiency, allowing pathologists to focus on more cases in less time.

• Quick prototyping with AI enables rapid development and deployment of applications, reducing project timelines from months to days.

"Quick prototyping approach turns some in that could have been easily many months to a year of a one-off project into a few days to bring an application to life."

• The ability to quickly prototype and test AI applications accelerates innovation and improves workflow efficiency.

• The Auto Talks application automates the toxicology workflow for pathologists.
• AI gathers necessary information and drafts interpretations, allowing pathologists to review rather than create from scratch.

"I as a pathologist see a list of cases I need to interpret... AI has already written yes it's positive for tramol and its metabolite."

• AI's capability to draft interpretations streamlines the workflow, shifting the focus to review and validation.

• The application allows pathologists to provide feedback on AI-generated interpretations, facilitating continuous improvement.

"We like this because a this changes my workflow from writing to seeing everything in one pain like data here data here and the draft."

• Feedback mechanisms ensure the AI model evolves and improves based on real-world usage and expert input.

• Feedback from pathologists is used to manually refine AI prompts, enhancing accuracy and completeness.

"We just manually take that into account and say okay when we write the next prompt we're going to call out you know you should include mention of this."

• The iterative refinement process ensures the AI model aligns closely with clinical needs and standards.

• The AI model achieves high accuracy in drafting interpretations, significantly reducing the need for manual adjustments.

"It's probably 96% of the time you don't need to touch anything about it and by the way I mean like that's very good."

• High accuracy rates demonstrate the effectiveness of iterative improvements and the model's alignment with clinical requirements.

• The pharmacogenetics dashboard provides insights into patients' genetic profiles and drug metabolism capabilities.
• The dashboard maps patients' genotypes to their ability to metabolize drugs, offering personalized medication recommendations.

"We can look at okay let's look at the genes let's look at the variance of the genes and let's look at where the variant is present in the assessment."

• Mapping genetic variants to drug metabolism helps personalize treatment plans, improving patient outcomes.

• The dashboard includes recommendations for drug dosages based on genetic profiles, ensuring safer and more effective treatments.

"This person has this genotype and this Gene and for this drug they are a poor metabolizer so the recommended start for this patient would be perhaps a lower dose."

• Personalized dosage recommendations based on genetic profiles enhance medication safety and efficacy.

• The integration of pharmacogenetic data into clinical workflows exemplifies the potential of AI to personalize and improve healthcare delivery.

• Pharmacogenomics is a new field focusing on the genetic factors influencing drug metabolism.
• AI streamlines the interpretation of genetic data, enhancing efficiency in healthcare.
• The integration of LLMs (Large Language Models) and APIs automates the retrieval and analysis of pharmacogenetic information.

"It's quite a new field and, you know, with new fields, there's consistent data and literature upgrades and publications."

• The quote highlights the dynamic nature of pharmacogenomics and the need for constant data updates.

"This just streamlines it and, yeah, it turns something that would take many hours, if not weeks, manually into less than five minutes."

• AI significantly reduces the time required for data analysis, improving efficiency in healthcare processes.

• The integration of new data types presents challenges, especially in fields not well-represented in training sets.
• Careful selection of data sources is crucial to ensure scientific rigor and accuracy in healthcare recommendations.
• The use of respected datasets like HGVS and PharmGKB ensures reliability in genetic interpretation.

"We have to give it to physicians at our school or at our institution and say, is the scientific rigor of this information enough to be giving patients feedback on this?"

• The quote underscores the importance of validating data sources to maintain scientific integrity.

"Luckily, we haven't been using vast data sets; we've been using HGVS and then we've also been using one for the interpretation which are very well respected within the medic field."

• Emphasizes the reliance on established and respected datasets to ensure accuracy in genetic analysis.

• Human oversight is essential in AI-driven workflows to ensure accuracy and reliability.
• Physicians review AI-generated interpretations before they are integrated into patient records.
• Feedback loops allow continuous improvement of AI models through prompt engineering.

"The draft interpretation doesn't get pushed to Epic unless a physician has read it and gone through all of it."

• Human review is critical to maintaining the quality and reliability of AI-generated data.

"We can apply analytics to that and just improve our prompt engineering as we go along."

• Continuous improvement through feedback and analytics enhances the effectiveness of AI tools.

• AI tools assist in identifying contradictions in legal documents, improving efficiency in legal departments.
• The use of LLMs and vectorization enables precise retrieval of information from large document sets.
• Ranking systems help prioritize contradictions, facilitating efficient resolution.

"What this does is it compares contradicting legal statements among 1,52 legal statements."

• AI tools streamline the identification of contradictions in extensive legal document collections.

"What AI, what this LLM is allowing us to do with llama index and embedding and vectorization of this content in higher dimensional space, it allows us to find the contextual semantic meaning of our queries."

• Advanced AI techniques enable precise and contextually relevant information retrieval.

• Larger embedding models with more dimensions reduce hallucinations and enhance precision.
• Embedding models with more granular dimensions improve the specificity of AI queries.
• The choice of embedding dimensions is crucial in balancing broad concept capture and detailed query precision.

"The larger embedding models seem to do a lot better for reducing hallucinations and genericness of the outputs."

• Larger, more detailed embedding models enhance the accuracy and reliability of AI outputs.

• AI tools significantly improve time efficiency and accuracy in legal document analysis.
• Legal departments have been receptive to AI tools, recognizing their potential as adjunct tools.
• AI tools enhance job quality and efficiency without replacing human roles.

"I think these tools are showing us that it's not going to replace us as quickly as everyone thinks it might."

• AI tools are viewed as enhancing, rather than replacing, human roles in legal departments.

• Retool facilitates the integration of various AI tools, improving usability and accessibility.
• A unified platform enhances user experience and compliance with institutional requirements.
• Retool provides a user-friendly interface for deploying AI applications within institutions.

"Retool just allows you to bring it all together and use it within one platform and then also allow people within the institution to use these apps."

• Retool simplifies the deployment and use of AI applications, enhancing accessibility and compliance.

"Retool has come in handy with that, also just allowing us to do HPoc compliant things by keeping it on our own servers."

• Retool's compliance features make it suitable for institutional use, ensuring data security and integrity.

• The process involves manually retrieving approximately 1,052 papers and processing them through Llama Index for machine compatibility.
• Llama Index helps in chunking documents, allowing retrieval of specific text chunks in response to user queries.
• Retrieval-Augmented Generation (RAG) involves retrieving and augmenting generation by embedding text into higher-dimensional space.

"We manually have to go out and retrieve about 1,052 papers. Once we have that, we process it through Llama Index. It's just a way to index and sort these PDFs so that they are machine-friendly and machine-compatible."

• This quote explains the initial step of retrieving and processing documents to make them usable for machine learning applications.

"In higher dimensional latent space, cat and dog will be close, but train will be further away."

• The quote illustrates how similar concepts are grouped together in latent space for efficient retrieval and processing.

• Large language models (LLMs) are effective for natural language processing but not for tasks easily done by Python scripts.
• Medicine can benefit from LLMs by structuring unstructured clinical notes, aiding in billing and patient care.
• LLMs can help convert progress notes into structured ICD codes for efficient billing in healthcare.

"If it can be done in a Python script, I wouldn't bother asking LLM to try and accomplish that."

• This quote emphasizes the limitations of LLMs for tasks that can be easily automated by simpler programming tools.

"We are taking unstructured clinical notes and making them structured and tabularized for machine learning purposes."

• The quote highlights the application of LLMs in transforming medical data for better usability and efficiency.

• OpenAI's structured output mode is useful, but a multi-step LLM process is employed for reliable structured output.
• Multi-architectural systems of LLMs perform tasks like named entity recognition and structured data generation.
• Validation loops and Python scripts are used for refining outputs, moving beyond simple one-hot prompts.

"We are trying to build a multi-step LLM process of almost Chain of Thought reasoning."

• This quote describes the complex process of using multiple LLMs in sequence to achieve structured data output.

"The idea of having a one-hot prompt is just, it's stuff of a legend."

• The quote underscores the impracticality of relying on single prompts for complex data processing tasks.

• Larger context windows and improved reasoning capabilities in LLMs are anticipated advancements.
• Integration of electronic health records (EHR) with multi-agent systems could revolutionize healthcare.
• Autonomous labs like Emerald Cloud Lab could enable dynamic clinical trials and personalized medicine.

"I'm excited for larger context windows; it's going to be cool to have that."

• This quote reflects the anticipation for advancements in AI that allow handling of larger datasets and more complex queries.

"Medicine is going to become extremely personalized with AI agents and robotics, and that connection is going to be very beautiful."

• This quote envisions a future where AI and robotics significantly enhance personalized healthcare.

• The focus is on enabling rapid experimentation and prototyping for quick deployment and testing of AI models.
• Retool aims to be the application layer for AI, facilitating seamless model upgrades and task performance evaluation.

"What we're really focused on is enabling this really rapid experimentation for folks to kind of prototype and get things going quickly."

• This quote highlights the commitment to providing tools and platforms that support fast and efficient AI development.

• Initial experiments involved using fake data and Azure HPP compliant GPT models to maintain HIPAA compliance.
• Importance of data de-identification when using real patient data to ensure privacy and compliance.
• Use of Azure to self-host and ensure data remains in-house, maintaining compliance and control.

"A lot of what we do is with either fake data and also we use fake data majority of the for the retour as well like in the beginning we did a lot of fake patient data and then we also have the Azure HPP compliant gpts on our Microsoft Azure account here."

• The use of fake data and Azure HPP compliant GPTs ensures compliance with HIPAA regulations while allowing for experimentation.

"We do do a deidentification of patient data yes if we're using real patients."

• De-identification of patient data is crucial for compliance and privacy when handling real patient data.

• Retool is used extensively for UI building, data storage, input-output management, and data analysis.
• Retool's ease of use and learning curve makes it a preferred tool for fast experimentation and application development.
• Retool enables non-developers like lawyers and physicians to interact with AI systems effectively.

"We we're using retool for for all of what you just said, and also to store input outputs and then you know do data analysis on that later on."

• Retool's versatility allows it to be used for multiple purposes, including data analysis and storage.

"Retool in and of itself is extremely easy to learn and use."

• Retool's user-friendly nature allows for quick adoption and implementation without extensive training.

• A contradiction tool is used to rank contradictions from zero to ten, with ten being catastrophic for legal compliance.
• The legal department collaborates to define criteria for ranking contradictions.
• The ranking system helps prioritize issues that need urgent attention from a legal perspective.

"That rank in the right hand corner is essentially once we've discovered contradictions, we feed those contradictions back to an llm and we say Okay based upon all of these contradictions rank which ones would be from zero to 10."

• The ranking system helps in assessing the severity of contradictions for legal compliance.

"The specific criteria was worked out with the legal team."

• Collaboration with the legal team ensures that the ranking criteria align with legal requirements.

• Retool Vectors is used to host a database and embed documents for query purposes.
• The process involves indexing PDFs, vectorizing them, and embedding them into the model for efficient querying.
• Retool facilitates the storage and retrieval of data, making it accessible for users without technical expertise.

"We have all these PDFs we process them through llama index once they've been in a indexed format then we vectorize and embed those into the model."

• The process of indexing and embedding data allows for efficient querying and retrieval.

"Retool just kind of facilitates the storing of everything."

• Retool's role in data storage simplifies the process for end-users, making it accessible and user-friendly.

• The development team consists of a small group, emphasizing the efficiency of Retool in enabling rapid development.
• Collaboration with physicians and lawyers for validation and refinement of outputs.
• Retool's user-friendly platform reduces the need for a large development team.

"Me and mcaffrey made the auto talks in five weeks together."

• A small team can efficiently develop complex applications using Retool.

"Retool is a very very user friendly platform."

• Retool's ease of use reduces the need for extensive technical expertise in the development process.

• Various tools are mentioned for different purposes: Llama Index for data indexing, OpenAI embedding models for embedding, Pinecone or Chroma DB for vector databases.
• Experimentation with different language models is encouraged to find the best fit for specific tasks.
• Retool's flexibility allows for integration with multiple tools and technologies.

"I would say llama index for indexing your your data, then I would use open AI embedding models."

• The use of specialized tools for data indexing and embedding enhances the efficiency of data handling.

"Experimenting with them is something we've really tried to make pretty simple for that reason."

• Experimentation with different models allows for optimization and fine-tuning of applications.

• Retool supports AI-driven data retrieval from tables and documents.
• The ability to query both structured and unstructured data enhances the flexibility and power of applications.
• Retool's integration capabilities allow for the incorporation of various data sources into applications.

"You can use queries inside of your retool app to just pull in that data and use it if you just want to query the structured data."

• Retool's querying capabilities allow for the integration of structured data into applications.

"The fact that retool can bring you know structured data in a database or unstructured data in a vector database like any kind of data you have together."

• Retool's ability to handle both structured and unstructured data makes it a versatile tool for application development.