Improving patient care with large-scale data engineering and ML (PRO207)
Improving Patient Care with Large-Scale Data Engineering and Machine Learning
Introduction
The presentation focused on how to improve patient care using large-scale data engineering and machine learning techniques.
The presenters, Joe Langreo and Jenna Eun, are part of the AWS Professional Services team.
The presentation covered four main topics:
Why is AI challenging in healthcare today?
Using traditional methods, including traditional AI/ML approaches, and the common use cases and challenges faced by customers.
How does GenAI and transformer models fit into this, and how do they help address those common use cases?
Improvements and learnings from a project the presenters worked on.
Challenges of AI/ML in Healthcare
Over 80% of healthcare data is trapped in unstructured freeform medical texts, such as progress notes, medical notes, and visit summaries.
This unstructured data is often multimodal and contextual, making it difficult to extract and transform into structured data for use in EHR systems.
The process of extracting this data is time-consuming and labor-intensive, which can be challenging for smaller medical facilities or hospitals that may not have the necessary staff or expertise.
Common Healthcare Use Cases
Patient Care Coordination: For example, in organ transplant and cancer treatment planning, important information is trapped in unstructured data, making it difficult to create a comprehensive plan for the patient.
Patient Transfer between Healthcare Organizations: When transferring a patient from one facility to another, the data required for the transfer is often trapped in different EHR systems, making the process time-consuming.
Population Healthcare Analysis with Government Departments: Analyzing healthcare data at a large scale, such as tens of thousands or hundreds of thousands of patients, is challenging due to the unstructured nature of the data.
Traditional Approaches and Their Limitations
Manual Approach:
Collect document images from EHR systems.
Curate the data manually, requiring medical expertise.
Rely on human experts for next steps, such as accepting or transferring the patient.
Traditional AI/ML Approach:
Incomplete extraction of document images using OCR (Optical Character Recognition).
Bag-of-words technique for data insights, which lacks contextual information and relationships.
Use of simpler ML algorithms that are not contextually aware, resulting in the loss of important information.
GenAI and Transformer Models Solution
Data Extraction using Amazon Textract:
Able to handle handwritten text, distorted text, and low-quality document images.
Recognizes the layout of the document, including headers, footers, tables, and forms.
Mimicking Human Reading using GenAI:
Identify the key questions a human expert would ask when reading the documents.
Convert these questions into prompts and use GenAI agents to ask and answer the questions.
Compile the answers from the GenAI agents to create a summary of the relevant information.
Transformer-based Predictive Modeling using Amazon SageMaker:
Use the summary information as input to a transformer model, skipping the feature engineering step.
Take advantage of pre-trained transformer models with clinical data for transfer learning.
Future Improvements
Explore Visual Language Models: Evaluate the use of multimodal language models that can process both images and text to potentially improve the data extraction process.
Implement Guardrails: Validate the outputs of the GenAI agents to ensure there are no hallucinations or incorrect information.
Optimize Costs: Utilize AWS Trainium to potentially save up to 50% on the cost of training the models.
Key Takeaways
Automate the data engineering process using GenAI-driven techniques to mimic human reading and extraction of relevant information.
Leverage AWS Cloud technologies, such as Lambda, Step Functions, and SageMaker, to enable massive parallelization and scalability of the data processing pipeline.
These cookies are used to collect information about how you interact with this website and allow us to remember you. We use this information to improve and customize your browsing experience, as well as for analytics.
If you decline, your information won’t be tracked when you visit this website. A single cookie will be used in your browser to remember your preference.