AWS Public Sector Blog

Improving patient care in Canada with Amazon Comprehend Medical

Amazon Comprehend Medical is a natural language processing (NLP) service that simplifies the use of machine learning (ML) to extract relevant medical information from unstructured text often found in clinical charts or doctor’s notes. Since the service launched in the AWS Canada (Central) Region in June 2019, it opened up possibilities for Canadian healthcare organizations to better serve patients.

By understanding and analyzing the insights and relationships that are “trapped” in free-form medical text, including hospital admission notes and a patient’s medical history, doctors and clinical researchers can improve patient care. Amazon Comprehend Medical quickly and accurately gathers information, such as a patient’s medical condition, medication, dosage, strength, and frequency from a variety of sources like doctors’ notes, clinical trial reports, and patient health records.

Bringing this service to the Canadian Region allows hospitals to develop advanced computing technologies and train advanced ML models to help with diagnosis and treatment.

Speeding time to treatment

Vancouver General Hospital (VGH) and University of British Columbia (UBC) researchers are among the organizations who leverage Amazon Comprehend Medical and Amazon SageMaker to create their own machine learning models that can triage x-rays to provide a better healthcare experience.

Imagine this scenario: a patient comes into the hospital with symptoms of pneumonia. The doctor takes an x-ray. The patient’s x-rays are then analyzed by a ML model trained on Amazon SageMaker that interprets the x-ray for the presence of infection. The algorithm then determines a priority for the x-ray to be seen by a radiologist. Ultimately, the patient in need would be seen more quickly and put on a treatment plan in less time than would have ordinarily taken to capture, assess, and diagnose. By using ML technology, fewer patients are left sitting in the waiting room or sent home waiting for answers. Treatment can be started efficiently, potentially saving lives.

This scenario was one of the first ML models the team experimented with at Vancouver General Hospital.

Having access to the AWS toolset, including Amazon Comprehend Medical and Amazon SageMaker along with storage and compute, brings forth three major potential benefits for VGH:

  1. Efficiency – Having access to petaflops of processing power enables and empowers clinical data scientists to train models quickly and iterate on models efficiently. Accurate models are created faster, potentially improving care for more patients sooner.
  2. Cost savings – Managing, securing, and servicing large data centers are expensive but having access to AWS on an as-needed basis decreases the headaches and costs associated with dealing with the technical aspects of managing these high-end systems. Ultimately, clinical data scientists can focus more time on patient care and outcomes.
  3. Accuracy – AWS uses storage, security, processing, and compute systems that enable clinical data scientists to produce high-accuracy models. The better the accuracy, the better the patient care.

Training machine learning models on AWS

In order to train the image-detection model to be able to make these predictions, the research team at UBC is training the ML algorithms on Amazon SageMager. Amazon SageMaker is a fully managed service that covers the machine learning workflow, from choosing an architecture and model, to training models, iterating on results and optimizing accuracy for deployment in the clinical environment.

“Amazon SageMaker alleviates our need for an on-site compute cluster and gives us a lot of flexibility when it comes to compute performance and cost. We do a lot of our work in short sprints, so being able to ‘rent’ time on the server only when we need it, is a big improvement on an upfront investment in GPUs that may not see much consistent usage,” said Brian Lee, researcher at the University of British Columbia.

In order to provide their model with pre-classified data (ie. labeling images with the presence or absence of pneumonia), the team uses Amazon Comprehend Medical. This allows them to only acquire images that have a specific diagnosis or something particular from the radiologist’s reports.

The VGH research team also developed their own anonymization tool called SapienSecure that extracts Personal Identifiable Information (PII) data from medical images and written medical reports. The tool will be released publicly by SapienML later this year, and has functionality to integrate directly into the AWS platform. This step improves patient safety and security, but also allows for easier adoption of AWS tools by other hospitals and institutions.

“SapienSecure is our way of giving back to the clinical data science community by offering a way to accurately anonymize medical data, so that researchers and technology developers can feel confident that their data is anonymized before training a model,” said Dr. William Parker, researcher and Radiology Resident at UBC. “It plugs into AWS, so hopefully it will help researchers to take advantage of the power of AWS’s cloud offerings.”

Learn more about Amazon Comprehend Medical.