Explainable Machine Learning for Clinical Decision Support with Amazon SageMaker

Healthcare organizations require explainability in machine learning models used for clinical decision support, especially for triage decisions based on patient admission notes. Amazon Web Services deployed a fine-tuned Hugging Face BigBird BERT model on Amazon SageMaker to predict patient outcomes such as in-hospital mortality from clinical notes. Amazon SageMaker Clarify was used to provide SHAP value-based explanations for model predictions to enhance trust and enable clinicians to understand AI decisions for individual patients. The solution allows clinicians to review feature-level explanations for predictions, supporting better informed clinical decisions and adoption in hospital triage workflows.

Industry
Healthcare
Published
August 2023

Reported outcomes

Strategic outcomes

Customer experience & trustImproved trust in clinical AI decisionsCustomer experience & trustEnabled explainable triage adoptionScale & capacityDemonstrated scalable explainable ML deploymentNew product / capabilityAdded transparent clinical reasoning support

Primary read

Use case focus

Showing 3 of 4

  • 1Explainable AI
  • 2Clinical Decision Support
  • 3NLP
Healthcare providers need interpretable machine learning models to support clinical decisions from unstructured clinical notes with transparent explanations to gain trust and adoption by medical staff.
  • Deployed a fine-tuned Hugging Face BigBird BERT model on Amazon SageMaker for predictions based on admission notes.
  • Enabled SageMaker Clarify to generate SHAP explanations providing detailed feature contributions to outcomes.
  • Set up a real-time explainability endpoint integrated with hospital decision systems to render interpretable results for individual patients.
  • Used a Jupyter notebook for model deployment and explanation demonstration, supporting integration into electronic health records and clinical workflows.
  • Improved interpretability and trust of AI models in clinical settings enabling clinicians to make better-informed decisions.
  • Facilitated adoption of AI in hospital triage processes by providing explainable predictions based on medical notes.
  • Demonstrated a scalable approach for explainable ML deployments in healthcare with Amazon SageMaker.
  • Enhanced patient care by combining AI predictions with transparent clinical reasoning support.
Architecture

Architecture consists of a SageMaker endpoint hosting the fine-tuned Hugging Face BERT model with SageMaker Clarify integrated for real-time SHAP value explainability. Real-time requests are sent from clinical systems or Jupyter notebooks, and responses provide prediction scores with feature explanation visualizations to clinicians.

Sources & evidence1
Groundedness: 3/5

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases
This website uses cookies to enhance the user experience. Learn more.