Munich Leukemia Lab Advances Leukemia Diagnosis with Amazon SageMaker Machine Learning

Munich Leukemia Lab (MLL) partnered with the Amazon Machine Learning Solutions Lab to build a machine learning pipeline leveraging Amazon SageMaker to classify 30 leukemia subtypes using next generation sequencing (NGS) data. Manually classifying leukemia subtypes is complex, slow, and requires expensive specialized equipment and highly skilled experts, leading to turnaround times up to ten days. MLL and AWS developed a feature extraction process transforming varied NGS data into tabular form with over 70,000 features, followed by training a LightGBM model with SageMaker Hyperparameter Optimization achieving 82% accuracy in 5-fold cross-validation. The interpretable model uses SHAP to explain feature impacts per patient, aiding clinical decision making. The solution accelerates diagnosis with high accuracy, reducing time and cost while supporting precision treatment strategies for 30 leukemia subtypes.

Industry
Healthcare
Location
Germany
Published
May 2021

Reported outcomes

82%

accuracyQuality & accuracy

Strategic outcomes

New product / capabilityBuilt automated leukemia subtype classification pipelineSpeed & agilityEnabled near real-time diagnosisBetter decisions & insightEnabled data-driven clinical decisionsCustomer experience & trustImproved interpretability for clinicians

Catalog median for quality & accuracy deployments: +90% across 282 reported metrics. Compare benchmarks →

Primary read

Use case focus

Showing 2 of 2

  • 1Predictive Analytics
  • 2Diagnostic Assistance
  • Leukemia subtype diagnosis is currently manual, time-consuming, relies on complex laboratory techniques and expert morphological review.
  • Turnaround times to diagnosis can be up to ten days.
  • Imbalanced and high-dimensional NGS data is challenging for ML models to classify accurately.
  • Developed an automatic leukemia subtype classification pipeline using Amazon SageMaker and SageMaker Notebooks.
  • Extracted and transformed multi-modal genomics data into rich tabular features for machine learning input.
  • Trained LightGBM model with SageMaker Hyperparameter Optimization to improve accuracy.
  • Used explainable AI with SHAP for prediction interpretability to assist clinicians.
  • Achieved 82% classification accuracy over 30 leukemia subtypes, validated by 5-fold cross-validation.
  • Reduced diagnostic turnaround time from days to near real-time analysis.
  • Enabled data-driven, algorithmic diagnostic decisions supporting precision medicine at a leading leukemia lab.
Architecture

The solution uses Amazon SageMaker for model training and hyperparameter tuning, SageMaker Notebooks for development, and stores data securely in AWS S3 buckets compliant with healthcare data regulations. The model integrates multi-modal NGS features derived from WGS and WTS data. Explainability is provided using SHAP for model interpretability.

Sources & evidence1
Groundedness: 5/5

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

Explore related AI use cases

Was this useful?