Genomics England has initiated an ambitious project in collaboration with data science teams and professional services from AWS to improve the identification of cancer subtypes and survival prediction using machine learning (ML). The initiative aims to combine genomic data and histopathological images to achieve higher accuracy in the models.
In the first proof of concept, the Pathology-Omic Research Platform for Integrative Survival Estimation (PORPOISE) platform was used to analyze data from breast cancer and gastrointestinal cancer. While this model was advanced, it showed certain limitations by excluding gene expression data from the analysis.
To overcome these limitations, AWS developed a new model called Hierarchical Extremum Encoding (HEEC), designed to enhance accuracy and interpretability. HEEC integrates hierarchical representations at multiple spatial levels and uses decision trees to reduce the risk of overfitting. The results have shown that HEEC significantly improves accuracy compared to the best individual modal model by combining multiple data modalities.
In a later phase, the Hierarchical Image Pyramid Transformer (HIPT) model was implemented, trained in a self-supervised manner, to enhance the results of the previous phases. Preliminary results indicated a significant improvement in the accuracy of survival analyses.
Architecturally, the proof of concepts implemented a modular architecture on AWS using SageMaker, which allows for separating data processing and model training, with scalability and efficiency advantages. The architecture also utilizes containers and CI/CD pipelines to automate and manage the deployment of resources sustainably and securely.
The implementation of these technologies provides Genomics England with advanced tools to explore the potential of machine learning in precision medicine, with the goal of improving cancer patient outcomes.
“At Genomics England, our mission is to realize the enormous potential of genomic and multimodal information to advance precision medicine,” commented Dr. Prabhu Arumugam, Director of Clinical Data and Imaging at Genomics England.
via: MiMub in Spanish