Modular approach to comprehensive oncological decision support
Principal Investigator
Name
Daniel Schneider
Degrees
Ph.D.
Institution
Innovation Center Computer Assisted Surgery, Leipzig University
Position Title
PostDoc
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCO-813
Initial CDAS Request Approval
Jul 22, 2021
Title
Modular approach to comprehensive oncological decision support
Summary
Our research group ‘Models for personalized medicine‘ develops a modular approach to comprehensive oncological decision support. Central to the project is a digital patient representation semantically linking all known information about a patient and their clinical history, the so-called digital twin. Hierarchically structured data processing modules with confined specific tasks interact with the digital twin and with each other. These modules fulfill assignments including data preparation and summarization, outcome prediction, and treatment guidance. The proposed framework can be used as assistance system in clinical practice.
We propose a pilot study involving a tumor board review assistance application using the PLCO trial data. Due to its wide range of features from various clinical sources and followup information, the PLCO data is particularly suitable for comprehensive decision support studies. For the proposed study, various modules with specific data processing or inference tasks will be developed and semantically linked into the framework.
Next to the implementation of clinical guidelines, data processing based on machine learning approaches plays an integral part in our proposal. Datasets from clinical studies and health records are usually hierarchically structured and incomplete, containing collections of different interrelated entities such as demographics, symptomes, diagnoses, treatments, or lab results. Conservative machine learning methods treat the clinical data simply as ‘bag of features‘ and thus, completely neglect relations originating from the data generation process. For that reason we intend to use graph-based models and recent graph neural networks (GNN), which provide promising approaches to learn effective data representations on medical data such as provided by the PLCO trial.
For clinical adoption of the proposed application, additional requirements such as human comprehensibility and uncertainty estimation need to be satisfied. While the former prerequisite can be achieved with feature contribution analysis via e.g. SHAP values, the latter may be fulfilled with methods of variational inference.
We propose a pilot study involving a tumor board review assistance application using the PLCO trial data. Due to its wide range of features from various clinical sources and followup information, the PLCO data is particularly suitable for comprehensive decision support studies. For the proposed study, various modules with specific data processing or inference tasks will be developed and semantically linked into the framework.
Next to the implementation of clinical guidelines, data processing based on machine learning approaches plays an integral part in our proposal. Datasets from clinical studies and health records are usually hierarchically structured and incomplete, containing collections of different interrelated entities such as demographics, symptomes, diagnoses, treatments, or lab results. Conservative machine learning methods treat the clinical data simply as ‘bag of features‘ and thus, completely neglect relations originating from the data generation process. For that reason we intend to use graph-based models and recent graph neural networks (GNN), which provide promising approaches to learn effective data representations on medical data such as provided by the PLCO trial.
For clinical adoption of the proposed application, additional requirements such as human comprehensibility and uncertainty estimation need to be satisfied. While the former prerequisite can be achieved with feature contribution analysis via e.g. SHAP values, the latter may be fulfilled with methods of variational inference.
Aims
- develop and validate GNN and other deep learning-based modules on the PLCO dataset for data preparation, extraction of higher order features, treatment advice, and outcome prediction (mortality and recurrence)
- compare performance of data-driven graph representation learning (via GNN) with knowledge-based graph modeling (e.g. Bayesian networks) for above tasks
- add variational inference methods for uncertainty estimation (e.g. Bernoulli dropout) and explainability techniques (e.g. SHAP-values)
- compile and semantically link the set of modules as a tumor board review decision support pilot application based on the PLCO data
Collaborators
N.A.