The project will create multi-dimensional data-driven predictive simulation computer models enabling – for the first time – personalised stroke treatment, addressing patient’s needs in four stages: prevention, acute treatment, rehabilitation and reintegration. Linköping University participate mainly with methods for semantic integration of data sources and models for predicting individualised risk of stroke.
Semantic integration of data sources
The amount of data in the healthcare system is enormous, but the data is usually stored in different format for different healthcare organisations or even different patients. However, to use this diverse healthcare data effectively to create and validate models, the data from the different sources needs to be transformed into a consistent and comparable format. To solve this problem we work with ontology-based data integration where data from all sources are mapped to the PRECISE4Q ontology. The data have a range of different formats, using different scales and on different levels of granularity, etc, due to varying requirements when data was originally collected. A set of reasoning tasks is therefore developed aiming at bridging such differences in data sources. E.g., taxonomic reasoning using the PRECISE4Q ontology aggregates data into more generic categories. A more complex example is the harmonisations of quantitative and categorial representations of the same phenomenon (e.g., Glucose < 65 mg/dL = hypoglycemia). Data sources may group data differently in source information structures and represent epistemic features and negation differently. For the set of source data variables:
- Differences in data structure will be identified and categorized
- Transformation reasoning requirements will be developed depending on data source structure and modeling requirements phrased as competency questions
- Transformation rules will be defined and verified
- A reasoning chain will be set up according to the requirements
Predicting individualised risk of stroke
To be able to better prevent that a person develop stroke we create a model that is capable to predict the individualised risk of stroke with a time horizon of 3 to 5 years. Using the model, we will be able to run individualized simulations through time of the evolution of stroke risk under different treatment conditions (medication and changes in patient lifestyle). The outputs of this model will enable a healthcare professional to develop personalized intervention regimes for their patients. The model will be a hybrid model. A mechanistic model will be created by combining multi-level models and long-term prediction models. This model will be fused with machine learning models resulting in a multi-level and multi-scale simulation model. The model will be validated using registry data. Once validated, the simulation model will predict stroke risk based on key patient features that evolve over time, in response to different treatments, such as diet, exercise, and medication.
