Machine Learning-Based Surrogate for Computational Modelling of Human Cardiac Cells
Ylinen, Olli (2026)
2026
Sähkötekniikan DI-ohjelma - Master's Programme in Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-06-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606087148
https://urn.fi/URN:NBN:fi:tuni-202606087148
Tiivistelmä
Computational modelling of human cardiac cells enables cheap, fast and easy experiments compared to modelling with real human cardiac cells that are done in a laboratory or with patients. Experiments like modelling specific drug effects, unique patients or diseases can be done with computational models. However, in cases where a large number of model runs is needed, the current models are slow to run.
The current human cardiac cell models are derived from measurements, where different parts of the cell are measured after which differential equations are fitted to those measurements. The whole model is then put together with combining the differential equations for different parts. These models can be quite complex so they need to be solved numerically, which can take tens of seconds for a single run.
The aim of the thesis was to make a tool to estimate quickly and accurately how electrophysiological changes made to the cell model would impact the end state of the model. To tackle this problem, the problem was turned into a supervised learning problem and chose 22 cell model parameters to be changed and 5 different model outputs for duration of 1.5 seconds to be predicted. Based on this the generated dataset was large consisting of 50 000 samples for training and 5000 for validation and 20 000 for testing. A convolutional neural network based surrogate model was developed based on the generated data to emulate the original cell models responses.
After iteratively developing and training the surrogate model, it achieved over 7 times lower absolute error values compared to the baseline. The model runtime using a modern CPU decreased around 1000 times compared to the differential equation model. Furthermore, the runtime could be decreased an additional multiple times with the use of a GPU. This showed that machine learning based models could have huge decrease in runtime compared to current models with the cost of small errors.
The current human cardiac cell models are derived from measurements, where different parts of the cell are measured after which differential equations are fitted to those measurements. The whole model is then put together with combining the differential equations for different parts. These models can be quite complex so they need to be solved numerically, which can take tens of seconds for a single run.
The aim of the thesis was to make a tool to estimate quickly and accurately how electrophysiological changes made to the cell model would impact the end state of the model. To tackle this problem, the problem was turned into a supervised learning problem and chose 22 cell model parameters to be changed and 5 different model outputs for duration of 1.5 seconds to be predicted. Based on this the generated dataset was large consisting of 50 000 samples for training and 5000 for validation and 20 000 for testing. A convolutional neural network based surrogate model was developed based on the generated data to emulate the original cell models responses.
After iteratively developing and training the surrogate model, it achieved over 7 times lower absolute error values compared to the baseline. The model runtime using a modern CPU decreased around 1000 times compared to the differential equation model. Furthermore, the runtime could be decreased an additional multiple times with the use of a GPU. This showed that machine learning based models could have huge decrease in runtime compared to current models with the cost of small errors.