Vi Thanh Pham PhD Defense

Invitation to PhD defence
Vi Thanh Pham - Statistical methods for analyzing ECG signals using deep learning and functional data analysis

Date Monday, March 24, 2025, at 14.00
Venue Øster Farimagsgade 5, 1353 Copenhagen K, room 2.0.63

After the defence, a reception will be held in 5-2-46.

Academic advisors:

Andreas Kryger Jensen, Section of Biostatistics, University of Copenhagen
Thomas Alexander Gerds, Section of Biostatistics, University of Copenhagen

Assessment committee:

Esben Budtz-Jørgensen, Section of Biostatistics, University of Copenhagen, Denmark
Benoit Liquet-Weiland, School of Mathematical and Physical Sciences, Macquarie University, Australia
Morten Mørup, Section for Cognitive Systems, Technical University of Denmark, Denmark

Summary

This PhD thesis develops statistical methods that integrate deep neural networks in statistical models to predict various health outcomes, with a focus on  electrocardiogram (ECG) data as predictors. Traditional ECG signal analysis relies mostly on fixed algorithms, which can be affected by patient motion during observation, leading to inaccurate readings, but also being specifically tailored to predetermined outcomes through non-data-driven feature extraction methods.

This research aims to eliminate such artifacts and bypass the prespecified feature selection step entirely by associating full ECG signals directly to clinical  outcomes of interest using methods from functional data analysis. Additionally, the project seeks to evaluate ECG measurements to predict short-term and long-term risks of various cardiovascular diseases, mortality, and other conditions by adapting existing methodologies and developing new algorithms. 

The analysis of ECG data presents significant challenges due to inherent noise, high dimensionality, and changes over time. The primary objectives of this PhD thesis include a method for alignment of multivariate functional data in time and the development of neural networks capable of processing functional inputs and time-to-event outcomes while adjusting for the time warping. Proper alignment of ECG data in time ensures accurate diagnosis, comparability, noise reduction, and data integration. Combining functional data analysis and neural networks into a hybrid model offers robust models that are less prone to overfitting.  Furthermore, the application of survival analysis to ECG data allows for the prediction and understanding of critical cardiac events, enabling timely interventions at the patient level.