Project #1: Machine Learning Tools for Automated Image-based ECG

In collaboration with Fabio Bonassi and Jiawei Li.

Cardiovascular diseases are the leading cause of death worldwide, and the electrocardiogram (ECG) remains essential for their diagnosis. While most AI research has focused on raw ECG signals, hospitals and telehealth systems often store ECGs as images, creating new challenges and opportunities for automated analysis. This project will investigate deep learning methods tailored to image-based ECGs, aiming to improve interpretability, robustness to noise and artifacts, and clinical relevance. As part of the work, you will gain experience with cutting-edge computer vision techniques, work on some of the world’s largest open ECG datasets, and contribute to developing AI tools with direct clinical impact.

Some relevant literature:

Prerequisites: Linear algebra, programming (Python), statistical machine learning, and deep learning. Knowledge of signal processing techniques is a merit.

Project #2: Federated Learning for Physiological Signals

In collaboration with José Mairton Barros da Silva Jr.

Strict regulations on patient data protection in Sweden and many other countries limit how medical data can be stored, processed, and shared, creating challenges for developing and maintaining high-quality machine learning algorithms for ECG interpretation. This thesis will investigate federated learning methods as a privacy-preserving alternative to centralized training, with a focus on ECG analysis. The goal is to develop frameworks that enable collaborative model training across healthcare institutions while ensuring compliance with data protection laws and maintaining model performance. The student will implement and evaluate federated learning approaches in which models are trained locally and only model parameters—not patient data—are shared, and may also explore techniques such as differential privacy and strategies for robust performance evaluation.

Some relevant literature:

Prerequisites: Linear algebra, programming (Python), statistical machine learning, and deep learning. Knowledge of signal processing techniques is a merit.

Project #3: Machine Learning Tools for Breathing Signals

In collaboration with researchers from Karolinska Institutet Artin Arshmian and Reza Baboukani.

Breathing patterns are not only vital for sustaining life but also carry rich information: a recent study in Current Biology (Soroka et al., 2025) showed that long-term nasal airflow contains unique “respiratory fingerprints” that can identify individuals with near-biometric accuracy (∼97%) and correlate with physiological traits such as body mass index (BMI) and emotional state. Yet, these results were restricted to nasal breathing, leaving open questions about whether similar signatures exist in oral breathing and whether combining nasal/oral breathing with heart rate signals can improve the prediction of physiological markers. In this project, the student will analyze a rich dataset collected under both nasal and oral breathing conditions with synchronized heart rate recordings during rest (5 minutes) and task performance (45 minutes), developing and applying statistical and machine learning models to assess the predictive power of these signals for physiological markers, sex, and age, as well as their potential for individual identification across breathing modes. The work bridges biomedical signal processing, statistical learning, and digital health, with potential implications for novel biomarkers.

Some relevant literature:

Prerequisites: Linear algebra, programming (Python), statistical machine learning. Knowledge of deep learning and signal processing techniques is a merit.

Project #4: Distilling ECG segmentation models for differentiable ECG segmentation

Traditional ECG segmentation methods typically rely on combinations of filter banks and thresholding with manually set parameters, which, while effective, limit flexibility and integration with modern machine learning frameworks. This project aims to distill knowledge from well-established ECG segmentation algorithms into differentiable models and then combine these with limited cardiologist-annotated ECG data to train end-to-end segmentation networks. The approach leverages the interpretability and robustness of classical signal-processing techniques while enabling gradient-based optimization and deep learning integration.

Some relevant literature:

Prerequisites: Linear algebra, programming (Python), statistical machine learning, and deep learning. Knowledge of signal processing techniques is a merit.

Project #5: Reinforcement learning to conciliate making measurements and predictions for ECG

This project explores how reinforcement learning can be used to reconcile ECG measurements (e.g., intervals, amplitudes) with predictions of cardiac abnormalities. The idea is to generate a dataset containing both automated measurements and diagnostic predictions, and then train a reinforcement learning model—using methods such as GRPO (Generalized Reinforcement Policy Optimization)—to encourage consistency between the two outputs. By aligning measurement-based reasoning with prediction-based classification, the resulting model could provide more interpretable and clinically reliable ECG analysis.

Some relevant literature:

Prerequisites: Linear algebra, programming (Python), statistical machine learning, and deep learning. Knowledge of signal processing techniques is a merit.

Project #6: Adversarial training for age prediction

Project #7: Adversarial training for autorregressive models