We perform teaching and research in machine learning strategies for the pattern analysis of various kinds of data. This comprises statistical models for clustering, graphical models for network inference and algorithmic methods to efficiently find these structures in the data.

Opportunities

For more information please contact the corresponding advisor(s)

Semester Project

Privacy-preserving negotiations

Carlos Cotrini Prof. Dr. Joachim Buhmann

The rise of online marketplaces in the sharing economy has brought about a new way for individuals to exchange goods and services by finding matching users. Platforms such as car sharing services and online dating services have gained popularity due to their ability to connect users with similar interests.
In this research proposal, we aim to formalize this search process using contextual bandits, where the learner searches for compatible users and the arms are other users. However, an important aspect that differentiates this setting from previous work on bandits is that the arm features are not readily visible to the learner. Furthermore, arms may be sensitive about their privacy, making it challenging to strike a balance between maximizing the learner's reward and respecting the arms' privacy requirements.
To address these challenges, we propose combining the UCB algorithm with local differential privacy to achieve reasonable regret bounds for the learner while respecting the arms' privacy requirements. The ultimate goal of this research is to provide a solution that empowers individuals to find compatible users while maintaining their privacy.

Proposal

Federated Multi-Task Learning of Surgical Phases

Ami Beuret Prof. Dr. Joachim Buhmann

Surgical phase recognition is the task of classifying video frames to one of the predefined phases of surgical workflow. Surgical phase recognition is one of the important building blocks of computer-assisted surgery. Recently, surgical phase recognition has seen a significant progress [Garrow et al., 2021] due to the availability of large detests.
Given the sensitive nature of healthcare data, the use of federated learning has been previously explored [Antunes et al., 2022, Corinzia et al., 2019]. Surprisingly, the applications of federated learning on surgical data remains quite unexplored [Kassem et al., 2022]. This project aims at bridging this gap by using multi-task federated learning [Corinzia et al., 2019] on cholec80 surgical dataset [Twinanda et al., 2017]. Given that recorded data from different health institutions/surgoens are heterogeneous, a multi-task federated framework can tackle the problem more naturally.
**Very good skills with both tensorflow probability and pytorch distributions are required. With this thesis we aim for a publication in a medical venue.**

Proposal

Video-based ICU Stroke Detection

Robin Geyer Prof. Dr. Joachim Buhmann

This project is part of an ongoing cooperation between ETH Zürich’s Institute for Machine Learning and the Departments of Neurology and Cardiac Surgery of Charité University Hospital Berlin. Its aim is to enable early stroke detection in high-risk intensive care patients by detecting anomalous asymmetrical movement patterns leveraging pose estimation and landmark detection models. Most models trained on publicly available datasets experience a significant performance drop when evaluated on data collected on intensive care units. This can be attributed to the domain shift between training and test data. The student will investigate how to improve the performance of state of the art pose estimation algorithms for deployment on intensive care units by either fine-tuning an existing pose estimation method (or ensemble of methods) or post-processing its output such that its performance on intensive care data is enhanced. Additionally, the developed method will be used to create a pose-landmark dataset from the ICU-video data to compare and analyze movement patterns.

Implicit Neural Representation for Computed Tomography with Meta-Learning

Xia Li Prof. Dr. Joachim Buhmann

This research project aims to explore an effective meta-learning-based algorithm to learn general CT priors, which can then be adapted to a specific patient for personalized optimization given limited X-ray scanning (SVCT and CBCT). From the medical perspective, this project aims to combine the knowledge of general human organs prior to patient-specific symptoms, thus promoting the development of personalized medicine. We will develop and validate this method by using fractional CT images from patients treated at the Paul Scherrer Institut~(PSI). Depending on the progress, this method can be also extended to the 4D scenario, when time-resolved fluoroscope mode of the CBCT projection is available online.

Robust Nocturnal Hypoglycemia Prediction

Alina Dubatovka, Simon Föll Prof. Dr. Joachim Buhmann

Diabetes mellitus aﬀects approximately 460 million people worldwide. As for the individual, diabetes mellitus is characterized by elevated blood glucose levels (hyperglycemia). Treatment with insulin or insulin secretagogues decreases blood glucose and may lead to blood glucose levels below the normal range, a condition called hypoglycemia. Hypoglycemia is among the most relevant acute complications of diabetes mellitus. Because almost 50% of all episodes of severe hypoglycemia occur at night, it can cause impaired cognitive function, convulsions, coma, and cardiac arrhythmias, potentially resulting in sudden death (dead-in-bed syndrome).

Next