Thesis Project

Lucas Kook

"The focus is on the development of probabilistic predictive models allowing to identify the most important predictors or even causal factors which can be used for probabilistic predictions of the treatment outcome."

Homepage

Thesis: Deep and robust distributional regression with applications to intervention planning in acute ischemic stroke

Supervisors:
Prof. Dr. Torsten Hothorn (UZH)
Prof. Dr. Beate Sick (ZHAW)

Funding: The PhD project is funded via the FreeNovation grant with the title is "Improving treatment decisions for acute stroke patients through Deep Learning based risk analysis". The principal investigator of the project is Prof. Dr. Beate Sick. Collaboration partners are Prof. Dr. Susanne Wegener (USZ) and Dr. Helmut Grabner (ZHAW). The focus is on the development of probabilistic predictive models allowing to identify the most important predictors or even causal factors which can be used for probabilistic predictions of the treatment outcome.

Project description: In numerous medical applications we face the problem of predicting a patient’s outcome based on clinical images and tabular data. Classical statistical models fall short of being able to include data as complex as images without any form preprocessing (e.g., feature engineering). Contrary to classical statistical model, deep convolutional neural networks seamlessly make use of image data for prediction. However, deep neural networks lack interpretability and oftentimes use loss functions or evaluation metrics with undesirable statistical properties, such as impropriety. Models that borrow interpretability from a statistical approach and flexibility from a deep learning approach are seldom found in the literature. In this doctoral thesis we aim to unite classical statistical models with contemporary deep learning models. We build upon the general class of transformation models and show how a modular approach to model building enables the aforementioned trade-off between interpretability and flexibility. The resulting deep transformation models will be applied to data from the University Hospital Zurich.

The data set is comprised of patients from USZ admitted between December 2017 and March 2019 with acute ischemic stroke and contains multiple magnetic resonance image modalities (DWI, ADC, T2, FLAIR) alongside clinical covariates, such as blood pressure, age, BMI among others. The response is an ordinal score measuring the functional outcome 3 months after stroke (modified Rankin scale, mRS). The mRS ranges from 0 (no symptoms) to 6 (death due to stroke). The first goal is to develop a prediction model for the functional outcome (mRS) after acute ischemic stroke. The second goal is to identify potentially causal structures that explain the functional outcome through the clinical and/or image data to help guide neurological experts in intervention planning for future patients with acute ischemic stroke.

Intervention planning for patients with acute ischemic stroke based on observational data is a highly ambitious goal, because it involves the search for causal structures in vastly complex and high dimensional data. In general, it is known that identifying causal relationships from observational data without access to interventional data is impossible without making strong assumptions about an underlying causal model. Recently, Buehlmann and colleagues proposed anchor regression and anchor boosting to address a "diluted form of causality" by casting the problem of finding causal variables into finding a set of covariates under which the residual distribution is invariant towards perturbations. This solution results in predictions that are robust against perturbations in future data. Among others, we investigate anchor regression as a potential method to address the problem of intervention planning in patients with acute ischemic stroke.

Planned duration: 01.03.2020 - 01.09.2023