Towards trustworthy and accurate deep learning models for medical time series data


Doctors spend hours upon hours looking through data to deliver the correct diagnosis to patients. Cardiologists may look for heart arrhythmia in electrocardiography (ECG) signals, while neurologists look for epileptic seizures or abnormal sleep patterns in electroencephalography (EEG) signals. We are constantly learning about the lack of doctors in the medical sector, and thus imagine the impact, if we could automate some of the tasks that now consume a significant amount of time from doctors. This is achievable using machine learning algorithms, but the current accuracy and trustworthiness of models leave a lot to be desired.

Over the recent decades, deep learning has become the most predominant driver of progression in machine learning. Much of this advancement has been made in computer vision, and natural lan- guage processing tasks – largely due to the fact that enormous datasets are easily available within these topics. Time series tasks have not to the same extent enjoyed similar advances in the deep learning era, which can partially be prescribed to the lack of large datasets with high-quality labels. This is especially true in the medical field where data is difficult to get a hold of and high-quality labels require expensive expert knowledge. Therefore, we hypothesize that larger and deeper networks are simply not the solution to this problem.

Instead, we take a step back and consider how, for instance, novice neurologists learn to look at EEG data. Contrary to deep learning models, they do not need to see millions of examples of similar EEG sequences to understand what is going on. Instead, an experienced neurologist can explain the overall concepts through a few examples. The novice neurologists can then extrapolate this knowledge onto new unseen EEG sequences. The aim of this project is to move deep learning models closer to this conceptual and much more sample efficient type of learning. We propose that to do this, we need to alter the representations learned by the deep learning models to causally informed representations.

The first issue that one faces on this journey is that we currently have very little understanding of how deep learning models learn. Deep learning models are to a large extent black boxes, which means that it is difficult to understand why the models make their decisions. This is an issue for two reasons: in medical applications, it is difficult for doctors and patients alike to trust the decisions of deep learning models if they are unable to understand them. Furthermore, the lack of explainability of deep learning models makes it difficult to “diagnose” the models’ behaviour. The first step is therefore to gain a better understanding of the learned representations. A substantial part of the project will therefore be focused on enhancing the explainability of representations of deep learning models.

Once we understand the learned representations, the next step is then to alter them. Therefore, the second part of the project is focused on representation learning – more specifically on causal repre- sentation learning. Causality is traditionally widely used in statistics and has recently been gaining more and more interest in the machine learning community. Learning the causal model of a certain data type opens the door to a wide range of applications. Knowing the causal generative model can help us diagnose patients and enhance our understanding of the mechanisms of the brain or various diseases. Furthermore, the framework of causal learning allows us to explicitly incorporate our prior knowledge into the models.

With this project, we aim to bridge the gap to a future where deep learning models can learn in a more sample efficient manner. Furthermore, we hope to enhance the trustworthiness of deep learning models by making them explainable and enhancing the causal understanding.

PhD project

By: Thea Brüsch

Section: Cognitive Systems

Principal supervisor: Tommy Sonne Alstrøm

Co-supervisor: Mikkel Schmidt

Project title: Causal Explainability for Time Series



Thea Brüsch
PhD student
DTU Compute


Tommy Sonne Alstrøm
Associate Professor
DTU Compute
+45 45 25 34 31


Mikkel Nørgaard Schmidt
Associate Professor
DTU Compute
+45 45 25 52 70