This question is central to Anna Rodum Bjøru’s recently completed PhD thesis on Explainable AI (XAI) at NTNU.
Rodum Bjøru currently works as a data scientist at Vitalthings.
Her doctoral thesis addresses the following challenge: Advanced machine learning models are becoming increasingly accurate, but also less interpretable and more opaque. This makes them difficult to understand. We know what they conclude, but not always why. This is a problem when artificial intelligence is to be used in decision-making processes that affect people.
In his research, Rodum Bjøru places particular emphasis on causal explanations, which deal with cause and effect rather than merely statistical patterns. Such explanations are easier for people to understand because they resemble the way we naturally think.
XAI is a key focus area at Vitalthings because it is becoming increasingly important to be able to understand AI algorithms and, consequently, to explain to clinicians why the algorithms produce the results they do, and how this can help clinicians make sound decisions.
Read more about Rodum Bjøru’s research here: https://nva.sikt.no/registration/019df8bd5841-bd9c6683-f5b3-49da-b9c6-4750ab625899
PhD thesis, NTNU, 2026:
«Causal Post-hoc Explainable AI» by Anna Rodum Bjøru
- Advanced machine learning models are increasingly being used in decision-making processes across many areas of society. Although their performance is impressive, a significant drawback of modern machine learning is that complex model architectures are often not interpretable. This poses an immediate obstacle to a proper understanding of the decision-making logic underlying the decisions that are made.
- In response to this challenge, the field of Explainable Artificial Intelligence (XAI) has emerged, and a key objective within XAI is to ensure that automated predictions can be accompanied by corresponding explanations.
- Post-hoc explainability is one of the main approaches within XAI, and is characterised by the fact that explanations are generated only after the model to be explained has been fully trained. This type of explanation therefore has no impact on the model’s performance otherwise.
-
At the same time, the field of causality has developed a structured framework for formulating explanations, based on the modelling of causal relationships within a defined set of variables of interest. This makes it possible to model both intervention-based and counterfactual scenarios, with the aim of clearly identifying cause as an explanation of effect. Causal and counterfactual explanations are considered particularly useful for achieving a good understanding, with a firm foundation in real-world physical contexts.
- This thesis explores causal and counterfactual explanations as a starting point for the development of post-hoc XAI.
The patient monitor of the future
Guardian M10 offers continuous and precise monitoring of vital signs. The system monitors breathing rate and breathing pattern contactlessly, without the need for physical sensors.
EN 
NO