Public defence in Information and Computer Science, M.Sc. (Tech) Iiris Sundin

Public defence from the Aalto University School of Science, Department of Computer Science
Doctoral theses hanging on the wall

Title of the doctoral thesis: Interactive Knowledge Elicitation for Decision-Support Models in Precision Medicine

Doctoral student: Iiris Sundin
Opponent: Dr. Danielle Belgrave,DeepMind, United Kingdom
Custos: Prof. Samuel Kaski, Aalto University School of Science, Department of Computer Science

Artificial intelligence (AI) has a great potential in the field of precision medicine. It can be used to choose individualized treatments to diseases based on a patient's characteristics, or to help in designing new drugs, among other things. Nevertheless, in many problems in precision medicine the performance of such AI system still suffers from lack of data, because it is difficult to learn a good model that can generalize to unseen cases from small data. As a result, the model may make mistakes when predicting the effect of a treatment. In these cases, human experts' knowledge can provide a valuable source of information to improve the prediction accuracy of an AI system. 

This thesis studies human-in-the-loop machine learning and how to use it in precision medicine. The goal is to make a machine learning model perform better at a task when it is allowed to interact with a human expert. The thesis focuses on what kind of questions the AI should ask an expert, and how to choose which questions are asked. As example use-cases the thesis studies individualized treatment of cancer and AI assisted design of small molecules. The results show that interaction with an expert allows predicting more accurately how a patient would respond to a treatment. The usefulness of the interaction is further improved by concentrating interaction to questions that help to identify the best treatment for a patient. In the second use-case, human-in-the-loop approach is applied to the design of small molecules for drug development. The developed method allows a chemist to give feedback to an AI system that designs molecules. The feedback is used to infer the chemist’s implicit goal, and to guide the system to suggest new molecules that better match the chemist’s goal. This thesis provides a basis for developing interactive AI systems that will allow integrating expert’s knowledge into them. The results show that augmenting machine learning with human expertise will enhance performance of AI systems in cases where there is too little data for an AI system to learn from.

Thesis available for public display 10 days prior to the defence at:

Contact information:

Doctoral theses in the School of Science:

  • Published:
  • Updated: