Stockholm university

Language models help doctors make better decisions

Aron Henriksson, researcher and teacher at the Department of Computer and Systems Sciences (DSV), has been promoted to associate professor. His research is in natural language processing, and has important implications for healthcare.

Illustration of multiple screens showing heart frequencies.
Photo: Mingwei Lim/Unsplash.

Congratulations, Aron! Please tell us about your research.

“Thank you! My main research area is natural language processing, primarily applied to health. The idea is basically to develop tools capable of making use of the large amounts of clinical data that are generated in healthcare and stored in electronic health records. Much of this data is in the form of free-text, such as clinical notes about patients. This clinical text contains valuable information that can be analysed for different purposes. It can, for example, be used for creating clinical prediction models and clinical decision support systems. I find it highly motivating to work with health applications of natural language processing and machine learning. This allows me to contribute to improving healthcare and help physicians make better, more timely decisions based on automatic analysis of data.”

Portrait photo of Aron Henriksson, new associate professor at DSV.
Aron Henriksson. Photo: Private.

Can you tell us more about the research topics you are working on?

“Currently, my research focuses on large language models adapted to the clinical domain. These are machine learning models that learn by observing how language is used in enormous amounts of text data. Language models are often designed to be generic, but research has shown that domain-specific language models often perform significantly better in the target domain. In order to be used effectively in the clinical domain, the models need to account for the very particular type of language found in health records.”

“I also study how clinical language models can be leveraged in multimodal systems that utilize the heterogeneous data in electronic health records – both structured data and more unstructured data such as clinical text. Combining different types of data in machine learning models often leads to better predictive performance. In addition to health, I’m interested in applying natural language processing to other areas, such as data-driven requirements engineering and technology-enhanced learning.”


What’s it like to work with sensitive data about people and their health?

Working with sensitive data means that privacy is very important to consider. So when developing clinical language models, the goal is of course that they should perform well on some downstream task, for example predicting which diagnoses to assign to a patient. But in order to allow models to be shared, it is also important to train them in a privacy-preserving manner. This reduces the risk of exposing sensitive information in the training data and can be achieved in different ways, for example by automatically de-identifying the data used for training the language models.”


But you also have other important responsibilities to attend to…?

“I’m actually on parental leave with our second child, who is now 11 months old. So taking care of my family is my main ‘job’ at the moment, but I’m also trying to carry out some of my research commitments in parallel. I supervise doctoral students and I am the PI of a research project funded by Region Stockholm. This project is a collaboration with researchers at Karolinska Institutet and physicians at Karolinska University Hospital. We focus on outcome prediction for COVID-19 patients and early prediction of sepsis.


Have you published any papers in this research project?

“Yes, we have presented a number of papers in scientific conferences and I’m currently working on an article for submission to a special issue of a journal. It’s a follow-up study in which we are evaluating how well our multimodal model is able to predict which COVID patients at the emergency department are in need of hospital care. Some do not need to be hospitalized and can be safely discharged. But there is also a risk of ‘incorrectly’ discharging patients from the emergency and then having to readmit them later. If the model is able to distinguish these cases well, it could help to avoid unnecessary hospitalizations and thereby reduce healthcare costs. It could also lead to improved patient outcomes by ensuring that those in need of care are not sent home from the emergency.”


When and where did your research journey start?

“I started working at DSV as a research assistant in 2010 after completing a master’s degree at KTH. I was formally admitted as a PhD student in 2012 and defended my thesis in 2015. Before that, I did a bachelor’s degree in computer science at RMIT University in Melbourne, Australia. During my PhD studies, I also spent some time as a visiting researcher at the University of California San Diego in the US and at the University of Trento in Italy.”


How do you view your role as an educator?

“I teach several courses that are, in some ways, related to my research and interests in storing and analysing large amounts of data. Together with colleagues at DSV, I have developed and am responsible for two courses: ‘Big Data with NoSQL Databases’ and ‘Scalable and Responsible AI in Organizations’. Developing and teaching these courses has allowed me to broaden my knowledge and learn about areas that are adjacent to my own research. I enjoy taking complex topics and teaching them by breaking them down in a structured way. I also always try to give students a clear idea of how different parts build on, and relate to, each other.”

This article is also available in Swedish

More information

Contact information for Aron Henriksson

The title of his PhD thesis is “Ensembles of Semantic Spaces: On Combining Models of Distributional Semantics with Applications in Healthcare”.
It can be downloaded from Diva

Read about the courses “Big Data with NoSQL Databases”
and “Scalable and Responsible AI in Organizations”

About the Department of Computer and Systems Sciences (DSV)

Text: Åse Karlén