When Alexander Karlsson was looking for a topic for his master’s thesis in statistics and machine learning at the Department of Computer and Information Science, he talked to his sister.
“She is a dietician in the Unit for Clinical Nutrition at the Sahlgrenska University Hospital in Gothenburg. I would love to work in medical care and so I asked her if she knew of any researcher who was working on an interesting question relating to data processing”, says Alexander Karlsson.
This brought him into contact with a researcher at the same unit as his sister. The aim of his master’s thesis became to investigate whether a statistical model could improve a method for measuring the body’s muscle mass.
“When dieticians give dietary advice, it’s important that they know the body composition of the patient, i.e. the amounts of fat, bone and muscle in the body. The amount of muscle is particularlyimportant. This has proved to be a significant indicator of muscular function, and of how careful the patient is about diet in general.”
BIA or radiology
The Unit for Clinical Nutrition uses two methods to estimate body composition. One of these is a technique known as bioelectrical impedance analysis, or BIA. In this case, electrical signals are transmitted through the body and the strengths of the signals that emerge at other parts of the body are measured. The electrical properties of muscle differ from those of bone and fat, and this affects the strength of the signal. By analysing the signals, it is possible to build up a complex image of body composition – the problem is how to interpret them.
“BIA is easy to use. The patient stands on something that looks like an electronic weighing scale, grips two handles, and the measurements are completed after 20 seconds. The scale itself has a built-in model that estimates the muscle mass, but the results are not very reliable. This is the estimate that my work aims to improve.”
The other method used to measure body composition is radiology. This is more expensive and more complex to use, but the results are more exact.
Neural networks – an obvious tool
At the Sahlgrenska University Hospital, Alexander Karlsson gained access to a dataset with information from both BIA and radiology for different individuals, with about 5,700 observations. The dataset contained other information such as sex, age, height and weight.
“The thesis describes how I took the electrical signals from BIA and ran them through a statistical model that I had created. I try to convert the electrical signals from different parts of the body, together with sex, age, height and weight of each individual, to the estimates obtained from radiology.
To put it extremely simply, Alexander Karlsson has created an advanced equation in which age, weight, height and sex are among the variables. The equation converts values from BIA such that they approach the values from radiology.
“A large part of my thesis deals with processing the data in the right way. I spent a lot of time on understanding the data and how they can be used. Talking to my external supervisor at Sahlgrenska was a great help in this.”
The correlation between the values from BIA and radiology is highly complex.
“This is why neural networks are an obvious tool to use, since these models are very flexible. In brief, it’s based on sending signals through a network of weights and nodes. It’s not easy to interpret exactly how the signals are transformed within these complex networks and that’s actually not so important: what’s important is the result.”
Do your results have any practical applications?
“Oh yes! Instead of having all patients undergo both radiology and BIA, it would be possible to implement my algorithm in the clinic and patients would only need to use the BIA unit. We’re discussing how to proceed at the moment. It would be amazing to see my degree project in clinical use.”
More information about the Statistics and Machine Learning master’s programme at Linköping University.
Alexander Karlsson has been awarded the Christer Gilén Scholarship in statistics and machine learning for 2020
Alexander Karlsson has been awarded the Christer Gilén Scholarship in statistics and machine learning for 2020 for his master’s thesis “Improving predictions of muscle mass from an impedance device – Cross-calibration of bioelectrical impedance analysis and dual Xray absorbiometry using a Bayesian approach”.
The Christer Gilén Scholarship has been awarded since 2019 to master’s theses: one in statistics and machine learning, another in economic governance, organisation and innovation. Each scholarship has a value of SEK 10,000.
From the citation:
Alexander Karlsson has in a very commendable and independent manner written the thesis “Improving predictions of muscle mass from an impedance device – Cross-calibration of bioelectrical impedance analysis and dual Xray absorbiometry using a Bayesian approach”.[...]
The proposed methods within statistics and machine learning are useful innovations in a research field that aims to improve the determination of muscle mass from an impedance scale in medical care. The new methods give clearly better estimates than previous methods used in research.
Translated by George Farrants