Machine learning can yield "proxy measures" for brain-related health issues

Machine learning can yield "proxy measures" for brain-related health issues, say researchers

(Medical Xpress)—A team of Canadian researchers is offering a sneak peek at a novel technique of "proxy measures" for brain-related health issues. In their paper, published in the journal Nature Methods, the authors explain that they applied "novel machine learning approaches" to health-related data collected for the Canadian population. The results of a large number of different diseases and outcomes were analyzed in order to identify trends, provide predictive accuracy, and inform future study.

As an example, the team analyzed a series of different health-related variables to predict cardiovascular risk. While the research team was not able to provide a definitive answer on the question of which health factors are best for cardiovascular risk, they claim that the findings do identify some of the top contenders for future study. This is important, as they state that there are no one-size-fits-all methods for cardiovascular risk prediction. "There are many different tools that are used in different research labs and institutions," Tom R. H. Chang is associate professor at the University of Waterloo and leads the team of researchers. "Having no unified approach makes it difficult to compare research and leads to differences which can confound risk assessment."

In addition, he says, having standard methods to accurately predict "outcomes that matter to the public" can help policymakers formulate policies and prioritize spending on health issues. As a demonstration of the method's usefulness, the team cites the case of the "Healthy Nation" initiatives in Canada from the 1960s, which aimed to monitor and compare the health of Canadians. With that knowledge, doctors and policymakers were able to detect any trends in the population—and were able to change their public policy based on these trends to improve population health.

The problem with that approach, though, is that most of these older initiatives were "hard sciences" projects that were not always readily available to the public or easily implemented. "People like the fact that our method is providing a metric that is practical to measure and could be implemented in a clinical setting," says Chang. "There is more and more science being done in 'big data' or clinical domains but it isn't necessarily readily usable to the average citizen. This is one way of providing a standardized measure that is easy to comprehend and is widely applicable."

The methods also have a number of advantages over the "traditional" methods of risk assessment, which are based mostly on the correlation between clinical and biochemical factors and disease development. The classic approach, for example, relies mainly on the fact that some heart conditions such as heart disease are more likely to be seen in people who smoke—as certain chemicals in cigarettes have been demonstrated to damage heart tissue over time. "We've identified some of these biochemical correlates in the past that are risk factors," Chang says. "That is the classic way of risk assessment and we show that that is also predictive for other diseases. What we do is provide an alternative approach."

Chang says that the technique has potential uses far beyond cardiovascular risk assessment—and could be applied to a variety of health-related conditions, including cancer, diabetes, kidney disease, sleep disorders, and even mental health disorders such as depression. The Canadian researchers were only able to analyze data from approximately 10% of the population, as the Canadian census does not collect a large amount of health data for specific purposes. So, for all of those applications outside of cardiovascular risk, the technique could be adapted to collect health data for a much larger proportion of the population. The researchers note that they chose to focus on the Canadian population, because they could provide the most robust information on the topic. But the goal is to apply the method to the entire international population, with the hope that the results will benefit researchers in other countries as well.

"If we put this into clinical practice," says Chang, "we could start to see what factors have strong prognostic power and this, in turn, would then give us a lot more ideas for research. We could look at different clinical trials, genetic approaches, and lifestyle approaches to potentially intervene in different disease processes. Or we could just start to make changes right away."

The technique has "a lot of potential" as well in terms of public health, says Chang. By using the data to identify emerging problems and trends, health care professionals could start to offer preventative tests and medical treatments before they worsen. For example, if a study has identified an increase in the incidence of cardiovascular disease among younger people, then health care professionals would be alerted to start screening younger people for risk factors and offer preventative treatment.

Chang adds that there could also be advantages for people with mental illness or depression, as treatment techniques would be adjusted, possibly leading to fewer hospitalizations. So while the technique may be useful for identifying emerging problems in populations, it could also be of use in treating those problems.

This study is one of few in which clinicians have used a "proxy measure" to analyze "big data" in order to identify important health trends. However, there have been a number of previous studies that have done this kind of "proxy measure" research in "normal science" domains, like computer science, says Chang. For example, researchers have used data from social networks like Facebook, web searches, and cell phone records in order to track consumer trends and changes in technology usage.

One key difference between those studies is that while researchers are often comparing their methods to the "hard sciences" or to traditional methods for risk assessment, they're doing it in order to use the data in different ways. In the research presented in this study, the team says they're taking a different approach, as they're using "novel machine learning approaches" that they can apply to the data that is collected for different medical conditions. For example, the researchers say that they plan on using other studies into social media usage by applying the technique to those data.

Chang adds that the results of the approach they're exploring could potentially benefit scientists and clinicians in more ways than just being "good measures" for specific health conditions. The team is aiming to not only apply the technique to data collected for specific conditions, but to apply it to a variety of "health" disciplines, such as sports, sleep, nutrition, and exercise. "We are showing this approach is scalable from a scientific application to clinical application," says Chang.

About the author

Kara Bergus is a science writer for TechRepublic. She has a background in physics, biomedical engineering, and applied mathematics. Kara's interests include the future of the biomedical space and the use of emerging technology in the clinical space for patient care. Follow her on Google+ or Twitter.

Full Bio

Kara Bergus is a science writer for TechRepublic. She has a background in physics, biomedical engineering, and applied mathematics. Kara's interests include the future of the biomedical space and the use of emerging technology in the clinical space for patient care. Follow her on G...

Displaying 1 comment

It could be a very useful tool in the future. If I could add some of my thoughts about this paper,

1. The data was collected for 2 primary purposes. One was to help in the future of medicine by having some insight into the future health of the nation.

2. The other purpose was to help in improving medical research by helping to understand what conditions were the most effective at showing up in which age groups.

3. I believe that this "proxying measure" is a good measure for both of the above. But what this study shows is that there are issues that come with the measurement of the data. In the paper it shows the problems with accuracy. It could also be used for other measurements if you look at it like that.

4. By using the data the way it was intended and as it is known as by medical scientists and research scientists. This could help to be able to improve care on a personal level. Which then will lead to better medical studies, which leads to better outcomes. Which would only improve the world in the long run...

I actually agree with the title of this "article" in one respect.

That is "Machine learning can yield "proxy measures" for brain-related health issues".

The first thing that came out of my head after seeing the title of the article was the idea that of finding measures in proxies of mental health and illnesses.

When one sees the title of the paper they think that they are going to learn about something new about the brain, but really it seems that it is more about having a measurement system for mental health conditions or illnesses.

You can go and look up "proxy measures" and get lots and lots of information. But the one thing that almost all of them have in common is that "proxy measures" are not directly to the cause of the problems.

For example in the article I mentioned in this article, mental health has to do with our emotions. So it is interesting that proxy measures are being developed for that.

But the thing about proxy measures is that they are not measures of the cause.

So when you get lots of information on what the proxy measures are and what the problems are being measured for, how are they coming up with the proxy measure, then you get something like this.

I get things by searching in terms like, "proxy measures of brain function", "proxy measures of brain health", "proxy measures of brain problems". I get lots of information, and in all of the information I am seeing you learn about what the brain is doing. So you are getting data about problems on the brain, but no solution.

So I came upon "proxy measures" in the title of the paper and I was like "That is exactly what he is talking about".

It is also something that could be taken to the next level.