Uncovering potential biomarkers for dry eye disease with machine learning
Published:
Researchers from SimulaMet are using machine learning to identify potential biomarkers for dry eye disease, hoping to increase the understanding of the disease. The latest study from the project shows promising results.
Dry eye disease affects tens of millions of people worldwide. The most common type of dry eye disease is called Meibomian Gland Dysfunction (MGD). This is a potentially painful condition that is diagnosed through a combination of several clinical tests and where existing treatment options are insufficient for some patients.
Tears, and more specifically their protein composition, provide information about the conditions on the eye’s surface. Some specific proteins can play an important role in the MGD condition.
Identifying these proteins can give a deeper understanding of the disease. Researchers from SimulaMet, in collaboration with eye doctors and other medical experts, have developed a machine learning framework that does exactly this.
Identifying the relevant proteins
In the most recent study from the project, they found that several of the proteins detected by the framework have been associated with MGD in previous studies, indicating that they are detecting clinically relevant proteins. They also discovered some proteins that are less explored in regards to the disease.
“These undiscovered proteins would be especially interesting to look further into, with the aim to increase our understanding of the disease and develop improved treatment methods.”, says PhD student Andrea Storås, first author of the article revealing these results.
In the proposed approach, machine learning models are first trained to grade the severity of the disease based on the amounts of tear proteins. Next, explainable artificial intelligence methods are applied to investigate which proteins the models regarded as most important when grading the disease.
The importance of explainable AI in the medical field
Even though machine learning models show impressive results in solving challenging tasks like this, a major concern is the lack of interpretability. This is where the explainable AI methods come in, as a tool for explaining the models and their predictions.
“It is very important, in the medical field especially, that users understand how machine learning systems analyse the data and come to their conclusions”, says Andrea.
Machine learning models can learn spurious patterns in a dataset, and propagate existing biases in the data. There is a risk that models show discriminating behaviour or severely underperforms on some groups in the society.
“Understanding how the models interpret the data can also lead to new knowledge because the model might identify patterns that humans have overlooked.”
What's next?
In addition to looking further into the undiscovered proteins, future work will include data from a larger population and clinical studies.
“Through clinical studies we can test whether or not the identified proteins can serve as targets for treatments.”
One example is to develop a treatment that regulates the amount of these proteins in tears, either directly or indirectly, by affecting signaling pathways that involve the proteins. This could possibly, over time, have an effect on the disease.
There is also a potential of applying this approach further to other relevant medical conditions.
Andrea brings up other relevant use cases, such as inspecting protein and metabolite patterns in patients with various stages of cancer. Or increase the understanding of serious mental disorders like schizophrenia and bipolar disorder, where earlier research has already shown that some metabolites differ in concentration for these patients (cambridge.org).
