Exploring the intersection of neuroscience
and machine learning

A conversation with Elisa Serra on deep-phenotyping depression and the future of clinical diagnosis

October 2023

SB: Hi Elisa, thank you so much for your availability at answering my questions and congratulations again for winning the ECNP Poster Award at the 36^th ECNP Congress 2023!
ES: Hi Silvia, thank you, I’m happy of being here.

Tell me a bit about yourself. What is your background and how did you get to work on this topic?
I got my BSc in Psychology and Neuroscience from the University of Glasgow and my MSc in Cognitive Neuroscience and Clinical Neuropsychology from the University of Padova. I developed my passion for coding and statistical analyses while working on my undergraduate thesis and decided to follow this path for my Master’s thesis, which helped me learn more about data-driven approaches in Computational Neuroscience. After graduation, I did a post-graduate internship in prof. Francesco Benedetti’s lab at San Raffaele Hospital in Milan where I carried out the work described in my ECNP abstract under his and Dr. Benedetta Vai’s supervision. I am really fascinated by how these data-driven approaches can be used to address clinically relevant questions, and Benedetti’s lab offered me the ideal environment to work on this topic.

Deep-phenotyping depression (both unipolar and bipolar) is extremely important and your work represents an interesting step forward in this regard. Would you like to tell me something about it?
By using machine learning algorithms, we tried to understand whether we could find specific features that could distinguish unipolar from bipolar depression in a data-driven manner. We looked at brain structural and cognitive features, both separately and combined. Interestingly, the only model that was able to differentiate individuals with unipolar from those with bipolar depression above chance level was the model including both structural and cognitive information, suggesting that multimodal information was more efficient in classifying patients with different conditions than just unimodal information.

What were the main challenges of this project?
Working with machine learning requires a lot of data resources and considerable computational capacity. Learning how to deal with these things was quite challenging but it also ended up being one of the most enjoyable parts of this project. Also, trying to explain my results in a balanced manner was really important to me. I didn’t want to paint them as the only solution to improve diagnosis. Instead, I wanted to emphasize their potential to complement already established clinical approaches.

Do you see any clinical translation of these findings into clinical practice?
The ultimate goal is to develop tools to support precise diagnosis in clinical practice, as part of a larger project funded by the Italian Ministry of Health. I think that before implementing this method in the clinical setting we should validate it on high quality data and large sample sizes, which are essential to train these algorithms effectively.

What are the next steps in your career?
I have just started a PhD at the University of Oxford in collaboration with the Center for Music in the Brain in Aarhus. My PhD project is a bit different from what I did for this poster: I am going to use a musical task to evoke different music-related memory systems while recording the brain activity of participants in a magnetoencephalography. The main goal is identifying unique neural signatures of early stages of dementia. I am going to use data from a wide range of participants, spanning from healthy young and older adults to individuals at different stages of cognitive decline including dementia and then use different computational approaches to evaluate the existence of different neural patterns indicative of early stages of dementia.

Well, that sounds super interesting. Best of luck with your PhD and thank you for your time!
Thank you!