Project Update: Deep Learning to Extract Biomarkers for the Diagnosis and Personalized Treatment of Neuropsychiatric Disorders

Mental disorders have a significant impact on people's lives. Swift recovery depends on accurate diagnosis and treatment. However, the absence of objective criteria to determine a patient's specific psychiatric condition (and the most suitable treatment) poses a challenge. Can deep learning play a role in the diagnosis and treatment of neuropsychiatric disorders?

Early November, we discussed this with prof. dr. Guido Van Wingen. Guido is a professor of Neuroimaging in Psychiatry at the Faculty of Medicine at the University of Amsterdam (AMC-UvA). He conducts research using neuroimaging techniques to investigate the underlying mechanisms of psychiatric disorders, particularly in cases of depression and obsessive-compulsive disorder. He explores how biological and psychological factors related to psychiatric vulnerability influence the brain. Additionally, he conducts research on predicting successful treatment outcomes and how it normalizes brain function.

Background

This research, which began in 2019, is part of the Data2person call aimed at stimulating multidisciplinary research contributing to the development of effective, efficient, and responsible personal empowerment methods for a healthy society in the future. Guido collaborates on this project with Philips, a company that utilizes advanced technologies and deep insights into clinical applications and consumer needs to develop integrated solutions. Philips is a leading company in diagnostic imaging, image-guided treatments, medical IT applications, patient monitoring, home care systems, and consumer health applications. Selene Gallo completed her postdoctoral research last year, while Ahmed El-Gazzar is currently working on completing his thesis. The actual research phase is now completed.

MRI + AI

The goal of this research is to further develop advanced artificial intelligence techniques for analyzing brain scans and subsequently use these techniques for objective diagnostics and predicting the response to electroconvulsive therapy in patients with severe depression.

"Based on clinical data and questionnaires, we couldn't accurately predict whether a patient would benefit from a specific treatment. That's why we started using MRI images about ten years ago to see if we could predict a patient's response to a particular treatment. We began with electroconvulsive* therapy." Artificial intelligence has proven capable of doing what was previously challenging. Studies show that machine learning analysis of both functional and structural features (biomarkers) in an MRI scan of the brain can also have diagnostic, prognostic, and predictive value in psychiatry. MRI, especially when combined with AI, has shown promise in predicting the success of treatment at the individual level. "In this research, we wanted to explore if deep learning could make this even better."

Can deep learning make it even better?

Advancements in artificial intelligence are progressing rapidly. In the first few years of the research, the research team experimented with the latest techniques. However, the expected progress and added value of deep learning did not materialize initially. "Unfortunately, after several years of hard work, we had to conclude that the technical aspects were working, but it wasn't better than what we could achieve with classical machine learning techniques." At the end of the project, in 2021, the team suddenly made significant progress.

Was it coincidence? A new deep learning technique at the last moment

Surprisingly, things could indeed be improved further! Over time, better deep learning algorithms emerged. "At the eleventh hour of the project, we made significant progress due to developments in the field of deep learning. We stumbled upon a new technique, which was only developed in 2021, and it allowed us to utilize a massive amount of MRI datasets from the UK Biobank in conjunction with a new type of algorithm."

The UK Biobank is a large-scale biomedical database and research resource that contains extensive genetic and health information from half a million British participants. The database is regularly updated with additional data and is globally accessible to approved researchers conducting vital research on the most common and life-threatening diseases. It contributes significantly to the advancement of modern medicine and treatment, enabling various scientific discoveries that improve human health.

"When we could use this enormous dataset along with suitable deep learning techniques, we found that we could achieve the desired and hoped-for progress."

An 80% Accurate Predictive Model

The research ultimately yielded promising results as the right method was identified, and MRI was shown to be valuable in psychiatry. With the advent of machine learning analysis, the clinical application of MRI in psychiatry has come closer to reality. Recent studies demonstrate that machine learning analysis of both functional and structural features (biomarkers) in an MRI scan of the brain can also have diagnostic, prognostic, and predictive value in psychiatry. "We can now correctly predict whether a treatment will work for 8 out of 10 patients. We know that, on a group level, 30-40% of patients benefit from treatment, but you don't know who beforehand. This result is a significant step beyond the 50% prediction by a practitioner who doesn't know if a treatment will be effective, which is purely based on chance."

New research proposal

Guido mentions that they are still at the beginning of the research journey; this was phase 1. They intend to move forward with a new proposal (with a national consortium) to advance toward clinical applications and apply the model to predict treatment outcomes in cases of depression. To create the right models, it's crucial to gather a substantial amount of imaging data in psychiatry. "The current project was very technical, but now we are going to apply these techniques and map the situation in the Netherlands. We want to focus not only on electroconvulsive therapy but also on treatments involving antidepressants, psychotherapy, and transcranial magnetic stimulation." This marks another significant step in enabling personalized care in psychiatry and reducing the suffering of patients and healthcare costs.

* Electroconvulsive therapy (ECT), or with an older designation electroshock therapy or simply electroshock, is a treatment of patients in which an attempt is made to treat certain psychiatric disorders by means of inducing an epileptic seizure, provoked by a pulse of electricity through the head.

Links to publications:

• fMRI-S4: learning short- and long-range dynamic fMRI dependencies using 1D Convolutions and State Space Models - https://arxiv.org/abs/2208.04166
• Dynamic Adaptive Spatio-temporal Graph Convolution for fMRI Modelling - https://arxiv.org/abs/2109.12517
• A Hybrid 3DCNN and 3DC-LSTM based model for 4D Spatio-temporal fMRI data: An ABIDE Autism Classification study - https://arxiv.org/abs/2002.05981