At 31, Mariam Khayretdinova, CEO of Brainify AI, is already on her second startup on a quest to help female depression through technology. This one is an artificial intelligence, deep machine learning biomarker with a focus on female depression.
“This major depressive disorder is majorly under-studied, despite the fact that females are twice as likely to develop depression at some point in their lives when compared to men,” she explained. Marium Khayretdinova wants to solve that problem.
Interview with Mariam Khayretdinova
Frank LaVigne: The topic of depression is relevant to the post-pandemic period. But why have you chosen to focus on female depression topic specifically?
Mariam Khayretdinova: Depression itself is a very heterogeneous disease and certainly, there are many limitations in how it is being diagnosed and treated. This heterogeneity allows us to claim that there are certain depression subtypes. Female depression is majorly under-studied, despite the fact that females are twice as likely to develop depression at some point in their lives when compared to men.
There is more to that, females are going through a unique to females hormonal journey throughout their lives. We know that in some women first depression episode falls during the menarche or years following the menarche (first instance of their menstrual cycle) when the first hormonal changes occur. It can further be stratified into premenstrual dysphoric disorder, antenatal and post-partum depression, pre- and post-menopause depression, which are all connected to certain hormonal changes in the female body and brain.
FL: You mentioned that Brainify.AI is an AI/ML biomarker platform that is focused on raw EEG. Why are you focused on EEG data, and not MRI or PET?
MK: While we acknowledge that MRI and PET methods might be more precise when it comes to imaging techniques. The EEG method is fast, reliable, and provides a good time resolution, which helps to analyze the connectivity within and between brain networks. For the future applicability of our product, the EEG method fits the goal, as it is non-invasive and can be used on vulnerable populations.
FL: What methods do you use for your AI/ML model architectures?
MK: We use Deep Convolutional Neural Networks that have shown promising results in pattern recognition and computer vision applications. This is due to their ability to automatically extract significant spatiotemporal features that best represent the data from its raw form without preprocessing or human decisions for selecting these features.
DCNNs have been used to identify biomarkers and diagnose mental disorders using computer tomography and MRI images. They have been successfully used to solve tasks related to predicting mental diseases from resting-state EEG recordings and to predict the sex and age of the brain. We believe that deep learning is a promising technology for extracting information from a complex data source, such as human brain EEG, without the need for manual feature engineering.
FL: Nowadays, the issue with the data became more apparent. The ironic part is that there is a lot of data, however, not all of it is available to the public or the sample sizes are small in each dataset. How do you overcome this issue with your AI/ML models?
MK: This is a very good question and the amount of data is critical for achieving high performance on ML tasks. Increasing the size of the existing datasets in science can be a game changer for developing stable and reliable treatment prediction biomarkers. Our team has currently built Data Whitening Model, which works with different EEG devices to build large and improved datasets from multiple sources (with approximately 15 000 recordings).
By this, I mean that we have developed a model that can solve the current application and focuses on combining already existing advances (i.e., automatic EEG artifact rejection) and new tools (i.e., automatic data quality assessment for multi-site data collection and data harmonization), which are deemed essential for collecting new large-scale biomarker data from multiple collection sites.
FL: Could you tell us a bit more about the models that Brainify.AI have built or is currently building?
MK: Apart from Data Whitening model, my team and I are refining several AI/ML models that can be useful for inclusion criteria into the clinical trials, as well as might serve as potential EEG biomarkers for treatment response prediction. Speaking of inclusion criteria, we have recently built Placebo Non-responders model (with 72% accuracy) that can detect people who will likely respond to a placebo. We believe it would be a useful asset for the ongoing drug trials to clasterize the patients who would potentially lower the studied drug efficacy.
For the treatment prediction, currently, we have two models, which are Brain Age Prediction model and Brain Sex Phenotype model. Brain Age Prediction model predicts the age of the brain based on the EEG features. The mean error score in predicted and chronological age is called brain age delta, and might serve as a potential biomarker for a number of mental health and neurodegenerative disorders.
Currently, our team has managed to show the best result in mean error score across the EEG studies on brain age prediction, with a mean error score of 5.9 years. We have outperformed the previously conducted research by 13% and we plan on beating our own result in the nearest future. Brain Sex Phenotype model predicts the person’s brain sex based on the EEG features from just 4 seconds of recording.
Brain Sex Phenotype might serve as a potential biomarker specifically for hormonally-related depression in women, as we believe it might be rather sensitive to the hormonal receptors. The accuracy of our Brain Sex Phenotype model is 86%, and we think of future ways for us to utilize this model in transfer learning to predict the treatment response.
In addition to being the CEO of Brainify, Mariam Khayretdinova is an ALM in Neuroscience (Dean’s Award for Excellence, Harvard) and holds an M.Sc. in Applied Mathematic.
