Hello, if anyone is applying for E9 regarding human biological data/neurocognition, please let me know. I would like to connect with someone who has a precision medicine tool.
@HUMANawg @MultiOmicsAWG @ALSDAawg
5 Likes
Hello, interesting coincidence. A few months ago, I mentioned that I needed EEG Data on space conditions because I developed some tools for neuroscience. One of them is so precise that, by modifying a few parameters, I diagnosed with 96% accuracy Schizophrenia, 96% Frontotemporal Dementia, 97% Alzheimer’s Disease, and 96% Parkinson’s Disease (with more data, I can easily increase the accuracy). This method is currently in the process of publication, but I already published the first version (slightly less accurate with 93% for schizophrenia):
I can also predict and classify cognitive abilities with nearly 100% accuracy in certain tests (I’m going to publish this in another journal after publish my pending article). If any of what I mention is useful, we could collaborate. On another note, I hadn’t shared the schizophrenia article yet because the publication I have pending, which contains Alzheimer’s Disease, I wanted to share specifically with the Alzheimer’s Disease and Brain Resilience group.
4 Likes
Feel free to check out my dissertation here: Mathematical Models of Behavioral and Neural Error Response for Applications in Psychosis Assessment
Not quite as high of accuracy as I’d like (in the 70s to 80s), but (1) I believe I could improve that in a follow-up study using more of the EEG channels than I did the first time around and (2) I made the classification problem more difficult, but I believe more ecologically valid, than is typical (which was made possible by data collection that I can’t take any credit for, so I’m not tooting my own horn here but rather am just grateful the original authors put in that upfront work).
- N=524 (followed over 10-20 years if I recall correctly), and diagnoses were established by multi-psychiatrist consensus at the 10-year mark. Many other studies tackling this kind of classification problem I think will struggle to generalize past their own samples, in part because of mislabeling (given how difficult it can be to establish an accurate diagnosis without spending obscene amounts of time and resources like this study did, which again, I sadly can take no credit for).
- I used only human-interpretable ML models (i.e., providing closed-form equations that, in addition to providing potentially interesting substantive insights, allow for the application of its results to a new, individual case without providing any proprietary data or retraining),
- It distinguishes schizophrenia from bipolar and first-degree relatives of people with psychosis, which is a difficult classification problem, but one that is super relevant clinically. (Typically, the problem isn’t differentiating someone who’s totally healthy and low risk from someone who isn’t, but rather from high risk people or people with similar but distinct disorders.)
- Some of my models were able to achieve adequate accuracy even just using a simple cognitive task (without the accompanying EEG data) that could be administered cheaply and easily in just about any setting. In future studies, it would be cool to try to leverage the EEG data to create a really accurate model that could then potentially be applied to situations where you only have the cognitive data.
Also, @angel, that is a really cool approach! Maybe we could reach out to the original authors, who collected the data used in my dissertation to see if they’d be okay with us reapplying such an approach in their data. I wanted to use information from the whole brain (or rather, all parts covered by the EEG) rather than the limited set of channels I ended up using. (A good dissertation is a done dissertation!)
4 Likes
(edited to add to my original post for clarity)
1 Like
It would be amazing to work together, studying 524 individuals would be extraordinary. When I developed the schizophrenia method, I tried to keep the logic as minimal as possible so that when I had access to large datasets, I could still work efficiently. However, I also found hundreds of relevant patterns that are easily analyzable by humans, the problem is precisely generalization in this kind of patterns. In this case, the tool I developed has a very rigorous validation process (excluding subject-by-subject) and scaling features that allow us to study the brain in great detail. I used Parkinson’s Disease, Alzheimer’s Disease, Schizophrenia, and Frontotemporal Dementia, but I can diagnose essentially any pathology or “mental state.” It’s a highly generalizable method with very low computational cost.