Week 5: Safaa Mouline

Now that I have the structural connectivities (SC) and functional connectivities (FC) processed for the MS dataset, this week consisted mainly of building different models with them to predict EDSS scores. The spearman correlation between predicted and true scores for a ridge regression model using is about 0.45 using SC. Using a logistic regression to classify these into mild (EDSS <= 2.5) and moderate (EDSS > 2.5), the accuracy was around 0.77. I hoped to create an even finer classification, with more EDSS buckets, but since low EDSS scores are overrepresented, even with undersampling, the lack of higher scores makes it hard for the models to generalize well. 

I also tried to create models of the female patients split by peri and post menopause (using their ages as an estimates since we do not have menopause data recorded) but at least in terms of predicting EDSS, these did not perform well. However, looking to future steps, I hope to do some feature importance analysis. For models that perform well, analyzing model weights reflects which features (brain ROIs or in the case of FC, brain networks) are most predictive of EDSS. Another future step that is slightly more involved is to fine-tune the krakencoder (a machine learning model that can translate between SC and FC) to predict EDSS scores to then use the estimated FCs from this for downstream models. 

Comments

Popular posts from this blog

Week 3: Eddie Wei - Multimodal Endoscope

Week One: Eddie Wei

Week 1- iqra