Contributions of lifestyle, education, and cardiovascular risk factors to the brain age gap

Figures

• 

  Fig. A.1. . Age distributions of participants for the training dataset (left) and age distribution after over/down sampling.

• 

  Fig. A.2. Predicted versus real age for the validation set for the case with lowest MAE. Validation set with 300 subjects and 803 images.

• 

  Fig. A.3. . Predicted age corrected versus real age for the test set (left) and Residuals corrected versus real (chronological) age (right) for the case with lowest MAE on the test set. Test set with 200 subjects and 555 images

• 

  Fig. A.4. . By fitting the data with linear regression we observe that for healthy participants the brain age gap is stable and close to zero (blue solid line). On the contrary, for participants who developed dementia the average value of brain age gap (dashed red line) is larger than the average value of healthy subjects (dashed blue line). Additionally, the brain age gap for people who develop dementia increases over time (red solid line). A Chow test returns an F value of 2.01 with a p-value of 0.002 indicative that the regression coefficients (intercepts and slopes) are significantly different between healthy and subjects who develop dementia during the study. This means the relationship between Age Gap and time of the study varies across the two datasets (healthy and dementia). 2605 Healthy Participants with 5259 images and 403 Dementia Participants with 638 images

• 

  Fig. A.5. . Effect of alcohol on brain age gap (left). 100 gr per day of alcohol is considered as the threshold for high consumption. Effect of diastolic pressure on the right (as threshold for high pressure is considered the value of 85) The assumptions of the multivariate linear regression for homoscedasticity and normality of the residuals is assessed by plotting the residuals over the fitted values and the density plot Figure Appendix A.6 and Figure Appendix A.7. Residuals are fitted with a straight line close to zero while the density plot is close to a bell shaped.

• 

  Fig. A.6. . Residuals over fitted value (left) and density plot of residuals (right) for linear mixed models regression of life habits and education.

• 

  Fig. A.7. . Residuals over fitted value (left) and density plot of residuals (right) for the linear mixed models regression of cardio-metabolic factors.

• 

  Fig. A.8. . Structure of CNN with six convolutional layers and 256 neurons at the output layer. Statistical tests, Appendix A1. A one-way multivariate analysis of variance (MANOVA) was conducted to examine whether there are differences between full sample (n = 11,627) and regression sample (n = 5,897) across seven independent continuous variables: alcohol gr per day, systolic blood pressure, diastolic blood pressure, glucose, cholesterol, sleeping hours per night, and age were jointly statistically significant.

• 

  Fig. A.9. . Flowchart showing the number of participants and the number of MRI brain images per each step.