Developmental Brain Age

This repository contains the release of the model developed for predicting brain age in developmental samples, created in:

Citation: Drobinin, V., Van Gestel, H., Helmick, C. A., Schmidt, M. H., Bowen, C. V., & Uher, R. (2021). The developmental brain age is associated with adversity, depression, and functional outcomes among adolescents. Biological Psychiatry. Cognitive Neuroscience and Neuroimaging.

Data description

Figure 1. A. Brain age prediction scatterplot. The diagonal line represents perfect prediction accuracy. Dots represent individual age predictions from structural MRI. Dots above the line represent predicted age older than chronological age. Dots below the line represent a prediction of a younger appearing brain. B. Age and sex distributions of all datasets. C. Data splits representing training, validation and testing data. Typically developing control participants used for model training and validation. Trained model applied in independent at-risk sample (FORBOW).

Abbreviations: CV = Cross Validation, ABIDE = Autism Brain Imaging Data Exchange, CMI = Child Mind Institute: Healthy Brain Network, CORR = Consortium for Reliability and Reproducibility, NIH = The NIH MRI study of normal brain development, PING = Pediatric Imaging, Neurocognition, and Genetics, FORBOW = Families Overcoming Risks and Building Opportunities for Well-being

Brain age prediction

Figure 2. Brain age prediction scatterplot showing predicted brain age by actual chronological age. A. Predicted age compared to scan age in validation (20% holdout) set. B. Brain age prediction in testing sample of at-risk adolescents (FORBOW). Line of best fit in magenta with confidence interval in blue. Warmer colors indicate a positive brain age gap (older appearing brain); cooler colors denote a negative brain age gap (younger appearing brain).

Top neuroanatomical contributions

Figure 3. Top neuroanatomical structures involved in accurate age prediction. A. Variable importance of top 10 cortical volume features. Left: Stacked bar plot with right hemisphere, followed by left hemisphere importance. Right: Variable importance of all cortical regions visualized on 2D brain. B. Variable importance of top 10 subcortical volume features. Left: Stacked bar plot with right hemisphere followed by left hemisphere where applicable. Right: Variable importance of all subcortical regions visualized on 2D brain.

Abbreviations: bankssts = Banks of the Superior Temporal Sulcus, DC = Diencephalon. CC = Corpus Callosum

Model description and requirements

• The model requires FreeSurfer processed data, exported to tabulated HCP-wide style format. One individual per row, with brain features as separate columns. See feature_list.txt for feature list and naming convention.

• Machine learning was performed within the tidymodels framework in R, using xgboost as the engine. Model preparation and cross-validation is described in model-prep.R, model tuning and fit is described in train-fit-model.R. XGboost version 1.0.0.2 tested, installation process described below:

packageUrl <- "https://cran.r-project.org/src/contrib/Archive/xgboost/xgboost_1.0.0.2.tar.gz"
# You then install this version of the package using
install.packages(packageUrl, repos = NULL, type = 'source')

Making predictions on new data

1. Load your tabulated data compliant with feature_list naming scheme. See above.
2. Load the model xgboost_9to19_brain_age_mod.rds

library(here) # OS agnostic relative paths (relative to project dir)
library(tidyverse) # data wrangling tools and pipes
library(tidymodels) # machine learning metaverse
library(xgboost) # main engine, needs version 1.0.0.2, see above

xgb_mod <- readRDS(
  file = here::here(
    "model",
    "xgboost_9to19_brain_age_mod.rds"))

3. Predict brain age in your data

brain_age_df <-
  xgb_mod %>%
  predict(new_data = your_df) %>%
  mutate(
    # provide the chronological age at time of scan
    truth = your_df$scan_age
  ) %>%
  # compute the brain age gap by subtracting chronological age from prediction
  mutate(gap = .pred - truth)

# 2024: Note dependencies are likely quite out of date, consider a more native xgboost prediction
test_out <- predict(xgb_mod$fit, newdata = as.matrix(your_df))

4. Evaluate brain age prediction accuracy

# compute common performance metrics: mae, rsq, rmse

brain_age_df %>%
  metrics(truth = truth, estimate = .pred)

5. (Optional) For improved visualization, bias correct brain age prediction using slope and intercept determined from prior independent validation.

# bias prediction method, steps in comments,
# previously determined values hardcoded.
# xgb_bias_mod <- lm(.pred ~ truth, data = xgb_validate)

# extract intercept and slope
bias_intercept <- 6.41 # xgb_bias_mod$coefficients[["(Intercept)"]]
bias_slope <-  0.55 # xgb_bias_mod$coefficients[["truth"]]

# create bias correct data frame
brain_age_corrected_df <- brain_age_df %>%
  mutate(
    # corrected brain age prediction
    corrected_pred =  (.pred - bias_intercept) / bias_slope
  ) %>%
  # corrected corrected brain age gap
  mutate(corrected_gap = corrected_pred - truth)