Prediction of East Asian Brain Age using Machine Learning Algorithms Trained With Community-based Healthy Brain MRI

Aging in the East Asian population is associated with health conditions that significantly challenge individuals and communities.1 On an individual level, aging is related to a progressive decline in physical health and increasing risk of neurodegenerative diseases; on a community level, the aging population is connected to more significant socioeconomic costs.2, 3 Therefore, advancing our understanding of brain aging is vital to help improve the detection of early-stage neurodegeneration and predict age-related cognitive decline. There are currently many efforts aiming to achieve the early detection of age-related diseases with the goal of preventing or delaying their progression. Brain magnetic resonance imaging (MRI) is a reliable method with which to evaluate brain development due to its high soft-tissue resolution, multi-parameter imaging advantages, and non-invasive nature.4

There has recently been significant research interest shown in measuring brain change through brain age prediction using machine learning methods, mostly based on structural T1-weighted MRI images.5, 6, 7 Most machine learning models learn patterns from data, then use these patterns to make predictions about new data. A vital advantage of these methods over traditional statistics is that they make it possible to capture differences at an individual level rather than solely at a group level, thereby increasing the potential for clinical translation.8 Brain age prediction studies commonly build regression machine learning models using structural MRI data from healthy controls.6 To minimize the prediction error of the model and improve its accuracy, different regression-supervised machine-learning algorithms such as linear regression (LR), support vector regression (SVR), gradient boosting regressor (GBR), and BayesianRidge (BR) have been utilized.

These algorithms are the conventional approaches used to develop a predictive model because of their simple development process, high interpretability, and improved stability.9 The LR, SVR, GBR, and BR are widely used in biological research since they can handle unclean and noisy data well, support different loss function, and have strong predictive ability for nonlinear data.10, 11, 12, 13

The primary outcome measure in brain age prediction is the difference between an individual's predicted brain age and chronological age, which is called the 'brain-age gap.’ Suppose that one’s predicted brain age is lower than their chronological age. That person might then be considered to have a “younger” brain than anticipated, be more pathology-resistant, or have accumulated less pathology. Meanwhile, if the predicted brain age is significantly higher than the individual’s chronological age, that person may have an “older” brain than is typical, and may be genetically predisposed to certain pathologies.14

Previous studies have reported high correlations between brain age predictions and chronological age in healthy subjects. Wang et al.15 examined the T1-MRI dataset of 3,688 dementia-free participants with a mean age of 66 years on a convolutional neural network (CNN) deep learning algorithm to predict brain age. Logistic regressions were conducted to evaluate the model performance in assessing the association of the age gap with incident dementia, and the mean absolute error (MAE) of 4.45 years was obtained. Paul et al.16 designed a model with an MAE of 3.60 years. That model was trained with 1,597 T1-MRI on a CNN algorithm. Aycheh et al.17 collected MR images of 2,911 cognitively normal subjects (ages 45–91) and then executed a non-LR gaussian algorithm to fit the final age prediction model. The experimental results indicated an MAE of 4.05 years. Other studies have also shown a relationship between brain age prediction with chronological age based on MRI raw data trained with machine learning.6, 7

In this paper, we build machine learning models to predict chronological age in the East Asian population using brain volume data (subcortical segmentation). Unlike other studies, we used four different machine learning algorithms LR, SVR, GBR, and BR to identify the most suitable algorithm for our dataset, and we evaluated the feature importance of the brain volume data during model training (Table 1).18

Table 1
The 20 features from 55 labels (Supplementary Fig. 1) included in the MRI image’s brain volume data

• Click for larger image
• Click for full table
• Download as Excel file

In total, 154 participants were allocated to the training and test datasets. The test dataset had 47 participants, while the training dataset had 107 participants. The results showed that the prediction accuracy of the model trained with the BR algorithm was better than those of the other regression models; it had an MAE score of 3.59 and a RMSE of 4.28 years (Fig. 2D). The respective MAE and RMSE values were 3.51 and 4.56 years for the GBR (Fig. 2C), 3.59 and 4.58 years for SVR (Fig. 2B), and 3.59 and 4.62 years for the LR (Fig. 2A). Regression graphs were plotted for the models based on the testing dataset to show the correlation between the real age of the participants and the predicted age (Fig. 2). The detailed results of the models are presented in Table 3.

Fig. 2
Comparison between the real and predicted ages according to four models based on the test dataset of the participants.
MAE: mean absolute error, MSE: mean squared error, RMSE: root mean squared error, R²: R-squared.

• Click for larger image
• Download as PowerPoint slide

This study aimed to predict the brain age of healthy adults in the East Asian population using machine learning algorithms based on brain volume data from T1-weighted MRI scans. Machine learning studies have established a number of LR models to study the relationship between the predicted age and real age using brain imaging features.20

Nevertheless, the characteristics of the brain features used to train the models can lead to different estimates of MAE and R² in age prediction, with other studies having a wide range of ages. Lancaster et al.6 used a dataset of 2003 healthy individuals (MRI brain volume data) aged 16–90 with an SVR algorithm in a study aiming to predict chronological age. They achieved an MAE of 5.08 years in predicting chronological age. Hwang et al.7 investigated feasibility and clinical relevance of brain age prediction using axial T2-weighted images of healthy subjects with a deep CNN algorithm. The CNN model was trained with 1,530 scans, and the MAE evaluated the performance between the predicted age and the chronological age based on an internal and external test dataset. The model showed MAEs of 4.22 years in the internal test set and 9.96 years in the external test set.

In our study, we showed that the selection of type features is generally more important than the choice of model. The models exhibited excellent performance with only 19 features (Fig. 1) from 55 brain volume data (Supplementary Fig. 1). The RF classifier algorithm (https://docs.google.com/document/d/1uzvkHBgNQ1waeb1Ex94dLLzz5Ds_LVM4/edit?usp=sharing&ouid=113534775365702156330&rtpof=true&sd=true) selected the essential features from the sci-kit-learn library. The algorithm-selected features correlated well with our predicting target (age).

This study evaluated four LR models—LR, SVR, GBR, and BR—for age prediction (55–83 years). Regression is an approach to modeling the relationship between dependent and independent variables. This relationship is modeled using a linear predictor function whose unknown model parameters are estimated from the data.10 All the algorithms were trained using the same brain volume data. The results showed that the correlation between the predicted and chronological ages in the BR algorithm was strong (R² = 0.3). In addition, the MAE was 3.31 years. The MAE value was lower than those of the other three models, indicating that the BR was a more suitable model for age prediction. Most neuroimaging-based age predictions using machine-learning regression algorithms are proposed as a biomarker of brain aging that are related to cognitive performance, health outcomes, and progression of neurodegenerative disease.21, 22

A limitation of our study was that our models showed poor performance for other ages because the dataset did not include sufficient numbers of participants. The regression can explain this toward the mean phenomenon. It is desirable to match training and testing data age distribution during model training. If enough participants were included in the training dataset, this could have improved the performance in other age predictions.

Unlike other studies, the present study aimed to evaluate different regression algorithms and predict brain age in the East Asian population based on subcortical brain segments extracted from T1-weighted MRI scans of healthy participants. Our results showed that the BR algorithm achieved better accuracy than other methods (Table 3). We also proved that 19 features (Fig. 1) played a vital role in predicting brain age during the model training using machine-learning regression algorithms.

In this study, we trained and tested machine learning regression algorithms LR, SVR, GBR, and BR with brain volume data from MRI scans of community-based healthy subjects. We found that the BR algorithm obtained better results than LR, SVR, and GBR. In particular, BR achieved an MAE and R² of 3 and 0.3, respectively. Secondly, our results showed that simple volumetric features, such as the right pallidum, white matter hypointensities, and left hippocampus, were important in predicting chronological age when comparing the left cerebral white matter and right white matter hypointensities on T1-MRI scans. The brain age gap between predicted and chronological ages can reduce the dimensionality of neuroimaging data to provide meaningful biomarkers for individual brain aging.

Supplementary Fig. 1

All 55 brain volume data features. The random forest classifier algorithm selected the essential features from the sci-kit-learn library.