Advanced Machine Learning for Brain Age Gap Estimation in Anorexia Nervosa: A Neuroimaging Approach
Poster No:
504
Submission Type:
Abstract Submission
Authors:
Yubraj Gupta1, Andy Schumann2, Feliberto de la Cruz1, Karl- Jürgen Bär2
Institutions:
1Jena University Hospital, Jena, Germany, 2Klinikum Universität, Jena, Thuringia
First Author:
Co-Author(s):
Introduction:
Anorexia Nervosa (AN) is a condition that primarily affects young women, leading to significant weight loss and cognitive deficits. Accurate and early detection of these neurobiological changes is vital. This study aims to develop a Brain Age Gap Estimation (BrainAGE) model using ML to investigate accelerated brain ageing in AN. BrainAGE, a proven biomarker in conditions like Alzheimer's, is utilized here to determine if AN is linked to faster brain ageing, thereby serving as an effective tool for assessing brain health in AN patients.
Methods:
We developed a BrainAGE model for AN leveraging structural magnetic resonance imaging (sMRI) from 2887 healthy female controls (HC, ages 18-40) and 44 female AN patients of equivalent age. HC data spanned eight global databases, ensuring diverse brain morphology representation, vital for model training and the discernment of AN-induced deviations (Figure 1(a)). AN data was sourced exclusively from Jena University Hospital (JUH). Each sMRI sample was processed via FreeSurfer v.7.3.2, utilizing 'recon-all' to automate morphological feature extraction. This workflow provided 378 features, including volumetric, surface area, and curvature measures from cortical/sub-cortical structures, guided by the Desikan-Killiany atlas.
To predict chronological age, we implemented Support Vector Regression (SVR) and Gaussian Process Regression (GPR), forming an N x P feature matrix (N = 2887, P = 378). We applied 10-fold cross-validation and PCA for dimensionality reduction, maintaining 99% feature variance, and identified 257 independent features. Hyperparameter tuning was conducted for both SVR and GPR models, optimizing epsilon, gamma, kernel, regularization parameters (SVR), and lengthscale, noise, and kernels (GPR). Data was partitioned into training, validation, and a 44-HC-JUH-DATA hold-out set. Models were trained on 2799 HC samples, validated, and tested on the hold-out and 44 AN samples. Statistical analysis included Pearson's correlation coefficient (PCC), MAE, MSE, and RMSE to assess prediction accuracy and infer potential brain ageing in AN.
To predict chronological age, we implemented Support Vector Regression (SVR) and Gaussian Process Regression (GPR), forming an N x P feature matrix (N = 2887, P = 378). We applied 10-fold cross-validation and PCA for dimensionality reduction, maintaining 99% feature variance, and identified 257 independent features. Hyperparameter tuning was conducted for both SVR and GPR models, optimizing epsilon, gamma, kernel, regularization parameters (SVR), and lengthscale, noise, and kernels (GPR). Data was partitioned into training, validation, and a 44-HC-JUH-DATA hold-out set. Models were trained on 2799 HC samples, validated, and tested on the hold-out and 44 AN samples. Statistical analysis included Pearson's correlation coefficient (PCC), MAE, MSE, and RMSE to assess prediction accuracy and infer potential brain ageing in AN.
Results:
In our study, SVR and GPR were leveraged to develop a BrainAGE model, demonstrating high efficacy with and without PCA dimensionality reduction for the prediction of age. Particularly, GPR showed superior performance. Figure 2(a) shows the obtained actual vs predicted BrainAGE prediction for validation, hold-out and patients.
According to Figure 1(b), the GPR model achieved notable correlation scores across all groups, affirming the significance of the full feature set for age prediction. Figure 2(b)'s boxplot displays prediction error distributions, with the validation and 44-HC-JUH-DATA hold-out sets showing tight interquartile ranges and minimal median errors, indicating precise age estimations. Conversely, the AN patient group exhibited increased median errors and variability, with significant differences from the hold-out set (p = 0.0254) and validation set (p = 0.0010), suggesting altered brain ageing in AN using GPR without PCA (Figure 2(b), Figure 1(c)).
Further, Figure 2(c)'s bar chart reveals that while the validation set achieved a strong correlation (0.84), indicating accurate age predictions, the AN patient group presented higher MSE (28.16) and lower correlation (0.42), implying potential brain structure differences in AN using GPR without PCA. This underlines that, despite the models' accuracy with healthy data, predictions for AN patients indicate a potential acceleration in brain ageing.
According to Figure 1(b), the GPR model achieved notable correlation scores across all groups, affirming the significance of the full feature set for age prediction. Figure 2(b)'s boxplot displays prediction error distributions, with the validation and 44-HC-JUH-DATA hold-out sets showing tight interquartile ranges and minimal median errors, indicating precise age estimations. Conversely, the AN patient group exhibited increased median errors and variability, with significant differences from the hold-out set (p = 0.0254) and validation set (p = 0.0010), suggesting altered brain ageing in AN using GPR without PCA (Figure 2(b), Figure 1(c)).
Further, Figure 2(c)'s bar chart reveals that while the validation set achieved a strong correlation (0.84), indicating accurate age predictions, the AN patient group presented higher MSE (28.16) and lower correlation (0.42), implying potential brain structure differences in AN using GPR without PCA. This underlines that, despite the models' accuracy with healthy data, predictions for AN patients indicate a potential acceleration in brain ageing.
Conclusions:
The models' ability to discern between healthy and atypical brain development in AN patients demonstrates the value of BrainAGE in neurodevelopmental assessment. The observed prediction errors in AN patients align with the hypothesis of accelerated brain ageing, establishing Brain Age Gap Estimation as a viable biomarker for monitoring AN. These findings may inform precision medicine strategies for AN, emphasizing the role of ML in psychiatric evaluation.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 1
Lifespan Development:
Early life, Adolescence, Aging
Modeling and Analysis Methods:
Bayesian Modeling 2
Classification and Predictive Modeling
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
Aging
Behavioral Therapy
Computational Neuroscience
Data analysis
Eating Disorders
Emotions
Machine Learning
MRI
Neurological
Psychiatric Disorders
1|2Indicates the priority used for review
Provide references using author date format
Cole, J.H., 2020. Multimodality neuroimaging brain-age in UK biobank: relationship to biomedical, lifestyle, and cognitive factors. Neurobiology of aging, 92, pp.34-42.
Franke, K., 2010. Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters. Neuroimage, 50(3), pp.883-892.
Griffiths-King, D., 2023. Predicting ‘Brainage’in late childhood to adolescence (6-17yrs) using structural MRI, morphometric similarity, and machine learning. Scientific Reports, 13(1), p.15591.
Koutsouleris, N., 2014. Accelerated brain aging in schizophrenia and beyond: a neuroanatomical marker of psychiatric disorders. Schizophrenia bulletin, 40(5), pp.1140-1153.
Lombardi, A., 2021. Brain age prediction with morphological features using deep neural networks: results from predictive analytic competition 2019. Frontiers in Psychiatry, 11, p.619629.