BrainAGE is a brain aging estimation approach based on Machine Learning models trained on high-dimensional neuroimaging data. This method is used to capture deviations between chronological and biological aging and can serve as a biomarker for monitoring health and detecting neurodegenerative conditions. Despite progress, current approaches exhibit a Mean Absolute Error (MAE) of approximately five years, limiting clinical applicability. The present work examines 25 workflows, generated by combining 11 ML algorithms with different feature sets extracted from T1-weighted MRI scans, with the goal of identifying the most effective estimation method. Models were trained on the Cam-CAN dataset, including 627 subjects, and validated against the HCP-Aging dataset, including 607 subjects, to test for scalability on unseen data. The SVM model with RBF kernel achieved the best performance on the Cam-CAN dataset (MAE = 5.89 years) but failed to generalize effectively to the HCP-Aging dataset. Overall, considering both the datasets used for training and validation, the XGB model emerged as the best performer, achieving a MAE of 7.45 years. Our results are in line with previous studies employing features derived from T1-weighted MRI scans for BrainAGE prediction and highlight the challenges of model transferability across datasets, underscoring the need of robust frameworks to ensure reliable brain age predictions.
BrainAGE è un approccio per la stima dell’invecchiamento cerebrale basato su modelli di Machine Learning addestrati su dati neuroimaging ad alta dimensionalità. Questo metodo viene utilizzato per catturare le deviazioni tra invecchiamento cronologico e biologico e può servire come biomarcatore per monitorare la salute e rilevare condizioni neurodegenerative. Nonostante i progressi, gli approcci attuali presentano un Errore Medio Assoluto (MAE) di circa cinque anni, limitando l’applicabilità clinica. Questo lavoro analizza 25 workflow, generati combinando 11 algoritmi di Machine Learning con diversi set di caratteristiche estratte da scansioni MRI pesate in T1, con l’obiettivo di identificare il metodo di stima più efficace. I modelli sono stati addestrati sul dataset Cam-CAN,che include 627soggetti, e validati sul dataset HCP-Aging, che comprende 607 soggetti, per testare la scalabilità su dati non visti durante l’addestramento. Il modello SVM con kernel RBF ha ottenuto le migliori prestazioni sul dataset Cam-CAN (MAE = 5.89 anni), ma non è riuscito a generalizzare efficacemente al dataset HCP-Aging. Complessivamente, considerando entrambi i dataset utilizzati per l’addestramento e la validazione, il modello XGB si è distinto per la performance migliore, raggiungendo un MAE di 7.45 anni. I nostri risultati sono in linea con studi precedenti che utilizzano caratteristiche derivate da scansioni MRI pesate in T1 per predizioni con BrainAGE e evidenziano le sfide legate alla trasferibilità dei modelli tra dataset, sottolineando la necessità di framework robusti per garantire previsioni affidabili sull’età cerebrale.
Comparison of Statistical Methods for Brain Age Prediction Using Neuroimaging Data
SAMMASSIMO, VALENTINA
2023/2024
Abstract
BrainAGE is a brain aging estimation approach based on Machine Learning models trained on high-dimensional neuroimaging data. This method is used to capture deviations between chronological and biological aging and can serve as a biomarker for monitoring health and detecting neurodegenerative conditions. Despite progress, current approaches exhibit a Mean Absolute Error (MAE) of approximately five years, limiting clinical applicability. The present work examines 25 workflows, generated by combining 11 ML algorithms with different feature sets extracted from T1-weighted MRI scans, with the goal of identifying the most effective estimation method. Models were trained on the Cam-CAN dataset, including 627 subjects, and validated against the HCP-Aging dataset, including 607 subjects, to test for scalability on unseen data. The SVM model with RBF kernel achieved the best performance on the Cam-CAN dataset (MAE = 5.89 years) but failed to generalize effectively to the HCP-Aging dataset. Overall, considering both the datasets used for training and validation, the XGB model emerged as the best performer, achieving a MAE of 7.45 years. Our results are in line with previous studies employing features derived from T1-weighted MRI scans for BrainAGE prediction and highlight the challenges of model transferability across datasets, underscoring the need of robust frameworks to ensure reliable brain age predictions.File | Dimensione | Formato | |
---|---|---|---|
Sammassimo_Valentina.pdf
accesso aperto
Dimensione
7.22 MB
Formato
Adobe PDF
|
7.22 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/77856