Stroke is the most common cause of permanent disability worldwide. It has a significant impact on cognitive function, making a precise prediction of neurological outcomes critical for individualized rehabilitation and better patient care. However, the high complexity inherent in stroke data makes prediction difficult. Deep learning models can generate very accurate predictions by analyzing large patient datasets and are currently revolutionizing precision medicine. Nonetheless, the small number of patients in stroke datasets limits the applicability of many deep-learning architectures. The present project presents a novel strategy that exploits transfer learning from the large-scale Human Connectome Project (HCP) dataset used to train a two-dimensional convolutional autoencoder (CAE) for dimensionality reduction. The CAE significantly decreases the complexity of stroke patient data and it can be combined with regularized linear regression for prediction of neuropsychological scores. Gains in predictive accuracy with the deep learning pipeline across behavioral domains are assessed against conventional machine learning methods used as baseline models. The applicability of deep learning methods is likely to improve outcome prediction and help rehabilitation planning following stroke.
Stroke is the most common cause of permanent disability worldwide. It has a significant impact on cognitive function, making a precise prediction of neurological outcomes critical for individualized rehabilitation and better patient care. However, the high complexity inherent in stroke data makes prediction difficult. Deep learning models can generate very accurate predictions by analyzing large patient datasets and are currently revolutionizing precision medicine. Nonetheless, the small number of patients in stroke datasets limits the applicability of many deep-learning architectures. The present project presents a novel strategy that exploits transfer learning from the large-scale Human Connectome Project (HCP) dataset used to train a two-dimensional convolutional autoencoder (CAE) for dimensionality reduction. The CAE significantly decreases the complexity of stroke patient data and it can be combined with regularized linear regression for prediction of neuropsychological scores. Gains in predictive accuracy with the deep learning pipeline across behavioral domains are assessed against conventional machine learning methods used as baseline models. The applicability of deep learning methods is likely to improve outcome prediction and help rehabilitation planning following stroke.
A Deep Learning Framework for predicting neuropsychological scores from resting state functional connectivity in stroke patients
RANGRAZI ASL, ASAL
2023/2024
Abstract
Stroke is the most common cause of permanent disability worldwide. It has a significant impact on cognitive function, making a precise prediction of neurological outcomes critical for individualized rehabilitation and better patient care. However, the high complexity inherent in stroke data makes prediction difficult. Deep learning models can generate very accurate predictions by analyzing large patient datasets and are currently revolutionizing precision medicine. Nonetheless, the small number of patients in stroke datasets limits the applicability of many deep-learning architectures. The present project presents a novel strategy that exploits transfer learning from the large-scale Human Connectome Project (HCP) dataset used to train a two-dimensional convolutional autoencoder (CAE) for dimensionality reduction. The CAE significantly decreases the complexity of stroke patient data and it can be combined with regularized linear regression for prediction of neuropsychological scores. Gains in predictive accuracy with the deep learning pipeline across behavioral domains are assessed against conventional machine learning methods used as baseline models. The applicability of deep learning methods is likely to improve outcome prediction and help rehabilitation planning following stroke.File | Dimensione | Formato | |
---|---|---|---|
Rangraziasl_Asal.pdf
accesso riservato
Dimensione
30.1 MB
Formato
Adobe PDF
|
30.1 MB | Adobe PDF |
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/20.500.12608/66545