A Tensorized Multitask Deep Learning Network for Progression Prediction of Alzheimer's Disease.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Current machine learning for Alzheimer's disease (AD) diagnosis often separates classification and cognitive score prediction.
• This separation overlooks the potential synergistic relationship between these tasks for enhanced diagnostic and prognostic capabilities.

Purpose of the Study:

• To develop a unified multitask deep learning framework for simultaneous Alzheimer's disease (AD) classification and longitudinal cognitive score prediction.
• To leverage multimodal data fusion, kernelization, and tensorization for improved accuracy in both diagnostic and predictive tasks.

Main Methods:

• A deep neural network (KTMnet) employing modality fusion, kernelization, and tensorization was developed.
• The model performs simultaneous multiclass classification and longitudinal regression within a unified multitask framework.
• Multimodality scenarios were investigated to exploit complementary features for predicting cognitive scores and classifying disease status from baseline data.

Main Results:

• The proposed KTMnet achieved an overall accuracy of 66.85 ± 3.77 for multiclass classification.
• Prediction of Mini-Mental State Examination (MMSE) scores showed an average RMSE of 2.32 ± 0.52 and a correlation of 0.71 ± 5.98.
• Results were benchmarked against state-of-the-art methods, demonstrating competitive performance.

Conclusions:

• The multitask framework effectively integrates classification and prediction, enhancing Alzheimer's disease (AD) diagnosis and prognosis.
• Optimizing hyperparameters for one task (classification or prediction) may not yield optimal results for the other, indicating a trade-off.
• This highlights the complexity of balancing diagnostic accuracy and predictive precision within a single model.