Paper Type
short
Description
Alzheimer’s Disease (AD) is the sixth leading cause of death in America. Distinguishing the transitional stage - Mild Cognitive Impairment (MCI) - from the cognitive decline of normal aging is vital for early detection of AD. Various clinical measurements provide different and complementary views of AD patients and can be effectively combined through multimodal learning. However, previous multimodal methods (1) ignore modality-specific information that have predictive value and (2) cannot be generalized easily to more than two modalities or (3) pose practical limitations due to complex feature engineering. This paper proposes a novel deep multimodal model called MIMSL that addresses all these limitations and is the first to integrate three important modalities (MRI images, genetic markers and clinical notes) jointly for MCI prediction. Our preliminary experiments on a real-world dataset demonstrate the benefits of the architectural innovations in MIMSL and show that MIMSL outperforms state-of-the-art multimodal models for MCI prediction.
Recommended Citation
Qiu, Lin; Rajan, Vaibhav; and Tan, Bernard, "Battling Alzheimer’s Disease through Early Detection: A Deep Multimodal Learning Approach" (2019). ICIS 2019 Proceedings. 9.
https://aisel.aisnet.org/icis2019/data_science/data_science/9
Battling Alzheimer’s Disease through Early Detection: A Deep Multimodal Learning Approach
Alzheimer’s Disease (AD) is the sixth leading cause of death in America. Distinguishing the transitional stage - Mild Cognitive Impairment (MCI) - from the cognitive decline of normal aging is vital for early detection of AD. Various clinical measurements provide different and complementary views of AD patients and can be effectively combined through multimodal learning. However, previous multimodal methods (1) ignore modality-specific information that have predictive value and (2) cannot be generalized easily to more than two modalities or (3) pose practical limitations due to complex feature engineering. This paper proposes a novel deep multimodal model called MIMSL that addresses all these limitations and is the first to integrate three important modalities (MRI images, genetic markers and clinical notes) jointly for MCI prediction. Our preliminary experiments on a real-world dataset demonstrate the benefits of the architectural innovations in MIMSL and show that MIMSL outperforms state-of-the-art multimodal models for MCI prediction.