Multimodal Attention Network for Dementia Prediction.

Area of Science:

• Computational neuroscience
• Medical informatics
• Public health

Background:

• Early dementia risk identification is vital for prevention and insurance in aging populations.
• Taiwan's National Health Insurance data presents challenges due to high-dimensional, sparse International Classification of Diseases (ICD) codes.

Purpose of the Study:

• To accurately predict future dementia incidence risk using neural networks.
• To address challenges posed by high-dimensional and sparse ICD code data.

Main Methods:

• Developed a multimodal attention network for dementia (MAND) inspired by click-through rate (CTR) prediction models.
• Incorporated an ICD code embedding layer and multihead self-attention to encode disease interactions.
• Investigated the applicability of various CTR methods for dementia risk prediction.

Main Results:

• MAND achieved an Area Under the Curve (AUC) of 0.9010, outperforming traditional CTR models.
• Attention score analysis identified key diseases correlated with dementia, aligning with existing research.
• The model demonstrated high flexibility and interpretability.

Conclusions:

• MAND offers an effective and accurate method for predicting dementia risk.
• The model's insights into disease correlations deepen understanding of dementia.
• Accurate prediction serves as an early warning system, supporting dementia prevention strategies.