Local-Global Memory Neural Network for Medication Prediction.

Area of Science:

• Computer Science
• Medical Informatics
• Artificial Intelligence

Background:

• Electronic medical records (EMRs) are crucial for medical data mining and sequential learning.
• Predicting future patient needs from EMRs is a significant challenge in healthcare.

Purpose of the Study:

• To propose a sequential neural network with dynamic content-based memories for predicting future medications using EMRs.
• To develop a model that learns hidden knowledge from EMRs through local and global memory layers.

Main Methods:

• A local-global memory neural network architecture was employed.
• The model utilizes local memory for individual patient patterns and global memory for group disease evidence.
• The model was adapted for unsupervised classification of EMR hidden states into medication progression phases.

Main Results:

• The proposed model demonstrated improved prediction performance for future medications compared to alternative methods.
• Experimental results on real EMR datasets validated the effectiveness of learning with external local and global memories.
• The unsupervised classification of EMR states into disease progression phases was also shown.

Conclusions:

• The local-global memory neural network effectively predicts future medications by leveraging comprehensive EMR data.
• This approach enhances medical data mining and sequential learning for improved patient care.
• The model's ability to identify disease progression phases offers new insights into treatment pathways.