Enhanced effective convolutional attention network with squeeze-and-excitation inception module for multi-label...

Area of Science:

• Artificial Intelligence in Healthcare
• Deep Learning for Medical Informatics
• Natural Language Processing in Clinical Settings

Background:

• Clinical Document Classification (CDC) is vital for managing vast medical data, improving patient care, research, and administrative tasks.
• Current deep learning models show promise but require enhanced accuracy for long clinical documents.
• The need for advanced AI solutions to effectively categorize complex medical texts is growing.

Purpose of the Study:

• To introduce the Enhanced Effective Convolutional Attention Network (EECAN) for improved automatic clinical document classification.
• To enhance feature representation and extraction from clinical documents using novel deep learning strategies.
• To address limitations in current methods for classifying lengthy and multi-label clinical texts.

Main Methods:

• Proposed the EECAN model, integrating a Squeeze-and-Excitation (SE) Inception module for adaptive feature recalibration.
• Introduced the Encoder and Attention-Based Clinical Document Classification (EAB-CDC) strategy within EECAN.
• Utilized sum-pooling and multi-layer attention mechanisms for discriminative feature extraction from clinical text representations.

Main Results:

• EECAN demonstrated superior performance compared to existing deep learning approaches on benchmark datasets (MIMIC-III, MIMIC-III-50).
• Achieved high AUC scores of 99.70% with sum-pooling and 99.80% with multi-layer attention.
• The model effectively transformed multi-label clinical texts' contexts without information loss.

Conclusions:

• The EECAN model offers a significant advancement in automated clinical document classification.
• Its high accuracy and efficiency suggest strong potential for integration into Electronic Health Record (EHR) systems.
• This approach can enhance healthcare decision-making support through improved medical information organization.