MAL-Net: A Multi-Label Deep Learning Framework Integrating LSTM and Multi-Head Attention for Enhanced Classification...

Area of Science:

• Nephrology
• Artificial Intelligence
• Medical Informatics

Background:

• IgA nephropathy (IgAN) is a primary cause of kidney failure, marked by complex heterogeneity.
• Current classification methods struggle with IgAN's diverse data and overlapping symptoms.
• There is a need for advanced tools to accurately subtype IgAN for better patient management.

Purpose of the Study:

• To introduce MAL-Net, a deep learning framework for multi-label IgAN subtype classification.
• To leverage multidimensional clinical data, including sensor-based inputs, for improved IgAN subtyping.
• To address the challenges of data heterogeneity and class imbalance in IgAN classification.

Main Methods:

• Developed MAL-Net, integrating Long Short-Term Memory (LSTM) and Multi-Head Attention (MHA) networks.
• Utilized a memory network for feature extraction from clinical sensors and records.
• Trained and validated the model on data from 500 IgAN patients, including demographics, labs, and symptoms.

Main Results:

• MAL-Net achieved 91% accuracy and an AUC of 0.97, outperforming six baseline models.
• Multi-Head Attention significantly improved classification, especially for rare IgAN subtypes.
• The F1-score for the Ni-du subtype increased by 0.8, demonstrating effective class imbalance mitigation.

Conclusions:

• MAL-Net offers a robust solution for multi-label IgAN subtype classification.
• The framework effectively handles data heterogeneity, class imbalance, and feature interdependencies.
• Integrating clinical sensor data enhances IgAN subtype prediction for improved diagnosis and prognosis.