Patient Clustering for Vital Organ Failure Using ICD Code With Graph Attention.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Severe organ failures (OF), including heart, respiratory, and kidney failure, are critical conditions with high mortality rates, frequently encountered in intensive care units.
• Understanding the underlying relationships and shared characteristics among different types of organ failure is crucial for improving patient outcomes and treatment strategies.

Purpose of the Study:

• To explore patient clustering for three severe organ failures (heart, respiratory, kidney) using graph neural networks and patient diagnosis history.
• To develop and evaluate a novel neural network-based pipeline for unsupervised organ failure patient clustering.

Main Methods:

• A neural network pipeline was developed, incorporating embeddings pre-trained using an International Classification of Diseases (ICD) ontology graph.
• An autoencoder-based deep clustering architecture, jointly trained with K-means loss, was employed for non-linear dimension reduction and patient clustering.
• The methodology was validated on the MIMIC-III dataset.

Main Results:

• The proposed pipeline demonstrated superior performance compared to other clustering models on the MIMIC-III dataset.
• Two distinct patient clusters were identified, exhibiting different comorbidity spectra potentially linked to disease severity.
• The clustering results were stable but did not strictly align with specific organ failure types, suggesting shared hidden diagnostic features.

Conclusions:

• The identified clusters highlight significant shared characteristics in the diagnosis of heart, respiratory, and kidney failures.
• These clusters can serve as indicators of potential complications and disease severity, facilitating personalized treatment approaches.
• This work pioneers an unsupervised approach from a biomedical engineering perspective for organ failure insights and provides pre-trained embeddings for future research.