Multimodal graph-based classification of esophageal motility disorders

Area of Science:

• Gastroenterology
• Medical Informatics
• Computational Biology

Background:

• Diagnosing esophageal motility disorders is challenging due to complex high-resolution impedance manometry (HRIM) data and variable clinical interpretation.
• Existing methods struggle with the nuances of HRIM data, necessitating advanced analytical techniques.

Purpose of the Study:

• To explore the feasibility of a multimodal machine learning (ML) classification approach for esophageal motility disorders.
• To integrate HRIM recordings with patient-specific information using graph-based modeling for improved diagnostic accuracy.

Main Methods:

• Analysis of HRIM recordings and patient data from 104 patients with esophageal motility disorders.
• Extraction of patient information from structured questionnaires and free-text notes using keyword detection and large language models.
• Representation of HRIM data as spatiotemporal graphs and application of graph neural networks (GNNs) for feature learning.
• Fusion of GNN-derived patient embeddings with clinical data for multi-class classification of swallow events.

Main Results:

• The multimodal approach, integrating patient data, outperformed models relying solely on HRIM features.
• Graph-based modeling demonstrated performance gains compared to vision-based classifiers.
• Ablation studies confirmed the complementary contributions of different data modalities and the graph-based approach.

Conclusions:

• Integrating patient-level data with graph-based HRIM signal representations shows promise for accurate esophageal motility disorder classification.
• Larger, more representative datasets are needed for further validation and to ensure generalizability of the findings.