Graph-spa: A Spatiotemporal Graph Neural Network based framework for ARDS prediction and interpretability.

Area of Science:

• Computational biology and medicine
• Artificial intelligence in healthcare
• Time-series analysis and deep learning

Background:

• Traditional deep learning models struggle with long-range temporal dependencies in multivariate time-series data, hindering early prediction of critical conditions like acute respiratory distress syndrome (ARDS).
• Existing methods lack the ability to dynamically model evolving interactions among clinical variables, which are crucial for timely and accurate event prediction.

Purpose of the Study:

• To introduce Graph-spa, a dynamic Spatiotemporal Graph Neural Network (STGNN) framework designed to improve early prediction of ARDS onset.
• To enhance model interpretability by identifying critical clinical features and their temporal interactions preceding ARDS.
• To provide a flexible and extensible framework for dynamic clinical event prediction in intensive care units (ICUs).

Main Methods:

• Developed Graph-spa, integrating temporal convolution with a dynamic STGNN that updates adjacency structures to capture complex temporal dependencies.
• Benchmarked Graph-spa against traditional deep learning models (GRU, LSTM, TCN, Transformer) and a baseline STGNN across three large clinical datasets (HiRID, MIMIC-IV, eICU).
• Employed mask-based interpretability methods for feature-time attribution and co-occurrence analysis to identify sustained feature activations preceding ARDS.

Main Results:

• Graph-spa consistently outperformed all baseline models in ARDS prediction across internal and external validations, achieving superior AUC F1-MCC scores.
• The dynamic adjacency mechanism in Graph-spa effectively captured evolving feature interactions, leading to more diversified connectivity patterns than the baseline.
• Interpretability analysis highlighted sustained potassium abnormalities and declining Glasgow Coma Scale scores as critical early indicators of ARDS.

Conclusions:

• Graph-spa represents a significant advancement in dynamic clinical event prediction, offering an end-to-end approach for early detection of organ failure.
• The framework's model-agnostic core modules (dynamic graph construction, attribution, co-occurrence mining) allow for broad applicability to various ICU dynamic classification and regression tasks.
• The study provides a valuable tool for early ARDS detection and demonstrates the potential for discovering sub-clinical signatures predictive of critical events.