MDMNet: Multi-dimensional multi-modal network to identify organ system limitation in cardiopulmonary exercise testing

Area of Science:

• Cardiology
• Physiology
• Biomedical Engineering

Background:

• Cardiopulmonary exercise testing (CPET) is crucial for assessing cardiorespiratory fitness.
• Identifying functional limitations from CPET data is challenging due to complex variable associations and inter-individual variability.

Purpose of the Study:

• To present a novel multi-dimensional multi-modal network (MDMNet) for identifying functional limitations of organ systems using CPET data.
• To address the complexities of intra-variable associations and inter-individual variability in CPET analysis.

Main Methods:

• Developed a dedicated embedding strategy to map raw CPET inputs into a learned feature space.
• Implemented a multi-dimensional feature extraction module with 1D (temporal/spatial) and 2D (Gramian Angular Field encoding) branches.
• Integrated extracted features with demographic information for multi-dimensional, multi-modal learning in the MDMNet.

Main Results:

• The MDMNet achieved high Area Under the Curve (AUC) scores: 0.948, 0.949, and 0.931 across three distinct tasks.
• Demonstrated superior performance in discerning inter-individual differences compared to existing methods.

Conclusions:

• The proposed MDMNet effectively identifies functional limitations from CPET data.
• The model shows significant potential for automated clinical applications of CPET analysis.