Multi-VentNet: Multimodal Fusion Framework for Predicting Need for Mechanical Ventilation

Invasive Mechanical Ventilation (MV) is a common and critical intervention for patients with respiratory failure in intensive care units (ICUs). Early identification of patients who will require MV is crucial to improve patient outcomes. Prior studies using single-modality models face challenges in fully capturing complex interrelationships between clinical data sources. To address these challenges, we propose Multi-VentNet, an integrative multimodal fusion approach that combines Electronic Health Records (EHR) and Chest X-rays (CXR) to predict the need for MV up to 24 hours in advance. Our approach consists of three key components: (1) Single-Modality Encoders for extracting modality-specific features from EHR and CXR data, (2) Cross-Modality Encoder to capture inter- and intra-modality relationships, and (3) Risk Integration Module that generates a final risk score by combining predictions from each modality. Experimental results demonstrate that Multi-VentNet achieved an area under the curve (AUC) of 0.8362, a specificity (SPC) of 79.63%, and a positive predictive value (PPV) of 26.41%, outperforming single-modality models by 0.0478 in AUC, 2.01% in SPC, and 1.51% in PPV. These results highlight the effectiveness of our multimodal framework in combining EHR and CXR data for more accurate and timely predictions of MV needs.