PULSE-LAB: A Multimodal Hybrid State-Space Model for Forecasting the Presence of Thoracic Pathologies from ECG Time Series and Laboratory Data

Ismael Villanueva-Miranda, Zifan Gu, Donghan Yang, Yang Xie, Guanghua Xiao

Keywords: Multimodal deep learning, Clinical forecasting, State-space models (SSM), Medical decision support, Explainable AI (XAI)

TL;DR: A Multimodal Hybrid State-Space Model for Forecasting the Presence of Thoracic Pathologies from ECG Time Series and Laboratory Data

Abstract: In acute care, a gap often exists between admission and confirmatory imaging. We address this gap by forecasting chest X-ray (CXR) findings hours in advance using admission-time biosignals alone. We present PULSE–LAB, a hybrid model that combines a Mamba state-space encoder for long-sequence ECG with an MLP over 50 routine laboratory tests. Using the public Symile–MIMIC dataset, we predict 14 thoracic findings on the first post-admission CXR and establish, to our knowledge, the first benchmark for this task. The model achieves a macro-AUROC of 0.62 and macro-recall of 0.63, with the strongest discrimination for acute conditions (e.g., Fracture, AUROC 0.81) and high recall for developing pulmonary processes (e.g., Pneumonia, Consolidation). Interpretability analyses reveal clinically coherent patterns: coagulation and injury markers drive fracture risk, while renal and oncotic markers relate to edema. Both the ordered lab panel and informative ECG leads contribute. These results indicate that admission-time signals capture early physiological trajectories that later manifest on imaging, suggesting utility for earlier risk stratification, imaging prioritization, and proactive care.

Submission Number: 110