MorphologyFM：用于心电图与脉搏血氧波形形态感知表征学习的基础模型

Jul 16, 2026

Upload by Hazel

Foundation models have recently emerged as a powerful paradigm for learning transferable representations from large-scale biomedical data, yet existing approaches for physiological waveforms primarily optimize reconstruction or forecasting objectives that do not explicitly preserve clinically meaningful waveform morphology. Electrocardiograms (ECGs) and pulse oximetry (SpO2) waveforms embed cardiovascular and hemodynamic information in waveform shape, intervals, slopes, and beat-to-beat variability. MorphologyFM introduces a morphology-aware multimodal self-supervised pretraining framework that learns representations transferable to downstream prediction tasks without manual annotation, and improves performance across multiple benchmarks.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages18

Reading Time45 min

Abstract

1 Introduction

Motivation and Problem Statement

Proposed Approach

arXiv :2607 .09749v1 [ ee ss . SP] 3 Jul 2026 MorphologyFM: A Foundation Model for Morphology-Aware Representation Learning from ECG and Pulse Oximetry Waveforms Saiyang Feng 1 , Yuanyun Zhang 1 , Shi Li2 , 1 University of the Chinese Academy of Sciences 2 Columbia University Abstract Foundation models have recently emerged as a powerful paradigm for learning transferable representations from large-scale biomedical data, yet existing approaches for physiological waveforms primarily optimize reconstruction or forecasting objectives that do not explicitly preserve clinically meaningful waveform morphology. Electrocardiograms (ECGs) and pulse oximetry (SpO2 ) waveforms encode rich cardiovascular and hemodynamic information through their morphological structure, including changes in waveform shape, intervals, slopes, and beatto-beat variability that underlie many clinical diagnoses. In this work, we introduce MorphologyFM, a multimodal foundation model pretrained on paired ECG and SpO2 waveforms from the MIMIC critical care database using a morphology-aware self-supervised learning objective. MorphologyFM combines morphology-guided masking, cross-modal representation learning, and contrastive latent alignment to learn representations that capture clinically relevant physiological structure without requiring manual annotations. We evaluate MorphologyFM across multiple downstream prediction tasks, including arrhythmia classification, hypoxemia prediction, mortality prediction, and length-of-stay estimation, demonstrating consistent improvements over representative self-supervised learning methods, including Masked Autoencoders (MAE), contrastive learning, Barlow Twins, and Joint Embedding Predictive Architectures (JEPA) . Furthermore, we show that jointly modeling ECG and SpO2 waveforms produces more transferable representations than single-modality pretraining and that morphology-aware objectives scale effectively with increasing amounts of unlabeled physiological data. Our results establish waveform morphology as a powerful inductive bias for self-supervised physiological representation learning and introduce MorphologyFM as a generalpurpose foundation model for continuous physiological monitoring. 1 Introduction Foundation models have emerged as the dominant paradigm for large-scale representation learning. Across natural language processing, computer vision, and multimodal learning, self-supervised pretraining on massive datasets has enabled models to learn representations that generalize across diverse downstream tasks with minimal supervision [1, 2] . These advances have been driven by improvements in optimization, architectural design, and scalable learning frameworks [3–5] . As a result, modern foundation models increasingly serve as general-purpose learning systems capable of acquiring transferable representations directly from raw observations. Healthcare has rapidly adopted this paradigm. Foundation models have now been developed for electronic health records (EHRs), clinical narratives, medical imaging, physiological monitoring, biological sequences, and multimodal patient data [6–8] . These systems have demonstrated strong Preprint. Under review. performance across prediction, retrieval, phenotyping, coding, clinical reasoning, and decisionsupport tasks [9–14] . Through large-scale self-supervised learning, foundation models acquire representations that capture complex statistical structure within healthcare data and transfer effectively across diverse downstream applications. Among healthcare modalities, physiological waveforms provide a unique and information-rich representation of patient state. Electrocardiograms (ECGs) characterize the electrical activity of the heart, while pulse oximetry (SpO2 ) waveforms capture peripheral vascular dynamics through photoplethysmography. Unlike structured EHR data, physiological waveforms are continuous, highfrequency signals whose diagnostic value lies no

FAQs

MorphologyFM的核心目标是什么？

MorphologyFM面向连续生理波形表征学习，强调显式保留临床有意义的波形形态特征，而不仅是重建或预测。它通过形态感知的自监督目标学习可迁移表征。

MorphologyFM使用哪些模态数据进行预训练？

预训练使用来自MIMIC重症监护数据库的配对ECG与SpO2波形。模型在多模态条件下学习联合表示。

MorphologyFM在下游任务上验证了哪些效果？

文中在多类下游预测任务上评估，包括心律失常分类、低氧血症预测、死亡率预测以及住院时长估计。结果显示相较代表性的自监督方法有一致提升。

Jul 16, 2026

Upload by Hazel