Few-Label Multimodal Modeling of SNP Variants and ECG Phenotypes Using Large Language Models for Cardiovascular Risk Stratification

TL;DR

This work addresses cardiovascular risk stratification under limited labeled data by proposing a few-label multimodal framework that fuses SNP variants and ECG phenotypes using large language models (LLMs) and chain-of-thought prompting for interpretability. A three-stage training pipeline—pseudo-label generation, top-k cluster refinement, and few-label fine-tuning with LoRA—enables robust genotype–phenotype modeling in low-resource settings, while CoT prompts provide clinically meaningful rationales. Using a harmonized dataset of 8,856 participants with SNP and ECG data plus clinical labels, the approach evaluates across three tiers of label availability, demonstrating that multimodal embeddings can approach full-data performance with substantially fewer annotations, especially for larger models like DeepSeek 1.3B. The findings highlight the potential of scalable, explainable LLM-based frameworks to advance personalized cardiovascular care in data-scarce environments and motivate future work on larger models and additional modalities. Overall, the study demonstrates that parameter-efficient LLMs, tiered pseudo-labeling, and CoT reasoning together yield strong, interpretable predictions and robust generalization across diverse patient profiles.

Abstract

Cardiovascular disease (CVD) risk stratification remains a major challenge due to its multifactorial nature and limited availability of high-quality labeled datasets. While genomic and electrophysiological data such as SNP variants and ECG phenotypes are increasingly accessible, effectively integrating these modalities in low-label settings is non-trivial. This challenge arises from the scarcity of well-annotated multimodal datasets and the high dimensionality of biological signals, which limit the effectiveness of conventional supervised models. To address this, we present a few-label multimodal framework that leverages large language models (LLMs) to combine genetic and electrophysiological information for cardiovascular risk stratification. Our approach incorporates a pseudo-label refinement strategy to adaptively distill high-confidence labels from weakly supervised predictions, enabling robust model fine-tuning with only a small set of ground-truth annotations. To enhance the interpretability, we frame the task as a Chain of Thought (CoT) reasoning problem, prompting the model to produce clinically relevant rationales alongside predictions. Experimental results demonstrate that the integration of multimodal inputs, few-label supervision, and CoT reasoning improves robustness and generalizability across diverse patient profiles. Experimental results using multimodal SNP variants and ECG-derived features demonstrated comparable performance to models trained on the full dataset, underscoring the promise of LLM-based few-label multimodal modeling for advancing personalized cardiovascular care.

Figures (4)

• Figure 1: SNPs influence ECG intervals (PR, QRS, QT) across biological levels: (A) Chromatin—identify genes affected by coding or non-coding SNPs; (B) Cell—determine cell types expressing the gene; (C) Tissue—link gene expression to tissue electrophysiology and ECG phenotypes.
• Figure 2: Linking SNP Variants and ECG Phenotypes for Explainable CVD Prediction.
• Figure 3: Example Creation of a Prompt for participants across tiers. The text in bold indicates the important features that the LLMs will choose to focus on, and the conclusions allow LLMs to draw the reasoning and link it to the relevant pseudo-label.
• Figure 4: Epoch-wise training loss comparison accross three models.