Research map / Architectures and efficiency / Attention and transformers

Extending Pretrained 10-Second ECG Foundation Models to Longer Horizons

NeurIPS 2026 · Architectures and efficiency › Attention and transformers · approach: State space and recurrent sequence models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Extends frozen 10-second ECG foundation models to variable-length 3-minute-plus recordings via a lightweight plug-in with global positional embeddings, soft prompts, and teacher-guided locality and dynamics-aware losses.

Problem

Pretrained ECG foundation models are snapshot models trained on fixed 10-second inputs, so they cannot natively aggregate information across long, variable-length recordings for clinical tasks like alarm reduction, ED disposition, triage, and AF detection.

Approach

Keep the 10-second backbone frozen and add (i) a two-level positional encoding (repeated local embeddings plus learnable global segment embeddings) with soft prompt tokens, and (ii) a teacher-guided rank-based soft contrastive loss and a stop-gradient cosine consistency loss over high-affinity segments, both using the frozen snapshot model's segment representations as supervision.

Key results

• On VTaC with CSFM-Base, the proposed method reaches 91.23 AUC at 180s versus 90.24 for the best baseline (VPT), and 81.11 overall average across tasks.
• Across four backbones and four tasks, the method reports the best overall average for CSFM-Tiny (79.98), CSFM-Base (81.11) and ECG-JEPA (78.04), while on MERL it is second-best (74.56) behind full fine-tuning (75.52).
• Ablation on VTaC with CSFM shows global positional encoding alone raises 180s AUC from 71.45 (repeat) to 88.92, and adding both local and dynamics losses reaches 89.50.

Surprising findings

• On MERL, full fine-tuning remains the strongest overall baseline, outperforming the proposed frozen-backbone method on overall average.
• Interpolating positional embeddings gives a strong 180s result on VTaC (88.18) but very poor 10s performance (60.78), showing sensitivity to input length.

Limitations

• Extending to arbitrary or substantially longer horizons under memory and compute constraints remains open.
• On MERL the proposed method does not beat full fine-tuning overall, and its gains depend on teacher representation quality.

Experimental setup

Datasets: VTaC MC-MED CPSC2021

Benchmarks: VTaC MC-MED (ED disposition) MC-MED (triage acuity) CPSC2021

Metrics: AUC accuracy

Baselines: Token Pooling Logit Pooling Full Fine-tuning Linear Probing Partial Tuning Bias Tuning Adapter VPT LSTM Tuning

Builds on: Visual Prompt Tuning CSFM ECG-JEPA MERL

Closest papers by meaning

• From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining (ICML 2025 · similarity 0.87)
• anyECG-chat: A Generalist ECG-MLLM for Flexible ECG Input and Multi-Task Understanding (AAAI 2026 · similarity 0.87)
• GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images (NeurIPS 2025 · similarity 0.86)
• Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model (ICLR 2025 · similarity 0.85)
• ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation (NeurIPS 2026 · similarity 0.85)
• H-Tuning: Toward Low-Cost and Efficient ECG-based Cardiovascular Disease Detection with Pre-Trained Models (ICML 2025 · similarity 0.85)