ECG-R1: From Hardware to Clinical AI

• ECG-R1 is a designation applied to distinct ECG systems, including wearable ECG-on-chip, mobile real-time monitoring, and a reasoning-oriented multimodal LLM.
• The clinical MLLM leverages protocol-guided supervision, modality-decoupled architecture, and reinforcement learning with ECG evidence rewards to ensure diagnostic accuracy and reduce hallucinations.
• Empirical evaluations show ECG-R1 achieving up to 80.57% diagnostic fidelity and enhanced cross-modal agreement, demonstrating superior robustness compared to earlier models.

ECG-R1 is a designation that has been used in arXiv literature for more than one class of electrocardiographic system. Its primary contemporary meaning is a reasoning-oriented multimodal LLM for reliable ECG interpretation, introduced as a protocol-guided, modality-agnostic MLLM that emphasizes measurable waveform evidence, diagnostic logic, and robustness to missing modalities (Jin et al., 4 Feb 2026). Earlier work used the same designation for embedded and mobile monitoring systems, including a highly integrated wearable ECG-on-chip and a portable real-time mobile ECG monitoring platform (Deepu et al., 2014). The term therefore denotes not a single historical lineage, but a recurring label applied to distinct ECG technologies spanning mixed-signal hardware, telemonitoring software, and multimodal clinical AI.

1. Terminological scope and documented usages

In the literature represented here, ECG-R1 appears in at least three clearly differentiated forms: as a wearable mixed-signal acquisition chip, as a mobile real-time monitoring system, and as a reasoning MLLM for interpretation. These uses share an ECG focus but differ in objective, technical substrate, and failure modes.

Usage	Paper	Defining characteristics
Wearable ECG-on-chip	(Deepu et al., 2014)	Integrated analog front end, 12-bit SAR ADC, on-chip QRS detector, FIFO/SRAM, SPI
Mobile real-time ECG system	(Yuksel, 19 Oct 2025)	Development kit, serial/PC pipeline, web service, SQL Server, mobile display and alerting
Protocol-guided MLLM	(Jin et al., 4 Feb 2026)	Multimodal ECG interpretation, missing-modality robustness, evidence-grounded RL

The coexistence of these meanings is important for technical reading. In hardware-centric work, ECG-R1 denotes a signal-acquisition and event-extraction substrate. In mobile monitoring, it denotes an end-to-end telemonitoring stack. In the 2026 interpretation paper, it denotes a clinical reasoning model operating over ECG images and time-series rather than a front-end sensor or firmware pipeline (Jin et al., 4 Feb 2026).

This multiplicity also clarifies a common source of confusion: references to ECG-R1 do not necessarily imply a single evolving platform. The available record instead shows repeated reuse of the label for distinct systems developed for different layers of the ECG pipeline.

2. ECG-R1 as a protocol-guided multimodal interpretation model

The most developed and explicit contemporary use of the term is "ECG-R1: Protocol-Guided and Modality-Agnostic MLLM for Reliable ECG Interpretation" (Jin et al., 4 Feb 2026). This system is motivated by the claim that existing multimodal LLMs often produce plausible but clinically incorrect ECG reports, with two major failure modes: severe hallucinations and instability when one modality is missing. The paper treats this as especially problematic because ECGs are commonly available both as 12-lead images and as time-series signals, and these two views should yield consistent interpretations of the same underlying waveform.

ECG-R1 addresses this by separating the two modalities architecturally. It uses a visual backbone for ECG images and ECG-CoCa for the time-series signal, with two independent projectors that map each modality into the LLM space. The modality embeddings are injected at explicit <ecg> and <image> tags rather than forcing one modality through a representation bottleneck intended for the other. This modality-decoupled design is paired with training procedures intended to make image-only, signal-only, and joint interpretations remain aligned.

The model’s supervision is protocol-grounded rather than purely report-imitation-based. A deterministic feature extractor computes structured physiological evidence from the raw time-series, including heart rate, RR intervals, PR interval, QRS amplitude and duration, T-wave measures, ST descriptors, QT/QTc, and related lead-wise sequences. These features are then combined with a monograph-derived protocol reorganized into five phases: technical/rate/rhythm; conduction/axis/intervals; chamber hypertrophy/voltage; ischemia/infarction/mimics; and electrolytes/QT. The resulting supervision follows a fixed schema with a > block containing six reasoning steps and an <answer> block containing the final diagnosis (Jin et al., 4 Feb 2026).

This design makes ECG-R1 notable within ECG AI because it attempts to constrain textual reasoning by explicit waveform evidence and diagnostic rules, rather than relying on fluent medical prose as a proxy for correctness.

3. Core methodology: protocol grounding, modality dropout, and evidence rewards

ECG-R1 is organized around three methodological innovations. The first is Protocol-Guided Instruction Data Generation. Using DeepSeek-V3.1-Terminus as the interpretation generator and MIMIC-IV-ECG as the source, the authors curate 30,000 protocol-guided instruction samples. The guiding idea is that interpretations should be anchored in measurable features and monograph-defined thresholds and logic, thereby reducing hallucinated causal attributions and surfacing clinically meaningful abnormalities that pure prompt-based generation may omit (Jin et al., 4 Feb 2026).

The second innovation is Interleaved Modality Dropout, or IMD. During both supervised fine-tuning and reinforcement learning, the model is exposed to transformations that drop one modality or swap the order of modality-token blocks. The paper formalizes this as a training distribution over four transformations, with risks defined over missing-modality and swapped-order conditions. The stated robustness result is that optimizing the mixture risk promotes robustness across the corresponding test environments, while consistency bounds are derived in terms of total variation discrepancies between image-only and time-series-only outputs, and between swapped-order outputs. In the paper’s framing, the intrinsic view gap is assumed small because both image and signal are renderings of the same waveform (Jin et al., 4 Feb 2026).

The third innovation is Reinforcement Learning with ECG Diagnostic Evidence Rewards, or EDER. The reinforcement-learning subset contains 3,948 examples. For each reference reasoning trace, an evidence extractor identifies step-specific key diagnostic phrases from each of the six reasoning steps. The step reward is exact evidence coverage, the process reward is the average over steps, the diagnosis reward is set-level Jaccard similarity over diagnoses, and a format reward enforces the presence of a <think> block. The total reward is the sum of format reward, diagnosis accuracy reward, and evidence reward with $\lambda = 1.0$. Optimization uses DAPO with a decoupled-clipping PPO objective and clipping parameters $\epsilon_{\text{low}} = 0.2$ and $\epsilon_{\text{high}} = 0.3$ (Jin et al., 4 Feb 2026).

Methodologically, this places ECG-R1 closer to a protocol-constrained diagnostic reasoner than to a generic visual-LLM. The explicit process reward is meant to discourage a known RL failure mode in medical generation: short, high-confidence conclusions that omit the intermediate evidence needed for verification.

4. Empirical performance, robustness, and the hallucination problem

The evaluation reported for ECG-R1 is unusually broad. On the ECG-Grounding benchmark with 2,381 samples, outputs are scored with seven rubric-based metrics: Diagnosis Accuracy, Analysis Completeness, Analysis Relevance, Lead Evidence Validity, Grounded ECG Understanding, Evidence-Based Reasoning, and Clinical Diagnostic Fidelity. The comparison spans proprietary MLLMs, general open-source MLLMs, medical MLLMs, and ECG-specialized MLLMs (Jin et al., 4 Feb 2026).

The reported results directly support the paper’s claim that severe hallucinations are widespread in non-specialized systems. GPT-5.1-Instant achieves the best diagnosis accuracy among non-specialized models, but only 31.48. Medical MLLMs remain below 30.00 diagnosis accuracy. Among ECG-specialized systems, the previous best GEM reaches 74.70 diagnosis accuracy, while ECG-R1 reaches 79.33 after supervised fine-tuning and 80.29 after reinforcement learning. The final RL model reports 80.57 ECG Feature Grounding, 79.08 Evidence-Based Reasoning, and 84.20 Clinical Diagnostic Fidelity, with a reported +17.49 average absolute gain over GEM on the grounding, reasoning, and fidelity metrics (Jin et al., 4 Feb 2026).

Robustness under modality loss is a central empirical result. GEM shows a maximum relative drop of 28.0% in diagnosis accuracy and 44.9% in analysis relevance when one modality is removed, especially in time-series-only settings. ECG-R1 is reported as substantially more stable. Cross-modal agreement between time-series-only and image-only outputs improves from GEM’s BLEU-4 of 0.33, ROUGE-L of 0.43, and SBERT-Score of 0.92 to ECG-R1’s 0.69, 0.73, and 0.97. The IMD ablation is particularly strong: without IMD, diagnosis accuracy in a modality-missing setting drops to 36.77, whereas with IMD it remains 77.91, while only slightly lowering peak omni-modality accuracy. The EDER ablation raises diagnosis accuracy from 79.96 to 80.29 and improves lead evidence validity, ECG feature grounding, and evidence-based reasoning, while also stabilizing generation length (Jin et al., 4 Feb 2026).

Clinical review is reported in two forms. Four licensed cardiologists reviewed 100 randomly sampled cases and rated ECG-R1 above GEM on analytical relevance, accuracy, completeness, reasoning quality, findings novelty, clinical value, and overall satisfaction; analytical accuracy was 4.34 for ECG-R1 versus 3.89 for GEM. At the same time, the paper explicitly frames ECG-R1 as a research system rather than a substitute for professional judgment, warns that the public should not directly trust model outputs without independent verification, and states that real deployment would require prospective validation, clinical governance, and regulatory review (Jin et al., 4 Feb 2026).

5. Earlier ECG-R1 systems in hardware and mobile monitoring

Before its use as an interpretation model, ECG-R1 was used for acquisition-centric systems. In "An ECG-on-Chip for Wearable Cardiac Monitoring Devices," ECG-R1 denotes a highly integrated low-power mixed-signal chip intended for wearable monitoring (Deepu et al., 2014). The chip integrates a programmable instrumentation amplifier, a band-pass filter, a 12-bit SAR ADC, a novel QRS detector, 8K on-chip SRAM, and CPU interfaces. The analog front end is described as a two-stage structure with a low-noise first stage and a programmable gain amplifier based on a flip-over-capacitor technique. The ADC samples at 256 Hz, and the QRS detector uses multiscale mathematical morphology rather than a computationally expensive digital filtering chain. Morphological opening and closing are used for baseline-drift removal, the filter length is set to 25 samples to match the expected QRS duration at 256 Hz, and adaptive thresholding updates the threshold to the newly detected peak level. R–R intervals are then measured by counting clock cycles between successive R peaks, while heart rate is calculated by counting the number of R peaks in the last 60 s.

At the system level, the chip includes an asynchronous FIFO, Gray-code pointer synchronization across clock domains, status flags such as full and nearly full, a duplex SPI slave interface, and an 8 Kb on-chip SRAM buffer implemented as 512 × 16 dual-port memory. It is fabricated in a 0.35 µm standard CMOS process; the analog core runs at 1 V, digital logic and SRAM at 3.3 V, total core area is 5.74 mm², and total power consumption is 9.6 µW. In this usage, ECG-R1 is a front-end processor that converts low-amplitude, noisy ECG into digitized waveform, QRS events, and heart-rate data suitable for wearable operation (Deepu et al., 2014).

A different usage appears in "Monitoring Real-Time ECG Signals on Mobile Systems," where ECG-R1 denotes a portable real-time ECG monitoring platform built around a development kit, a PC/server interface, and a mobile client (Yuksel, 19 Oct 2025). The acquisition side uses the EG01000 ECG preamplifier/development module with a three-lead ECG cable attached according to Einthoven’s triangle. The differential input is between the red and yellow leads, with the green lead as ground. The paper reports ECG signal amplitude of 0.2 mV to 2 mV and bandwidth of 0.05 Hz to 150 Hz. Data flow is described as ECG preamplifier → serial port or USB-TTL adapter → PC software → web service → SQL server → mobile or remote clients. The PC software is implemented in Visual Studio .NET 2015, serial communication is configured as 9600 baud, 8 data bits, 1 stop bit, and no parity, and marker bytes such as 0xF8, 0xFA, 0xFB, and 0x11 distinguish waveform samples, pulse values, information bytes, and "LEAD OFF" status.

This mobile ECG-R1 also combines hardware and software filtering. The ECG preamplifier includes a 0.05 Hz high-pass filter and a 150 Hz low-pass filter; software filtering uses a 4th-order Butterworth filter implemented with MATLAB’s filter toolbox. The system stores data through a web service in SQL Server, enables real-time and historical viewing on mobile clients, and issues alerts when values fall below a threshold or reach a critical value, although no numeric threshold is specified. The paper emphasizes portability, mobile access, and possible extension to patient tracking via GPS and sudden intervention such as defibrillation (Yuksel, 19 Oct 2025).

These earlier systems establish an important distinction. In their hardware and telemonitoring forms, ECG-R1 refers to acquisition, buffering, transmission, and event extraction. In its MLLM form, the label shifts upward in the stack to protocol-grounded clinical interpretation.

6. Position within broader ECG research

The contemporary interpretation-oriented ECG-R1 sits within a wider ECG research landscape that has increasingly emphasized interpretable structure, modality transfer, and robustness to real-world corruption. Work on automated and interpretable ECG profiling developed a CNN-HMM segmentation pipeline and a 725-element patient-level ECG profile for disease detection and tracking, explicitly favoring physiologically meaningful intermediate representations over raw black-box classification (Tison et al., 2018). This is conceptually aligned with ECG-R1’s reliance on explicit feature extraction and protocolized reasoning rather than unconstrained narrative generation.

A second adjacent theme is beat-structured representation learning. HeartLang treats heartbeats as words and rhythms as sentences, using a QRS-Tokenizer, vector-quantized heartbeat reconstruction, masked ECG sentence pre-training, and a reported heartbeat vocabulary of 5,394 effective discrete words derived from a configured codebook size of 8,192 (Jin et al., 15 Feb 2025). Although this work is not an interpretation MLLM, it similarly rejects arbitrary fixed-window processing in favor of clinically meaningful units. This suggests a broader movement in ECG AI toward semantically structured intermediate representations.

A third theme is robustness before interpretation. Blind ECG restoration using 1D operational Cycle-GANs addresses mixed, nonstationary artifacts in Holter and wearable recordings, evaluates on CPSC-2020, and reports cardiologist preference for restored ECGs in 95.51% of cases, with only 0.04% of arrhythmia beats restored as normal beats and no false arrhythmia beats created (Kiranyaz et al., 2022). Spatial incompleteness is addressed by WearECG, which reconstructs 12-lead ECG from leads II, V1, and V5 using a VAE and reports downstream macro-AUROC of 0.8333 on MIMIC, compared with 0.8465 for original 12-lead and 0.7837 for 3-lead input (Guan et al., 13 Oct 2025). Other work digitizes paper ECGs into sampled signals (Pazos-Santomé et al., 2024), supports contactless monitoring via ballistocardiography with HRV indices numerically close to ECG across 20 subjects (Remesh et al., 2023), or reconstructs ECG from millimeter-wave radar with median R-peak timing error of 3 ms and median Pearson correlation of 90% (Chen et al., 2021).

Taken together, these results indicate that the reliability problem addressed by ECG-R1 is only one layer of a broader pipeline. A plausible implication is that trustworthy ECG reasoning depends not only on grounded report generation, but also on signal quality, segmentation fidelity, modality completeness, and acquisition context. The 2026 ECG-R1 paper explicitly recognizes only part of this stack: it improves interpretation reliability and missing-modality robustness, but still requires independent verification and remains outside the scope of prospective clinical deployment (Jin et al., 4 Feb 2026).

In this broader sense, ECG-R1 marks a transition in the use of the label from low-power ECG capture and mobile telemetry toward evidence-grounded multimodal clinical reasoning. The shared name masks a substantive shift in what counts as an ECG system: from chips that detect QRS complexes, to software that streams traces to mobile devices, to models that attempt to justify diagnoses in protocol-aligned text.