Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis

# Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis
Source: [https://arxiv.org/html/2608.19297](https://arxiv.org/html/2608.19297)
Yihan Xie111footnotemark:1Hanwen Cui211footnotemark:1Runze Ye111footnotemark:1Juekai Lin1Haoyang Wang1Jinhao Mao1 Bo Zhang3Wenqiao Zhang122footnotemark:2Xiaogang Guo122footnotemark:2Jun Xiao122footnotemark:2Lei Zhang1email:[\{yihanxie, wenqiaozhang\}@zju\.edu\.cn](mailto:{yihanxie,%20wenqiaozhang}@zju.edu.cn)Affiliation:1Zhejiang University2Beijing Institute of Technology3University of Electronic Science and Technology of China

![Refer to caption](https://arxiv.org/html/2608.19297v1/Dataset.png)Figure 1\.Overview of our proposed Holtercare\-23K dataset\.###### Abstract\.

††footnotetext:∗These authors contributed equally to this research\.††footnotetext:†Corresponding authors\.

While multimodal large language models \(MLLMs\) excel in medical applications, most of them favor static images or short\-term signals\. In the critical field of dynamic electrocardiograms \(ECG\), models struggle with complex temporal reasoning and diagnostic report generation due to a lack of high\-quality datasets and benchmarks\. To address this, we introduce\(i\) Holtercare\-23K, a large\-scale multimodal dynamic ECG dataset comprising 22,980 QA pairs derived from 788 clinical Holter records and featuring a novel signal–video–text tri\-modal alignment\. Based on this dataset, we present\(ii\) Holtercare\-Bench, a multimodal benchmark that evaluates models on temporal localization, clinical diagnosis, and global summarization\. Zero\-shot evaluations of leading MLLMs reveal a significant performance gap in processing ultra\-long pathological sequences\. However, fine\-tuning representative models yields substantial improvements\. This work illuminates the limitations of current MLLMs in electrophysiology and provides a foundational benchmark for long\-term medical MLLMs\. Our project is available at[https://github\.com/ZJU4HealthCare/Holtercare\-Bench](https://github.com/ZJU4HealthCare/Holtercare-Bench)\.

## 1\.Introduction

Recent advances in multimodal large language models \(MLLMs\) have revolutionized medical artificial intelligence \(AI\)\. However, current research is heavily skewed toward spatial modalities, such as radiological images and pathology slides\. In cardiology, while short\-term electrocardiogram \(ECG\) datasets like PTB\-XL\([52](https://arxiv.org/html/2608.19297#bib.bib15)\)and MIMIC\-IV\-ECG\([11](https://arxiv.org/html/2608.19297#bib.bib14)\)have driven crucial progress, they only capture a few seconds of cardiac activity\. In clinical practice, continuous dynamic ECG Holter monitoring is the gold standard for diagnosing intermittent arrhythmias\. Unfortunately, the underdevelopment of the data ecosystem for Holter monitoring has left modern MLLMs untested and unoptimized for long\-term cardiac care\.

Analyzing dynamic ECGs presents unique challenges that cannot be addressed by short\-segment datasets or benchmarks\. A standard Holter record spans around 24 hours, containing hundreds of thousands of heartbeats\. Within this massive temporal context, critical pathological events—such as a brief episode of ventricular tachycardia—are extremely fleeting\. Finding them requires processing an ultra\-long sequence while maintaining temporal sensitivity at the millisecond \(ms\) level\. Furthermore, clinical diagnosis is not just about classification; it requires linking micro\-level waveform changes \(e\.g\., missing P\-waves\) and macro\-level disease labels\. Current MLLMs lack both high\-quality data and specific benchmarks for learning this complex clinical reasoning\.

To bridge this gap, we introduceHoltercare\-Bench, a multimodal benchmark for evaluating long\-term dynamic ECG analysis\. The main contributions of our work are as follows:

\(i\) Dataset\.We proposeHoltercare\-23K, a large\-scale dynamic ECG dataset containing 788 real\-world clinical records, lasting 13–24 hours each\. A major hurdle in applying MLLMs to electrophysiology is the modality mismatch: modern language models cannot naturally process raw voltage arrays\. To overcome this, we designHolterAgent, an automated data engine to convert Holter records into asignal–video–texttri\-modal format\. Additionally, we construct a total of 22,980 QA pairs across three distinct types \(Closed\-QA,Open\-QA, andReport Generation\) from expert clinical annotations and reports, bridging the gap between electrophysiological signals and modern MLLM architectures\. An overview of the complete Holtercare\-23K dataset is illustrated in Figure[1](https://arxiv.org/html/2608.19297#S0.F1)\.

\(ii\) Benchmark\.Based on Holtercare\-23K, we proposeHoltercare\-Bench, a comprehensive evaluation framework designed to quantify a model’s cognitive ability in long\-context cardiology\. Holtercare\-Bench consists of 12 fine\-grained tasks that mirror the actual diagnostic workflow of a cardiologist\. Instead of relying on a single generic metric, we deploy a tailored scoring system\. This ranges from matching accuracy for discrete decision\-making to an LLM\-as\-a\-judge system for evaluating the logical consistency of generated clinical reports\.

We evaluate a wide range of mainstream generalist and medical MLLMs on our benchmark\. Zero\-shot results expose a significant performance gap—most current models struggle to maintain temporal consistency over ultra\-long sequences\. However, fine\-tuning on our dataset leads to substantial performance gains, demonstrating the efficacy of Holtercare\-23K in enhancing complex clinical reasoning\. Our results establish a strong baseline and define the performance frontiers for future long\-context medical MLLMs\.

## 2\.Related Work

Medical MLLMs and ECG Representation Learning\.Recent MLLMs have progressively advanced from general vision–language understanding toward fine\-grained spatial–temporal perception, instruction\-driven visual reasoning, and adaptive multimodal architectures\. Representative efforts include VideoRefer\([63](https://arxiv.org/html/2608.19297#bib.bib64)\)and PixelRefer\([64](https://arxiv.org/html/2608.19297#bib.bib65)\)for spatial\-temporal object understanding, InstructSAM\([62](https://arxiv.org/html/2608.19297#bib.bib66)\)for instruction\-driven visual segmentation, VisualThink\-VLA\([9](https://arxiv.org/html/2608.19297#bib.bib68)\)for visual intermediate reasoning, and HyperLLaVA\([65](https://arxiv.org/html/2608.19297#bib.bib67)\)for dynamic visual–language adaptation\. These advances have also extended to the medical domain, where MLLMs such as Med\-Flamingo\([40](https://arxiv.org/html/2608.19297#bib.bib37)\), LLaVA\-Med\([21](https://arxiv.org/html/2608.19297#bib.bib5)\), MedVLM\([43](https://arxiv.org/html/2608.19297#bib.bib6)\), Lingshu\([56](https://arxiv.org/html/2608.19297#bib.bib4)\), and HealthGPT\([29](https://arxiv.org/html/2608.19297#bib.bib7)\)have advanced multimodal clinical understanding\. Specialized MLLMs have further emerged across diverse clinical domains, including LLaVA\-Rad\([5](https://arxiv.org/html/2608.19297#bib.bib49)\)in radiology, EyecareGPT\([23](https://arxiv.org/html/2608.19297#bib.bib59)\)in ophthalmology, and SkinGPT\([68](https://arxiv.org/html/2608.19297#bib.bib50)\)in dermatology\. More recent works have moved toward modality\-specific modeling and clinically grounded reasoning, including unified slice\-volume analysis in OmniCT\([28](https://arxiv.org/html/2608.19297#bib.bib60)\), multimodal chain\-of\-thought \(CoT\) reasoning in TumorChain\([24](https://arxiv.org/html/2608.19297#bib.bib61)\), Group Relative Policy Optimization \(GRPO\) in TIF\-GRPO\([27](https://arxiv.org/html/2608.19297#bib.bib62)\), and evidence\-driven multimodal reinforcement learning in E\-MRL\([25](https://arxiv.org/html/2608.19297#bib.bib63)\)\.

In the ECG domain, models like ECGFounder\([22](https://arxiv.org/html/2608.19297#bib.bib38)\)and ECG\-FM\([35](https://arxiv.org/html/2608.19297#bib.bib39)\)enhance pretraining through massive datasets and a combination of self\-supervised learning techniques\. Multimodal integration is also advancing: ECG\-SL\([60](https://arxiv.org/html/2608.19297#bib.bib51)\), HeartLang\([17](https://arxiv.org/html/2608.19297#bib.bib52)\), and ESI\([59](https://arxiv.org/html/2608.19297#bib.bib55)\)focus on signal\-semantic alignment, whereas ECG\-LM\([58](https://arxiv.org/html/2608.19297#bib.bib53)\), SuPreME\([4](https://arxiv.org/html/2608.19297#bib.bib54)\), MERL\([31](https://arxiv.org/html/2608.19297#bib.bib44)\), and KED\([50](https://arxiv.org/html/2608.19297#bib.bib40)\)leverage external clinical knowledge to improve representations\. Concurrently, generative paradigms bridge LLMs and ECGs through continuous voltage tokenization as in ECG\-Byte\([14](https://arxiv.org/html/2608.19297#bib.bib43)\), instruction tuning for report generation like MEIT\([53](https://arxiv.org/html/2608.19297#bib.bib41)\), or visual adaptation as in PULSE\([33](https://arxiv.org/html/2608.19297#bib.bib42)\)\. Despite these advances, current research generally addresses isolated tasks or short\-duration static inputs, lacking a unified framework for complex reasoning over ultra\-long Holter contexts\.

Table 1\.Systematic comparison of current ECG datasets\.DatasetLeadsDurationSizeAnnotationBeatRhythmReportStatic ECG DatasetsPTB\-XL\([52](https://arxiv.org/html/2608.19297#bib.bib15)\)1210 seconds21,837✓ecg\-arrhythmia\([67](https://arxiv.org/html/2608.19297#bib.bib16)\)1210 seconds45,152✓MIMIC\-IV\-ECG\([11](https://arxiv.org/html/2608.19297#bib.bib14)\)1210 seconds800K✓CPSC 2018\([32](https://arxiv.org/html/2608.19297#bib.bib30)\)126–60 seconds9,831✓LUDB\([18](https://arxiv.org/html/2608.19297#bib.bib17)\)1210 seconds200✓Dynamic ECG DatasetsMITDB\([39](https://arxiv.org/html/2608.19297#bib.bib23)\)230 minutes48✓INCARTDB\([57](https://arxiv.org/html/2608.19297#bib.bib19)\)1230 minutes75✓LTSTDB\([16](https://arxiv.org/html/2608.19297#bib.bib24)\)2–321–24 hours86✓✓Icentia11k\([49](https://arxiv.org/html/2608.19297#bib.bib18)\)170 minutes11K✓✓AFDB\([36](https://arxiv.org/html/2608.19297#bib.bib25)\)210 hours25✓✓LTDB\([37](https://arxiv.org/html/2608.19297#bib.bib20)\)214–22 hours7✓SDDB\([13](https://arxiv.org/html/2608.19297#bib.bib26)\)214–25 hours23✓NSRDB\([38](https://arxiv.org/html/2608.19297#bib.bib21)\)223–26 hours18✓LTAFDB\([46](https://arxiv.org/html/2608.19297#bib.bib27)\)224–25 hours84✓✓SHDB\-AF\([51](https://arxiv.org/html/2608.19297#bib.bib22)\)224 hours143✓QTDB\([19](https://arxiv.org/html/2608.19297#bib.bib28)\)215 minutes105✓Apnea\-ECG\([45](https://arxiv.org/html/2608.19297#bib.bib29)\)17–10 hours70✓✓Holtercare\-23K \(Ours\)313–24 hours788✓✓✓

ECG Analysis Datasets\.Public ECG datasets like PTB\-XL\([52](https://arxiv.org/html/2608.19297#bib.bib15)\)and MIMIC\-IV\-ECG\([11](https://arxiv.org/html/2608.19297#bib.bib14)\)have driven deep learning diagnostics but consist of short signal segments\. Consequently, they fail to capture long\-term arrhythmia dynamics\. For continuous monitoring, dynamic datasets like MITDB\([39](https://arxiv.org/html/2608.19297#bib.bib23)\), LTSTDB\([16](https://arxiv.org/html/2608.19297#bib.bib24)\), and LTAFDB\([46](https://arxiv.org/html/2608.19297#bib.bib27)\)have been introduced\. However, the critical limitations of these are shown in Table[1](https://arxiv.org/html/2608.19297#S2.T1)\. They are either too small to train large\-scale models or feature limited annotations confined to simple discrete labels and beat\-level localizations\. They severely lack fine\-grained descriptions required for complex cross\-modal analysis and clinical report generation\. This gap prevents models from learning causal clinical reasoning and underscores the critical need for large\-scale data with multimodal alignments\.

Medical Benchmarks\.Current benchmarks are inadequate for long\-term ECG analysis\. Recent large\-scale medical VQA benchmarks \(e\.g\., PMC\-VQA\([66](https://arxiv.org/html/2608.19297#bib.bib56)\), GMAI\-MMBench\([7](https://arxiv.org/html/2608.19297#bib.bib57)\)\) offer thorough evaluation frameworks across various clinical modalities; however, along with previous datasets such as VQA\-RAD\([20](https://arxiv.org/html/2608.19297#bib.bib45)\)and Slake\([30](https://arxiv.org/html/2608.19297#bib.bib46)\), they are strictly limited to static spatial images\. Even within the cardiovascular field, current benchmarks like ECG\-QA\([41](https://arxiv.org/html/2608.19297#bib.bib58)\)primarily emphasize short\-duration static waveforms rather than continuous monitoring\. In contrast, general long\-sequence benchmarks \(Video\-MME\([8](https://arxiv.org/html/2608.19297#bib.bib47)\), LVBench\([54](https://arxiv.org/html/2608.19297#bib.bib48)\)\) focus on ordinary videos lacking specialized medical reasoning\. Evaluating dynamic ECGs demands high\-dimensional pathological feature extraction and temporal dependency modeling, yet the community currently lacks a comprehensive benchmark to systematically evaluate MLLMs on continuous, long\-term clinical reasoning\.

## 3\.Dataset: Holtercare\-23K

To address the scarcity of long\-term dynamic ECG data for MLLM research, we constructHoltercare\-23K, a large\-scale, multimodal dynamic ECG dataset derived from 788 clinically collected dynamic ECGs, which contains 22,980 QA pairs divided into three categories\.

### 3\.1\.Dataset Overview

All data in Holtercare\-23K originates from real\-world hospital collections recorded in 2026\. Compared with existing public databases in Table[1](https://arxiv.org/html/2608.19297#S2.T1), the dataset comprises 788 independent cases, with each continuous monitoring session lasting between 13 and 24 hours, as detailed in Figure[3](https://arxiv.org/html/2608.19297#S3.F3)\(a\)\. To ensure profound medical utility, every record is equipped with a comprehensive, tri\-level annotation system:

\(i\) Beat\-Level Annotations\.As Figure[3](https://arxiv.org/html/2608.19297#S3.F3)\(b\) shows, these annotations pinpoint the exact occurrence timestamps for 18 distinct heartbeat categories \(e\.g\., normal beats, premature contractions\) and various non\-beat events \(e\.g\., P\-waves, T\-waves, artifacts\)\.

\(ii\) Rhythm\-Level Annotations\.These annotations capture continuous cardiac events and arrhythmias and provide specific string descriptions \(e\.g\., ventricular tachycardia, ST\-segment changes\) bound by precise timestamps, with the top 10 frequent events shown in Figure[3](https://arxiv.org/html/2608.19297#S3.F3)\(c\)\.

\(iii\) Report\-Level Summaries\.Each case includes a global report strictly verified by professional cardiologists\. These reports provide overall recording statistics \(e\.g\., total heartbeats, average heart rate, exact timestamps for maximum/minimum heart rates, and duration percentages for tachycardia/bradycardia\) along with final diagnostic conclusions\.

Regarding data privacy and compliance, all raw data underwent rigorous de\-identification\. We remove all sensitive personal identifiers but preserve critical medical context: age, gender, and anonymize electronic medical records \(EMRs\)\. These EMRs provide brief clinical histories and primary symptoms \(e\.g\., “history of hypertension,” “dizziness,” or “chest tightness”\), serving as vital supplementary inputs for multimodal clinical reasoning\. The entire dataset construction fully complies with medical ethics and data protection laws\.

### 3\.2\.Multimodal Data Engine: HolterAgent

To transform this massive, multi\-grained clinical data into QA pairs suitable for MLLM evaluation, we develop an automated multimodal data engine,HolterAgent\. The systematic workflow of HolterAgent, from signal preprocessing to QA generation, is illustrated in Figure[2](https://arxiv.org/html/2608.19297#S3.F2)\. The processing pipeline involves three core modules:

![Refer to caption](https://arxiv.org/html/2608.19297v1/Engine.png)Figure 2\.The multimodal generation and data construction framework of HolterAgent\.![Refer to caption](https://arxiv.org/html/2608.19297v1/Counts.png)Figure 3\.Statistical distribution of Holtercare\-23K\. \(a\) Density estimation of recording durations\. \(b\) Log\-scale counts of beat\-level annotations\. \(c\) Frequencies of the top 10 rhythm\-level events\.\(i\) Signal Preprocessor\.To initiate the pipeline, HolterAgent first employs theMNE\([12](https://arxiv.org/html/2608.19297#bib.bib36)\)package to parse the raw clinical EDF files and extract continuous electrophysiological signals\. Considering the inevitable motion artifacts and baseline wandering in real clinical environments, we subsequently utilize theNeuroKit2\([34](https://arxiv.org/html/2608.19297#bib.bib31)\)package to perform lead\-level deep denoising and baseline correction\. This cascaded process ensures the extraction of the purest pathological waveform features for downstream tasks\.

\(ii\) Video and Text Generator\.To accommodate diverse MLLM architectures, HolterAgent transforms these cleaned, high\-quality signals into two extra modalities: clinical text records and dynamic video streams\. For text modality, the engine extracts the specified lead data, scales it to millivolts \(mV\), and formats it to three decimal places\. For video modality, the engine employsMatplotlib\([15](https://arxiv.org/html/2608.19297#bib.bib32)\)to construct sliding windows \(e\.g\., a 10\-second duration\), rendering the long\-term signals into dynamic ECG video streams\.

\(iii\) QA Constructor\.Relying on the preprocessed signals and the tri\-level annotations, we employ GPT\-5\-mini\([42](https://arxiv.org/html/2608.19297#bib.bib11)\)to perform deep semantic parsing and logical restructuring\. HolterAgent automatically transforms this rich clinical context into various QA types, yielding a total of 22,980 high\-quality, multimodal QA pairs\.

To ensure rigorous model evaluation and strictly prevent data leakage, we partition the dataset at the independent case level \(788 cases in total\) rather than the QA pair level\. Specifically, we randomly hold out 20% of the cases to form the test set\. The remaining 80% of the cases are further divided into training and validation sets following a 9:1 ratio\. Consequently, the 22,980 QA pairs are seamlessly distributed into their respective splits based on their source cases, guaranteeing that no patient data overlaps between the training, validation, and test phases\. An illustrative overview of Holtercare\-23K is depicted in Figure[1](https://arxiv.org/html/2608.19297#S0.F1)\.

## 4\.Benchmark: Holtercare\-Bench

To systematically evaluate the cognitive and multimodal reasoning capabilities of MLLMs on long\-term dynamic ECGs, we designHoltercare\-Benchto mirror the diagnostic workflow of human cardiologists\. The benchmark is structured into three progressive evaluation tiers:Closed\-QA,Open\-QA, andReport Generation\. This yields a total of 12 fine\-grained multimodal tasks, each equipped with rigorous, objective evaluation metrics\.

### 4\.1\.Closed\-QA Tasks

Closed\-QAtasks evaluate the fundamental feature recognition and discrete decision\-making abilities of models\. By providing explicit multiple\-choice candidates, these tasks assess absolute precision using strict matching accuracy\. This module includes five sub\-tasks:

\(i\)Presence\.Requires the model to determine whether a specific rhythm or abnormal waveform exists within the complex sequence\.

\(ii\)Event Counting\.Requires the model to select the correct occurrence count for a specific anomaly category from multiple options\.

\(iii\)Event Timing\.Requires the model to select the precise ms\-level timestamp combination for all occurrences of an abnormal event from highly similar interfering candidates\.

\(iv\)HR Extremum Timing\.Requires the model to accurately match the ms\-level timestamp of the fastest or slowest heart rate \(HR\) amidst global rhythm variations\.

\(v\)Diagnosis\.Requires the model to select the comprehensive diagnosis that best matches the overall characteristics of the current long\-term strip from multiple similar arrhythmia category options\.

### 4\.2\.Open\-QA Tasks

Open\-QAtasks require models to autonomously generate free\-text answers\. We evaluate text quality usingBLEU\([44](https://arxiv.org/html/2608.19297#bib.bib34)\),ROUGE\-L\([26](https://arxiv.org/html/2608.19297#bib.bib33)\), andF1\-Bio\([47](https://arxiv.org/html/2608.19297#bib.bib35)\), along with a custom normalized metricScoreMAEfor numerical reasoning\. This module includes five sub\-tasks:

\(i\)Event Countingand \(ii\)HR Extremum Counting\.To evaluate numerical accuracy while preventing metric collapse for baselines, we transform the mean absolute error \(MAE\) into a normalized score:S​c​o​r​eM​A​E=100/\[1\+exp⁡\(\(M​A​E−μ\)/σ\)\]Score^\{MAE\}=100/\[1\+\\exp\(\(MAE\-\\mu\)/\\sigma\)\], whereμ\\mu\(median\) andσ\\sigma\(standard deviation\) anchor baseline scores around 50\. This non\-linear scaling clearly distinguishes fine\-tuned models without severely compressing baselines\.

\(iii\)Event Timingand \(iv\)Diagnosis\.Models autonomously output structured ms\-level temporal offsets or multi\-label diagnoses, which are scored for precision and completeness using the text and overlap metrics described above\.

\(v\)Evidence Reasoning\.Models must generate rationales \(e\.g\., “absence of P waves”\) to support their diagnoses, demonstrating a causal logic chain from raw waveforms to clinical decisions\.

Table 2\.Evaluation dimensions and granular criteria for reports generated onStatistical Overviewtasks\.DimensionEvaluation CriteriaWeightGlobal MetricsAccurate extraction of recording duration15Accurate extraction of the total heartbeat count10Heart Rate StatsAccurate reporting of the average heart rate10Accurate reporting of the maximum heart rate10Accurate reporting of the minimum heart rate10Rhythm BurdenCorrect extraction of tachycardia burden \(percentage of time HR \> 100 bpm\)15Correct extraction of bradycardia burden \(percentage of time HR < 60 bpm\)15Report Logic and NormsLogically organized and coherent natural language5Adherence to clinical reporting conventions5Restriction to statistical metrics without introducing clinical diagnoses5Total Score100

### 4\.3\.Report Generation Tasks

Report Generationtasks require the model to step beyond local features and perform global abstraction over long\-term records, directly mimicking real\-world Holter reporting\. To ensure professional evaluation, we employ medical text metrics \(BLEU\([44](https://arxiv.org/html/2608.19297#bib.bib34)\),ROUGE\-L\([26](https://arxiv.org/html/2608.19297#bib.bib33)\),F1\-Bio\([47](https://arxiv.org/html/2608.19297#bib.bib35)\)\) and an innovative LLM\-as\-a\-judge framework based on GPT\-5\-mini\([42](https://arxiv.org/html/2608.19297#bib.bib11)\), which yields a comprehensive metric denoted asScoreGPT\. The evaluation rubric and the automated scoring consistency have been cross\-verified by professional cardiologists, ensuring clinical accuracy, completeness, and compliance\. This module includes two sub\-tasks:

\(i\)Statistical Overview\.Requires the model to extract and organize key clinical metrics \(e\.g\., total heartbeat count, extreme heart rates, anomaly burdens\) from long\-sequence data, ensuring global logical self\-consistency\. Scoring follows the criteria in Table[2](https://arxiv.org/html/2608.19297#S4.T2)\.

\(ii\)General Summary\.Requires the model produce a summary in the style of a clinical report, identifying key findings by combining local temporal patterns with global statistics\. Scoring follows the criteria in Table[3](https://arxiv.org/html/2608.19297#S4.T3)\.

Table 3\.Evaluation dimensions and granular criteria for reports generated onGeneral Summarytasks\.DimensionEvaluation CriteriaWeightStatistical AccuracyAccurate extraction of global metrics4Accurate reporting of HR extremes6Correct extraction of tachycardia and bradycardia burdens4Ectopic Beat and Rhythm DetailCorrect classification of the baseline rhythm \(e\.g\., AFib\)8Accurate total count and burden of PACs/PVCs7Precise breakdown of ectopic patterns \(e\.g\., isolated, paired, runs\)10Significant FindingsComplete inclusion of severe events \(e\.g\., VT runs, VF, asystole\)10Correct interpretation of conduction blocks or pacing signals8Correct reporting of ST\-segment and T\-wave changes7Factual FidelityStrict consistency with clinical facts without hallucinated findings8Objective description without exaggeration or understatement6Clinical diagnoses strictly grounded in supporting statistical data6Structural LogicCoherent narrative structure without fragmented data enumeration4Structured clinical hierarchy, prioritizing major diagnoses over secondary findings3Descriptive NormsPrecise application of medical terminology4Consistent numeric formatting and standardized unit application \(e\.g\., “bpm” for rate, “%” for burden\)3Inclusion of appropriate clinical caveats2Total Score100

ScoreGPTscores both generation task on a 100\-point rubric, defined as follows:

- •Perfect \(90–100\):Perfect or near\-perfect compliance with this specific criterion\.
- •Substantial \(70–89\):Substantial compliance, with minor flaws, slight inaccuracies, or trivial omissions\.
- •Partial \(40–69\):Captures the relevant clinical concept but applies an incorrect metric type or statistical aggregation\.
- •Poor \(10–39\):Barely related or severely flawed, but not completely blank\.
- •Failure \(0–9\):Complete failure; the concept is entirely missing or severely hallucinated\.

## 5\.Experiments

To validate Holtercare\-Bench and the fine\-tuning potential of the Holtercare\-23K dataset, we conduct a two\-stage experiment: a zero\-shot baseline evaluation of mainstream models, followed by a comparative analysis of two representative o pen\-source models before and after fine\-tuning\.

Table 4\.Performance comparison of evaluated models onClosed\-QAtasks from Holtercare\-Bench, evaluated by accuracy\.Boldandunderlineindicate the best and second\-best results, and and indicate the text and video modalities, respectively\. SuperscriptFTdenotes fine\-tuned models\.ModelModalityPresenceEventCountingEventTimingHR ExtremumTimingDiagnosisGeneralist ModelsGPT\-5\-mini76\.3131\.8033\.5438\.5448\.03Claude\-4\.5\-Haiku51\.6325\.9931\.7128\.3428\.35Phi\-4\-mini\-3\.8B52\.7827\.8333\.2336\.3124\.80Phi\-4\-mini\-3\.8BFT64\.3153\.5263\.4154\.7849\.61InternVL\-3\.5\-8B67\.9714\.0740\.8526\.1142\.13MiniCPM\-V4\.5\-8B52\.2919\.5737\.5039\.1719\.69Gemini\-3\.0\-Flash58\.1722\.6364\.3356\.3739\.76Qwen3\-VL\-8B59\.1524\.4635\.9833\.1241\.73Qwen3\-VL\-8BFT94\.7781\.6599\.7069\.1195\.28Medical ModelsLLaVA\-Med\-V1\.5\-7B43\.7922\.3225\.9112\.1027\.95MedGemma\-1\.5\-4B\-IT62\.4220\.8030\.7934\.7133\.46HealthGPT\-M3\-3\.8B56\.0514\.0724\.0928\.0331\.10Lingshu\-7B57\.8420\.1827\.7435\.3531\.89MedVLM\-R1\-2B54\.2542\.8130\.1832\.1735\.43HuatuoGPT\-Vision\-7B50\.6518\.9625\.0036\.3121\.26

### 5\.1\.Experimental Setup

Evaluated Models\.We evaluate 13 cutting\-edge models, categorized intoGeneralist Models\(GPT\-5\-mini\([42](https://arxiv.org/html/2608.19297#bib.bib11)\), Claude\-4\.5\-Haiku\([2](https://arxiv.org/html/2608.19297#bib.bib10)\), Phi\-4\-mini\-3\.8B\([1](https://arxiv.org/html/2608.19297#bib.bib13)\), InternVL\-3\.5\-8B\([55](https://arxiv.org/html/2608.19297#bib.bib2)\), MiniCPM\-V4\.5\-8B\([61](https://arxiv.org/html/2608.19297#bib.bib3)\), Gemini\-3\.0\-Flash\([10](https://arxiv.org/html/2608.19297#bib.bib12)\), and Qwen3\-VL\-8B\([3](https://arxiv.org/html/2608.19297#bib.bib1)\)\) andMedical Models\(LLaVA\-Med\-V1\.5\-7B\([21](https://arxiv.org/html/2608.19297#bib.bib5)\), MedGemma\-1\.5\-4B\-IT\([48](https://arxiv.org/html/2608.19297#bib.bib8)\), HealthGPT\-M3\-3\.8B\([29](https://arxiv.org/html/2608.19297#bib.bib7)\), Lingshu\-7B\([56](https://arxiv.org/html/2608.19297#bib.bib4)\), MedVLM\-R1\-2B\([43](https://arxiv.org/html/2608.19297#bib.bib6)\)and HuatuoGPT\-Vision\-7B\([6](https://arxiv.org/html/2608.19297#bib.bib9)\)\)\.

Modality Adaptation\.Since these models cannot directly process raw electrophysiological signals, we utilize the tri\-modal alignment of Holtercare\-23K, evaluating video\-capable models \(e\.g\., Qwen3\-VL\-8B\) via video modality and others \(e\.g\., GPT\-5\-mini\) via transformed text modality\. Notably, due to inherent model input capacity limits, extended continuous ECG text sequences may be truncated, and video inputs are constrained by file size limits, necessitating accelerated playback or truncation\.

Fine\-Tuning Setup\.To explore the impact of our dataset on model performance, we select two representative open\-source models for instruction tuning:Phi\-4\-mini\-3\.8B\([1](https://arxiv.org/html/2608.19297#bib.bib13)\), evaluated on text, andQwen3\-VL\-8B\([3](https://arxiv.org/html/2608.19297#bib.bib1)\), evaluated on video\. We partition Holtercare\-23K into training, validation, and test sets to align the models with actual clinical workflows\.

Table 5\.Performance comparison of evaluated models onReport Generationtasks from Holtercare\-Bench\.ModelModalityStatistical OverviewGeneral SummaryROUGE\-L↑\\uparrowF1\-Bio↑\\uparrowScoreGPT↑\\uparrowF1\-Bio↑\\uparrowROUGE\-L↑\\uparrowScoreGPT↑\\uparrowGeneralist ModelsGPT\-5\-mini13\.6880\.5631\.0811\.0381\.3927\.28Claude\-4\.5\-Haiku9\.0873\.0328\.327\.7473\.5528\.89Phi\-4\-mini\-3\.8B25\.9282\.2631\.6311\.5165\.5726\.77Phi\-4\-mini\-3\.8BFT49\.9490\.6345\.8834\.6088\.6829\.95InternVL\-3\.5\-8B31\.1583\.5419\.5810\.5861\.1830\.22MiniCPM\-V4\.5\-8B14\.9775\.3815\.819\.4575\.4126\.31Gemini\-3\.0\-Flash16\.5671\.9622\.5910\.9775\.7728\.51Qwen3\-VL\-8B13\.3067\.5215\.4310\.9269\.8424\.64Qwen3\-VL\-8BFT51\.3193\.2341\.0357\.1695\.2540\.79Medical ModelsLLaVA\-Med\-V1\.5\-7B15\.3274\.8313\.468\.0562\.7016\.34MedGemma\-1\.5\-4B\-IT16\.0976\.8429\.937\.9466\.5436\.48HealthGPT\-M3\-3\.8B12\.8169\.8519\.2711\.4963\.9832\.26Lingshu\-7B13\.3663\.5914\.7310\.4561\.5131\.37MedVLM\-R1\-2B22\.9769\.9217\.2911\.9967\.9125\.90HuatuoGPT\-Vision\-7B12\.7374\.1125\.249\.4470\.3727\.66

Table 6\.Performance comparison of evaluated models onOpen\-QAtasks from Holtercare\-Bench\.ModelModalityEvent CountingEvent TimingHR Extremum TimingDiagnosisEvidence ReasoningMAE↓\\downarrowScoreMAE↑\\uparrowROUGE\-L↑\\uparrowF1\-Bio↑\\uparrowMAE \(10710^\{7\}ms\)↓\\downarrowScoreMAE↑\\uparrowROUGE\-L↑\\uparrowF1\-Bio↑\\uparrowROUGE\-L↑\\uparrowF1\-Bio↑\\uparrowGeneralist ModelsGPT\-5\-mini2\.2377\.7220\.2463\.023\.8170\.7526\.0673\.7621\.4887\.91Claude\-4\.5\-Haiku17\.4217\.396\.0174\.754\.4550\.679\.4272\.3415\.2579\.67Phi\-4\-mini\-3\.8B2\.0778\.2325\.2081\.435\.4321\.6733\.0783\.4621\.1987\.11Phi\-4\-mini\-3\.8BFT0\.6382\.4246\.1690\.882\.3094\.8081\.9597\.9136\.2094\.41InternVL\-3\.5\-8B16\.9418\.7017\.5675\.614\.4849\.6725\.4878\.8918\.5887\.91MiniCPM\-V4\.5\-8B14\.4126\.8611\.4075\.884\.7241\.7112\.0975\.1618\.5788\.32Gemini\-3\.0\-Flash2\.4876\.9131\.9981\.994\.4650\.3337\.7479\.8819\.4785\.27Qwen3\-VL\-8B7\.3457\.575\.2751\.484\.4152\.0116\.1272\.3312\.4557\.42Qwen3\-VL\-8BFT0\.6382\.4244\.8990\.332\.7491\.0182\.6797\.6738\.1194\.19Medical ModelsLLaVA\-Med\-V1\.5\-7B9\.6946\.7720\.2880\.854\.4849\.6729\.4483\.8718\.9386\.94MedGemma\-1\.5\-4B\-IT7\.8755\.1618\.9756\.994\.4750\.0018\.9460\.9714\.6872\.65HealthGPT\-M3\-3\.8B9\.3248\.4811\.4376\.494\.4750\.0012\.8462\.9421\.8586\.97Lingshu\-7B10\.1044\.895\.1558\.334\.3753\.3417\.2174\.4317\.1974\.70MedVLM\-R1\-2B11\.3939\.0923\.5680\.554\.4949\.3327\.1072\.7716\.7379\.58HuatuoGPT\-Vision\-7B8\.9950\.0025\.3675\.414\.5746\.6624\.6165\.1411\.6478\.22

Comprehensive evaluation results acrossClosed\-QA,Open\-QA, andReport Generationtasks encompassing both zero\-shot baselines and fine\-tuned models are detailed in Tables[4](https://arxiv.org/html/2608.19297#S5.T4),[5](https://arxiv.org/html/2608.19297#S5.T5), and[6](https://arxiv.org/html/2608.19297#S5.T6), respectively\.

### 5\.2\.Baseline Evaluation

The zero\-shot baseline results, as visualized in Figure[4](https://arxiv.org/html/2608.19297#S5.F4), reveal significant limitations in existing models when processing long\-term dynamic ECGs\. While advanced generalist models exhibit reasonable foundational comprehension onClosed\-QAtasks—such as GPT\-5\-mini achieving 76\.31% accuracy inPresencedetection—performance drops sharply on fine\-grained temporal problems within theOpen\-QAcategory\. Specifically,Open\-QAtasks likeEvent CountingandHR Extremum Timingprove highly challenging for most baselines\. Interestingly, vision\-language models like Gemini\-3\.0\-Flash demonstrate a relative advantage on theEvent Timingsub\-task ofOpen\-QA\. Yet, all models without fine\-tuning, including medical\-specific ones, struggle significantly with reasoning required inOpen\-QAand the ultra\-long context demands ofReport Generation, yielding low entity coverage and poorROUGE\-Lscores\.

![Refer to caption](https://arxiv.org/html/2608.19297v1/Comparison-1.png)Figure 4\.Zero\-shot performance of four representative models across diverse metrics\.
### 5\.3\.Fine\-Tuning Comparison

Our comparative analysis of Phi\-4\-mini\-3\.8B and Qwen3\-VL\-8B before and after fine\-tuning, shown in Figure[5](https://arxiv.org/html/2608.19297#S5.F5), demonstrates the dataset’s substantial potential\. Post\-tuning, both models exhibit massive performance leaps across all metrics and categories\. Most notably, Qwen3\-VL\-8B achieved an exceptional 99\.70% accuracy on theEvent Timingsub\-task ofOpen\-QAand 95\.28% on theDiagnosissub\-task ofClosed\-QA, while simultaneously reducingMAEon theEvent Countingsub\-task ofOpen\-QAfrom 7\.34 to 0\.63\. Furthermore, both fine\-tuned models successfully learn to synthesize ultra\-long sequences, achieving state\-of\-the\-artF1\-BioandROUGE\-Lscores on complexOpen\-QAreasoning andReport Generationtasks\. These results confirm that domain\-specific multimodal alignment successfully activates the models’ temporal perception and clinical reasoning capabilities, bridging the gap toward expert\-level diagnostic analysis\.

![Refer to caption](https://arxiv.org/html/2608.19297v1/Comparison-2.png)Figure 5\.Performance comparison of Phi\-4\-mini\-3\.8B and Qwen3\-VL\-8B before and after fine\-tuning\. Metrics includeAccuracy,ScoreMAE, andScoreGPT\. Tasks lacking these metrics are evaluated byF1\-Bio\.

## 6\.Conclusion

To address the challenges of MLLMs in long\-term ECG analysis, we introduceHoltercare\-23K, a large\-scale dynamic ECG dataset with a tri\-modal alignment, andHoltercare\-Bench, a multimodal benchmark for evaluating long\-term dynamic ECG analysis\. While baselines reveal current models struggle with precise temporal localization and causal reasoning in long sequences, our fine\-tuning demonstrates that high\-quality multimodal data unlocks their long\-context diagnostic capabilities\. Looking ahead, Holtercare\-Bench provides a solid foundation for long\-context medical AI research\. We hope our contributions accelerate the development of native dynamic ECG MLLMs\.

## References

- M\. Abdin, J\. Aneja, H\. Behl, S\. Bubeck, R\. Eldan, S\. Gunasekar, M\. Harrison, R\. J\. Hewett, M\. Javaheripi, P\. Kauffmann,et al\.Phi\-4 technical report\.arXiv preprint arXiv:2412\.08905\.Cited by:[§C\.3](https://arxiv.org/html/2608.19297#A3.SS3.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p3.1.2)\.
- Anthropic \(2025\)AnthropicSystem card: claude haiku 4\.5\.Note:[https://www\.anthropic\.com/system\-cards](https://www.anthropic.com/system-cards)Cited by:[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Baiet al\.\(2025\)S\. Bai, Y\. Cai, R\. Chen, K\. Chen, X\. Chen, Z\. Cheng, L\. Deng, W\. Ding, C\. Gao, C\. Ge,et al\.Qwen3\-vl technical report\.arXiv preprint arXiv:2511\.21631\.Cited by:[§C\.3](https://arxiv.org/html/2608.19297#A3.SS3.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p3.1.3)\.
- Caiet al\.\(2025\)M\. Cai, J\. Jiang, W\. Huang, C\. Liu, and R\. ArcucciSuPreME: a supervised pre\-training framework for multimodal ecg representation learning\.arXiv preprint arXiv:2502\.196683\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Chaveset al\.\(2024\)J\. M\. Z\. Chaves, S\. Huang, Y\. Xu, H\. Xu, N\. Usuyama, S\. Zhang, F\. Wang, Y\. Xie, M\. Khademi, Z\. Yang,et al\.Towards a clinically accessible radiology foundation model: open\-access and lightweight, with automated evaluation\.arXiv preprint arXiv:2403\.08002\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Chenet al\.\(2024a\)J\. Chen, C\. Gui, R\. Ouyang, A\. Gao, S\. Chen, G\. H\. Chen, X\. Wang, Z\. Cai, K\. Ji, X\. Wan,et al\.Towards injecting medical visual knowledge into multimodal llms at scale\.InProceedings of the 2024 conference on empirical methods in natural language processing,pp\. 7346–7370\.Cited by:[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Chenet al\.\(2024b\)P\. Chen, J\. Ye, G\. Wang, Y\. Li, Z\. Deng, W\. Li, T\. Li, H\. Duan, Z\. Huang, Y\. Su,et al\.Gmai\-mmbench: a comprehensive multimodal evaluation benchmark towards general medical ai\.Advances in Neural Information Processing Systems37,pp\. 94327–94427\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- Fuet al\.\(2025\)C\. Fu, Y\. Dai, Y\. Luo, L\. Li, S\. Ren, R\. Zhang, Z\. Wang, C\. Zhou, Y\. Shen, M\. Zhang,et al\.Video\-mme: the first\-ever comprehensive evaluation benchmark of multi\-modal llms in video analysis\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 24108–24118\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- Gaoet al\.\(2026\)M\. Gao, W\. Zhang, Y\. Yuan, Y\. Dai, B\. Yu, Z\. Lv, H\. Zheng, J\. Zhu, Z\. Ge, Z\. Wan,et al\.VisualThink\-vla: visual intermediate reasoning for effective and low\-latency vision\-language\-action policies\.arXiv preprint arXiv:2605\.30011\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Google DeepMind \(2025\)Google DeepMindGemini 3 flash model card\.Note:[https://deepmind\.google/models/model\-cards/gemini\-3\-flash/](https://deepmind.google/models/model-cards/gemini-3-flash/)Cited by:[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Gowet al\.\(2023\)B\. Gow, T\. Pollard, L\. A\. Nathanson, A\. Johnson, B\. Moody, C\. Fernandes, N\. Greenbaum, J\. W\. Waks, P\. Eslami, T\. Carbonati, A\. Chaudhari, E\. Herbst, D\. Moukheiber, S\. Berkowitz, R\. Mark, and S\. HorngMIMIC\-IV\-ECG: Diagnostic Electrocardiogram Matched Subset\.PhysioNet\.Note:Version 1\.0External Links:[Document](https://dx.doi.org/10.13026/4nqg-sb35),[Link](https://doi.org/10.13026/4nqg-sb35)Cited by:[§1](https://arxiv.org/html/2608.19297#S1.p1.1),[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.6.1.1),[§2](https://arxiv.org/html/2608.19297#S2.p3.1)\.
- Gramfortet al\.\(2014\)A\. Gramfort, M\. Luessi, E\. Larson, D\. A\. Engemann, D\. Strohmeier, C\. Brodbeck, L\. Parkkonen, and M\. S\. HämäläinenMNE software for processing meg and eeg data\.neuroimage86,pp\. 446–460\.Cited by:[§3\.2](https://arxiv.org/html/2608.19297#S3.SS2.p2.1)\.
- Greenwald \(1986\)S\. D\. GreenwaldThe development and analysis of a ventricular fibrillation detector\.Ph\.D\. Thesis,Massachusetts Institute of Technology\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.16.1.1)\.
- Hanet al\.\(2024\)W\. Han, C\. Duan, M\. A\. Rosenberg, E\. Liu, and D\. ZhaoEcg\-byte: a tokenizer for end\-to\-end generative electrocardiogram language modeling\.arXiv preprint arXiv:2412\.14373\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Hunter \(2007\)J\. D\. HunterMatplotlib: a 2d graphics environment\.Computing in science & engineering9\(3\),pp\. 90–95\.Cited by:[§3\.2](https://arxiv.org/html/2608.19297#S3.SS2.p3.1)\.
- Jageret al\.\(2003\)F\. Jager, A\. Taddei, G\. B\. Moody, M\. Emdin, G\. Antolič, R\. Dorn, A\. Smrdel, C\. Marchesi, and R\. G\. MarkLong\-term st database: a reference for the development and evaluation of automated ischaemia detectors and for the study of the dynamics of myocardial ischaemia\.Medical and Biological Engineering and Computing41\(2\),pp\. 172–182\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.12.1.1),[§2](https://arxiv.org/html/2608.19297#S2.p3.1)\.
- Jinet al\.\(2025\)J\. Jin, H\. Wang, H\. Li, J\. Li, J\. Pan, and S\. HongReading your heart: learning ecg words and sentences via pre\-training ecg language model\.arXiv preprint arXiv:2502\.10707\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Kalyakulinaet al\.\(2021\)A\. Kalyakulina, I\. Yusipov, V\. Moskalenko, A\. Nikolskiy, K\. Kosonogov, N\. Zolotykh, and M\. IvanchenkoLobachevsky University Electrocardiography Database\.PhysioNet\.Note:Version 1\.0\.1External Links:[Document](https://dx.doi.org/10.13026/eegm-h675),[Link](https://doi.org/10.13026/eegm-h675)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.8.1.1)\.
- Lagunaet al\.\(1997\)P\. Laguna, R\. G\. Mark, A\. Goldberg, and G\. B\. MoodyA database for evaluation of algorithms for measurement of qt and other waveform intervals in the ecg\.InComputers in cardiology 1997,pp\. 673–676\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.20.1.1)\.
- Lauet al\.\(2018\)J\. J\. Lau, S\. Gayen, A\. Ben Abacha, and D\. Demner\-FushmanA dataset of clinically generated visual questions and answers about radiology images\.Scientific data5\(1\),pp\. 180251\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- Liet al\.\(2023\)C\. Li, C\. Wong, S\. Zhang, N\. Usuyama, H\. Liu, J\. Yang, T\. Naumann, H\. Poon, and J\. GaoLlava\-med: training a large language\-and\-vision assistant for biomedicine in one day\.Advances in Neural Information Processing Systems36,pp\. 28541–28564\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Liet al\.\(2025a\)J\. Li, A\. D\. Aguirre, V\. M\. Junior, J\. Jin, C\. Liu, L\. Zhong, C\. Sun, G\. Clifford, M\. Brandon Westover, and S\. HongAn electrocardiogram foundation model built on over 10 million recordings\.Nejm ai2\(7\),pp\. AIoa2401033\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Liet al\.\(2025b\)S\. Li, T\. Lin, L\. Lin, W\. Zhang, J\. Liu, X\. Yang, J\. Li, Y\. He, X\. Song, J\. Xiao, Y\. Zhuang, and B\. C\. OoiEyecareGPT: boosting comprehensive ophthalmology understanding with tailored dataset, benchmark and model\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 3893–3902\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Liet al\.\(2026a\)S\. Li, Z\. Qiu, J\. Liu, W\. Zhang, T\. Lin, Y\. Xie, J\. An, B\. Yun, C\. Yang, J\. Xiao, G\. Guo, J\. Yao, W\. Liu, Y\. Gao, K\. Yan, W\. Cao, Z\. Zheng, T\. C\. W\. Mok, K\. Cao, Y\. Shi, J\. Zhang, J\. Zhou, B\. C\. Ooi, Y\. Xia, and L\. ZhangTumorChain: interleaved multimodal chain\-of\-thought reasoning for traceable clinical tumor analysis\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Liet al\.\(2026b\)S\. Li, Z\. Qiu, Z\. Wang, B\. Yun, Z\. Yi, J\. Xu, W\. Zhang, Y\. Xia, and L\. ZhangE\-mrl: cross\-view aligned evidence\-driven multimodal reinforcement learning for reliable 3d tumor analysis\.InMICCAI 2026,Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Lin \(2004\)C\. LinRouge: a package for automatic evaluation of summaries\.InText summarization branches out,pp\. 74–81\.Cited by:[§4\.2](https://arxiv.org/html/2608.19297#S4.SS2.p1.1),[§4\.3](https://arxiv.org/html/2608.19297#S4.SS3.p1.1)\.
- Linet al\.\(2026a\)T\. Lin, Z\. Qiu, J\. Cao, J\. Liu, W\. Yan, B\. Zhang, Y\. Zhong, W\. Zhang, Y\. Xia, and L\. ZhangRegulating anatomy\-aware rewards via trajectory\-integral feedback for volumetric computed tomography analysis\.InProceedings of the 43rd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.306\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Linet al\.\(2026b\)T\. Lin, Z\. Qiu, W\. Zhang, J\. Liu, Y\. Xie, M\. Gao, Z\. Fan, Z\. Li, S\. Li, Z\. Xie, P\. Lu, Y\. Zhuang, L\. Zhang, B\. C\. Ooi, and Y\. XiaOmniCT: towards a unified slice\-volume lvlm for comprehensive ct analysis\.InInternational Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Linet al\.\(2025\)T\. Lin, W\. Zhang, S\. Li, Y\. Yuan, B\. Yu, H\. Li, W\. He, H\. Jiang, M\. Li, S\. Xiaohui, S\. Tang, J\. Xiao, H\. Lin, Y\. Zhuang, and B\. C\. OoiHealthGPT: a medical large vision\-language model for unifying comprehension and generation via heterogeneous knowledge adaptation\.InProceedings of the 42nd International Conference on Machine Learning,pp\. 37975–37995\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Liuet al\.\(2021\)B\. Liu, L\. Zhan, L\. Xu, L\. Ma, Y\. Yang, and X\. WuSlake: a semantically\-labeled knowledge\-enhanced dataset for medical visual question answering\.In2021 IEEE 18th international symposium on biomedical imaging \(ISBI\),pp\. 1650–1654\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- Liuet al\.\(2024a\)C\. Liu, Z\. Wan, C\. Ouyang, A\. Shah, W\. Bai, and R\. ArcucciZero\-shot ecg classification with multimodal learning and test\-time clinical knowledge enhancement\.arXiv preprint arXiv:2403\.06659\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Liuet al\.\(2018\)F\. Liu, C\. Liu, L\. Zhao, X\. Zhang, X\. Wu, X\. Xu, Y\. Liu, C\. Ma, S\. Wei, Z\. He,et al\.An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detection\.Journal of Medical Imaging and Health Informatics8\(7\),pp\. 1368–1373\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.7.1.1)\.
- Liuet al\.\(2024b\)R\. Liu, Y\. Bai, X\. Yue, and P\. ZhangTeach multimodal llms to comprehend electrocardiographic images\.arXiv preprint arXiv:2410\.19008\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Makowskiet al\.\(2021\)D\. Makowski, T\. Pham, Z\. J\. Lau, J\. C\. Brammer, F\. Lespinasse, H\. Pham, C\. Schölzel, and S\. A\. ChenNeuroKit2: a python toolbox for neurophysiological signal processing\.Behavior research methods53\(4\),pp\. 1689–1696\.Cited by:[§3\.2](https://arxiv.org/html/2608.19297#S3.SS2.p2.1)\.
- McKeenet al\.\(2025\)K\. McKeen, S\. Masood, A\. Toma, B\. Rubin, and B\. WangEcg\-fm: an open electrocardiogram foundation model\.Jamia Open8\(5\),pp\. ooaf122\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Moody and Mark \(1983\)G\. B\. Moody and R\. G\. MarkA new method for detecting atrial fibrillation using rr intervals\.Proc\. Comput\. Cardiol\.10,pp\. 227–230\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.14.1.1)\.
- Moody and Mark \(1999a\)G\. B\. Moody and R\. G\. MarkMIT\-BIH Long\-Term ECG Database\.PhysioNet\.Note:Version 1\.0\.0External Links:[Document](https://dx.doi.org/10.13026/C2KS3F),[Link](https://doi.org/10.13026/C2KS3F)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.15.1.1)\.
- Moody and Mark \(1999b\)G\. B\. Moody and R\. G\. MarkMIT\-BIH Normal Sinus Rhythm Database\.PhysioNet\.Note:Version 1\.0\.0External Links:[Document](https://dx.doi.org/10.13026/C2NK5R),[Link](https://doi.org/10.13026/C2NK5R)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.17.1.1)\.
- Moody and Mark \(2001\)G\. B\. Moody and R\. G\. MarkThe impact of the mit\-bih arrhythmia database\.IEEE engineering in medicine and biology magazine20\(3\),pp\. 45–50\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.10.1.1),[§2](https://arxiv.org/html/2608.19297#S2.p3.1)\.
- Mooret al\.\(2023\)M\. Moor, Q\. Huang, S\. Wu, M\. Yasunaga, Y\. Dalmia, J\. Leskovec, C\. Zakka, E\. P\. Reis, and P\. RajpurkarMed\-flamingo: a multimodal medical few\-shot learner\.InMachine learning for health \(ML4H\),pp\. 353–367\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Ohet al\.\(2023\)J\. Oh, G\. Lee, S\. Bae, J\. Kwon, and E\. ChoiEcg\-qa: a comprehensive question answering dataset combined with electrocardiogram\.Advances in Neural Information Processing Systems36,pp\. 66277–66288\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- OpenAI \(2025\)OpenAIGPT\-5 system card\.Note:[https://cdn\.openai\.com/gpt\-5\-system\-card\.pdf](https://cdn.openai.com/gpt-5-system-card.pdf)Cited by:[Appendix B](https://arxiv.org/html/2608.19297#A2.p1.1),[§3\.2](https://arxiv.org/html/2608.19297#S3.SS2.p4.1),[§4\.3](https://arxiv.org/html/2608.19297#S4.SS3.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Panet al\.\(2025\)J\. Pan, C\. Liu, J\. Wu, F\. Liu, J\. Zhu, H\. B\. Li, C\. Chen, C\. Ouyang, and D\. RueckertMedvlm\-r1: incentivizing medical reasoning capability of vision\-language models \(vlms\) via reinforcement learning\.InInternational Conference on Medical Image Computing and Computer\-Assisted Intervention,pp\. 337–347\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Papineniet al\.\(2002\)K\. Papineni, S\. Roukos, T\. Ward, and W\. ZhuBleu: a method for automatic evaluation of machine translation\.InProceedings of the 40th annual meeting of the Association for Computational Linguistics,pp\. 311–318\.Cited by:[§C\.1](https://arxiv.org/html/2608.19297#A3.SS1.p1.1),[§4\.2](https://arxiv.org/html/2608.19297#S4.SS2.p1.1),[§4\.3](https://arxiv.org/html/2608.19297#S4.SS3.p1.1)\.
- Penzelet al\.\(2000\)T\. Penzel, G\. B\. Moody, R\. G\. Mark, A\. L\. Goldberger, and J\. H\. PeterThe apnea\-ecg database\.InComputers in Cardiology 2000\. Vol\. 27 \(Cat\. 00CH37163\),pp\. 255–258\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.21.1.1)\.
- Petrutiuet al\.\(2007\)S\. Petrutiu, A\. V\. Sahakian, and S\. SwirynAbrupt changes in fibrillatory wave characteristics at the termination of paroxysmal atrial fibrillation in humans\.Europace9\(7\),pp\. 466–470\.Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.18.1.1),[§2](https://arxiv.org/html/2608.19297#S2.p3.1)\.
- Ramshaw and Marcus \(1995\)L\. Ramshaw and M\. MarcusText chunking using transformation\-based learning\.InThird workshop on very large corpora,Cited by:[§4\.2](https://arxiv.org/html/2608.19297#S4.SS2.p1.1),[§4\.3](https://arxiv.org/html/2608.19297#S4.SS3.p1.1)\.
- Sellergrenet al\.\(2025\)A\. Sellergren, S\. Kazemzadeh, T\. Jaroensri, A\. Kiraly, M\. Traverse, T\. Kohlberger, S\. Xu, F\. Jamil, C\. Hughes, C\. Lau,et al\.Medgemma technical report\.arXiv preprint arXiv:2507\.05201\.Cited by:[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Tanet al\.\(2022\)S\. Tan, S\. Ortiz\-Gagné, N\. Beaudoin\-Gagnon, P\. Fecteau, A\. Courville, Y\. Bengio, and J\. P\. CohenIcentia11k Single Lead Continuous Raw Electrocardiogram Dataset\.PhysioNet\.Note:Version 1\.0External Links:[Document](https://dx.doi.org/10.13026/kk0v-r952),[Link](https://doi.org/10.13026/kk0v-r952)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.13.1.1)\.
- Tianet al\.\(2024\)Y\. Tian, Z\. Li, Y\. Jin, M\. Wang, X\. Wei, L\. Zhao, Y\. Liu, J\. Liu, and C\. LiuFoundation model of ecg diagnosis: diagnostics and explanations of any form and rhythm on ecg\.Cell Reports Medicine5\(12\)\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Tsutsuiet al\.\(2025\)K\. Tsutsui, S\. Biton Brimer, and J\. BeharSHDB\-AF: a Japanese Holter ECG database of atrial fibrillation\.PhysioNet\.Note:Version 1\.0\.1External Links:[Document](https://dx.doi.org/10.13026/n6yq-fq90),[Link](https://doi.org/10.13026/n6yq-fq90)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.19.1.1)\.
- Wagneret al\.\(2022\)P\. Wagner, N\. Strodthoff, R\. Bousseljot, W\. Samek, and T\. SchaeffterPTB\-XL, a large publicly available electrocardiography dataset\.PhysioNet\.Note:Version 1\.0\.3External Links:[Document](https://dx.doi.org/10.13026/kfzx-aw45),[Link](https://doi.org/10.13026/kfzx-aw45)Cited by:[§1](https://arxiv.org/html/2608.19297#S1.p1.1),[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.4.1.1),[§2](https://arxiv.org/html/2608.19297#S2.p3.1)\.
- Wanet al\.\(2025\)Z\. Wan, C\. Liu, X\. Wang, C\. Tao, H\. Shen, J\. Xiong, R\. Arcucci, H\. Yao, and M\. ZhangMEIT: multimodal electrocardiogram instruction tuning on large language models for report generation\.InFindings of the association for computational linguistics: ACL 2025,pp\. 14510–14527\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Wanget al\.\(2025a\)W\. Wang, Z\. He, W\. Hong, Y\. Cheng, X\. Zhang, J\. Qi, M\. Ding, X\. Gu, S\. Huang, B\. Xu,et al\.Lvbench: an extreme long video understanding benchmark\.InProceedings of the IEEE/CVF International Conference on Computer Vision,pp\. 22958–22967\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- Wanget al\.\(2025b\)W\. Wang, Z\. Gao, L\. Gu, H\. Pu, L\. Cui, X\. Wei, Z\. Liu, L\. Jing, S\. Ye, J\. Shao,et al\.Internvl3\. 5: advancing open\-source multimodal models in versatility, reasoning, and efficiency\.arXiv preprint arXiv:2508\.18265\.Cited by:[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Xuet al\.\(2025\)W\. Xu, H\. P\. Chan, L\. Li, M\. Aljunied, R\. Yuan, J\. Wang, C\. Xiao, G\. Chen, C\. Liu, Z\. Li,et al\.Lingshu: a generalist foundation model for unified multimodal medical understanding and reasoning\.arXiv preprint arXiv:2506\.07044\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1),[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Yakushenko \(2008\)E\. YakushenkoSt Petersburg INCART 12\-lead Arrhythmia Database\.PhysioNet\.Note:Version 1\.0\.0External Links:[Document](https://dx.doi.org/10.13026/C2V88N),[Link](https://doi.org/10.13026/C2V88N)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.11.1.1)\.
- Yanget al\.\(2025\)K\. Yang, M\. Hong, J\. Zhang, Y\. Luo, S\. Zhao, O\. Zhang, X\. Yu, J\. Zhou, L\. Yang, P\. Zhang,et al\.ECG\-lm: understanding electrocardiogram with a large language model\.Health Data Science5,pp\. 0221\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Yuet al\.\(2024\)H\. Yu, P\. Guo, and A\. SanoEcg semantic integrator \(esi\): a foundation ecg model pretrained with llm\-enhanced cardiological text\.arXiv preprint arXiv:2405\.19366\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Yuet al\.\(2023\)H\. Yu, H\. Yang, and A\. SanoECG\-sl: electrocardiogram \(ecg\) segment learning, a deep learning method for ecg signal\.arXiv preprint arXiv:2310\.00818\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p2.1)\.
- Yuet al\.\(2025\)T\. Yu, Z\. Wang, C\. Wang, F\. Huang, W\. Ma, Z\. He, T\. Cai, W\. Chen, Y\. Huang, Y\. Zhao,et al\.Minicpm\-v 4\.5: cooking efficient mllms via architecture, data, and training recipe\.arXiv preprint arXiv:2509\.18154\.Cited by:[§5\.1](https://arxiv.org/html/2608.19297#S5.SS1.p1.1)\.
- Yuanet al\.\(2026\)Y\. Yuan, W\. Li, Z\. Li, Y\. Lin, J\. Li, S\. Tang, J\. Xiao, Y\. Zhuang, and W\. ZhangInstructSAM: segment any instance with any instructions\.arXiv preprint arXiv:2605\.26102\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Yuanet al\.\(2025a\)Y\. Yuan, H\. Zhang, W\. Li, Z\. Cheng, B\. Zhang, L\. Li, X\. Li, D\. Zhao, W\. Zhang, Y\. Zhuang,et al\.Videorefer suite: advancing spatial\-temporal object understanding with video llm\.In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),pp\. 18970–18980\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Yuanet al\.\(2025b\)Y\. Yuan, W\. Zhang, X\. Li, S\. Wang, K\. Li, W\. Li, J\. Xiao, L\. Zhang, and B\. C\. OoiPixelrefer: a unified framework for spatio\-temporal object referring with arbitrary granularity\.arXiv preprint arXiv:2510\.23603\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Zhanget al\.\(2024\)W\. Zhang, T\. Lin, J\. Liu, F\. Shu, H\. Li, L\. Zhang, H\. Wanggui, H\. Zhou, Z\. Lv, H\. Jiang,et al\.Hyperllava: dynamic visual and language expert tuning for multimodal large language models\.arXiv preprint arXiv:2403\.13447\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.
- Zhanget al\.\(2023\)X\. Zhang, C\. Wu, Z\. Zhao, W\. Lin, Y\. Zhang, Y\. Wang, and W\. XiePmc\-vqa: visual instruction tuning for medical visual question answering\.arXiv preprint arXiv:2305\.10415\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p4.1)\.
- Zhenget al\.\(2022\)J\. Zheng, H\. Guo, and H\. ChuA large scale 12\-lead electrocardiogram database for arrhythmia study\.PhysioNet\.Note:Version 1\.0\.0External Links:[Document](https://dx.doi.org/10.13026/wgex-er52),[Link](https://doi.org/10.13026/wgex-er52)Cited by:[Table 1](https://arxiv.org/html/2608.19297#S2.T1.2.1.1.1.1.1.5.1.1)\.
- Zhouet al\.\(2023\)J\. Zhou, X\. He, L\. Sun, J\. Xu, X\. Chen, Y\. Chu, L\. Zhou, X\. Liao, B\. Zhang, and X\. GaoSkinGPT\-4: an interactive dermatology diagnostic system with visual large language model\.arXiv preprint arXiv:2304\.10691\.Cited by:[§2](https://arxiv.org/html/2608.19297#S2.p1.1)\.

Appendix

This is the appendix for “Holtercare\-Bench: A Multimodal Benchmark for Evaluating Long\-Term Dynamic ECG Analysis”\.

This appendix is organized as follows:

- •Section[A](https://arxiv.org/html/2608.19297#A1)provides the detailed distribution and frequencies of the clinical annotations within Holtercare\-23K\.
- •Section[B](https://arxiv.org/html/2608.19297#A2)presents the specific prompt templates utilized for LLM\-based QA generation and report evaluation\.
- •Section[C](https://arxiv.org/html/2608.19297#A3)details supplemental generation metrics,BLEUn\-gram overlap metrics, and additional analyses on modality alignment and statistical significance of fine\-tuning improvements\.

## Appendix ADetailed Distribution of Clinical Annotations

While Figure[3](https://arxiv.org/html/2608.19297#S3.F3)contains an overview of the most frequent categories, to provide a deeper understanding of the scale and clinical diversity of Holtercare\-23K, we present the exhaustive frequencies of our expert\-verified beat and rhythm annotations\.

Table[7](https://arxiv.org/html/2608.19297#A1.T7)details the occurrence counts for all beat\-level annotations\. This includes a massive scale of valid QRS complexes—such as normal beats \(N\), atrial fibrillation beats \(Af\), and ventricular premature beats \(V\)—along with non\-QRS elements and artifacts\.

Table 7\.Frequency of beat\-level annotations across Holtercare\-23K\.LabelDescriptionCountValid QRS ComplexesNNormal Beat71,429,269AfAtrial Fibrillation4,822,970SAtrial Premature Beat \(APB\)1,042,095PUnclassified Pacing840,298VVentricular Premature Beat \(VPB\)779,227SeAtrial Escape Beat587,037AFAtrial Flutter490,579JeJunctional Escape Beat179,605BBundle Branch Block2,051VeVentricular Escape Beat1,969JJunctional Premature Beat182?Questionable Beat30FFusion Beat1aPAtrial Single\-Chamber Pacing0vPVentricular Single\-Chamber Pacing0dPDual\-Chamber Pacing0AbAPB with Aberrant Ventricular Conduction0VaAberrant Ventricular Conduction0Non\-QRS ElementsXArtifact833,824SaNon\-conducted APB32,645pP\-wave0tT\-wave0Table[8](https://arxiv.org/html/2608.19297#A1.T8)outlines the comprehensive frequencies of the rhythm\-level annotation\. These capture complex, continuous cardiac events and diverse arrhythmias ranging from common premature atrial contractions \(PAC\) to severe, life\-threatening events like ventricular fibrillation \(VF\) and asystole\.

Table 8\.Frequency of rhythm\-level annotations across Holtercare\-23K\.EventCountEventCountPremature Atrial Contraction \(PAC\)888Sinus Arrest6Premature Ventricular Contraction \(PVC\)775Atrial Flutter \(AFL\)6PAC Couplets396Atrial Escape Rhythm4Atrial Tachycardia \(AT\)339Wandering Atrial Pacemaker \(WAP\)4PVC Couplets152Ventricular Escape Rhythm3Ventricular Tachycardia \(VT\)94Atrial Undersensing2PAC Bigeminy82T\-Wave Alternans \(TWA\)2PAC Trigeminy64Ventricular Capture Management \(VCM\)3PVC Trigeminy58Premature Junctional Contraction \(PJC\)2PVC Bigeminy55Atrial Capture Management \(ACM\)2Asystole44Defibrillation2Sinus Arrhythmia44ST Segment Changes2Long R\-R Interval33Accelerated Atrial Escape Rhythm1Ventricular Fibrillation \(VF\)23Anti\-Tachycardia Pacing \(ATP\)1Second\-Degree Atrioventricular \(AV\) Block15Ventricular Dissociation1Junctional Escape Rhythm12OptiVol Fluid Status1First\-Degree Atrioventricular \(AV\) Block9Supraventricular Tachycardia \(SVT\)1Accelerated Idioventricular Rhythm \(AIVR\)6Ventricular Sense Response \(VSR\)1Ventricular Flutter \(VFL\)6Chest Compression Waveform1Atrial Fibrillation \(AF\)6
## Appendix BPrompt for QA Generation and Report Evaluation

As outlined in our methodology, we use GPT\-5\-mini\([42](https://arxiv.org/html/2608.19297#bib.bib11)\)for both the automated construction of complex reasoning tasks and the LLM\-as\-a\-judge evaluation of long\-context reports\. To ensure transparency and reproducibility, we provide the exact representative prompts used in these two pipelines\.

QA Generation\.Figure[6](https://arxiv.org/html/2608.19297#A2.F6)displays a representative structured prompt utilized by HolterAgent to generateOpen\-QApairs, specifically showcasing theEvidence Reasoningsub\-task\. By feeding the model explicit patient information, the prompt strictly forces the LLM to extract timing, rhythm, and morphology clues directly from the provided text, effectively mitigating hallucinated waveform features\.

QA Generation PromptYour task is to generate high\-fidelityOpen\-QAdata inEvidence Reasoningtasks for Holter ECG diagnostic training\. Given a patient’s structured information, you must generate exactly one QA pair and construct a rigorous, evidence\-based clinical reasoning process that verifies the target diagnosis based solely on the provided inputs\. Zero hallucination is tolerated\. Adhere strictly to requirements below\.Patient Information•Age:\{\{AGE\}\}•Gender:\{\{GENDER\}\}•Electronic Medical Records:\{\{EMR\}\}•Beat\-Level Annotations:\{\{BEAT\_ANNOTATIONS\}\}•Rhythm\-Level Annotations:\{\{RHYTHM\_ANNOTATIONS\}\}•Report\-Level Summaries:\{\{REPORT\_SUMMARY\}\}•Target Disease:\{\{TARGET\_DISEASE\}\}Requirements\(1\)Requirements forquestions:•Introduce the patient’sageandgender\(e\.g\.,“The current monitoring segment comes from a 61\-year\-old male\.”\), explicitly mention the exacttarget disease, and ask whether it is present on the visible ECG waveform\.•Ask the reader to specify which waveform features support that determination \(e\.g\., prompting for rhythm clues, timing, P\-wave morphology, PR\-interval behavior, or QRS characteristics\)\.•Vary the phrasing and sentence structure naturally, butdo not change the core verification task\.\(2\)Requirements foranswers:•Begin by directly confirming the presence of the target disease \(e\.g\.,“Yes,\{\{TARGET\_DISEASE\}\}is present\.”or“Confirmed\.”\)\.•Provide a structured clinical reasoning process by summarizing clues extracted from the provided annotations and summaries\. Organize the reasoning into logical clinical dimensions where applicable, such as:–Timing / Rhythm Clues:Discuss rates, RR intervals, coupling intervals, and event burden based on the report and beat/rhythm Annotations\.–Morphology Clues:Describe wave shapes \(P, QRS, T, ST\-segment deviations\) and AV conduction characteristics relevant to the disease\.•Base all explanations solely on the provided information\.Do not hallucinate waveform features that cannot be inferred or documented from these inputs\.\(3\)The output must be strictly in the form of JSON:``` { "question": "...", "answer": "..." } ```

Figure 6\.Prompt for QA pairs generation inEvidence Reasoningtasks ofOpen\-QA\.Report Evaluation\.Figure[7](https://arxiv.org/html/2608.19297#A2.F7)presents a representative prompt design for our LLM\-as\-a\-judge evaluation system, specifically showcasing theGeneral Summarysub\-task ofReport Generation\. Acting as a strict clinical expert, the prompt evaluates the generated output against the ground truth reference across 17 granular criteria\. The specific fine\-grained criteria evaluated by the LLM judge are detailed previously in Table[2](https://arxiv.org/html/2608.19297#S4.T2)forStatistical Overviewand Table[3](https://arxiv.org/html/2608.19297#S4.T3)forGeneral Summary\. The prompt is explicitly designed to severely deduct points for fabricated information, missing values, or hallucinated clinical diagnoses\.

Report Evaluation PromptYou are a strict and professional cardiac expert grading an LLM\-generated Holter ECG report against a ground truth reference\. Actively look for missing numbers, wrong values, fabricated information, and formatting issues\. Deduct points severely for any discrepancy\.Do not default to 100\.Reference Report:\{\{REFERENCE\_REPORT\}\}Generated Report:\{\{GENERATED\_REPORT\}\}Evaluation Criteria\(1\)Accurate extraction of global metrics\(2\)Accurate reporting of HR extremes\(3\)Correct extraction of tachycardia and bradycardia burdens\(4\)Correct classification of the baseline rhythm \(e\.g\., AFib\)\(5\)Accurate total count and burden of PACs/PVCs\(6\)Precise breakdown of ectopic patterns \(e\.g\., isolated, paired, runs\)\(7\)Complete inclusion of severe events \(e\.g\., VT runs, VF, asystole\)\(8\)Correct interpretation of conduction blocks or pacing signals\(9\)Correct reporting of ST\-segment and T\-wave changes\(10\)Strict consistency with clinical facts without hallucinated findings\(11\)Objective description without exaggeration or understatement\(12\)Clinical diagnoses strictly grounded in supporting statistical data\(13\)Coherent narrative structure without fragmented data enumeration\(14\)Structured clinical hierarchy, prioritizing major diagnoses over secondary findings\(15\)Precise application of medical terminology\(16\)Consistent numeric formatting and standardized unit application \(e\.g\., “bpm” for rate, “%” for burden\)\(17\)Inclusion of appropriate clinical caveatsRequirements\(1\)Score each item from 0 to 100 based on the degree of compliance with the specific criterion\.•Perfect \(90–100\):Perfect or near\-perfect compliance with this specific criterion\.•Substantial \(70–89\):Substantial compliance, with minor flaws, slight inaccuracies, or trivial omissions\.•Partial \(40–69\):Captures the relevant clinical concept but applies an incorrect metric type or statistical aggregation\.•Poor \(10–39\):Barely related or severely flawed, but not completely blank\.•Failure \(0–9\):Complete failure; the concept is entirely missing or severely hallucinated\.\(2\)If a specific clinical finding \(e\.g\., ectopic patterns, severe events, ST\-segment changes\) is not mentioned in the reference report, it means the patient does not have it\.•If the generated report correctly omits it as well, this is a perfect match for that criterion\.Do not deduct points for missing a condition that isn’t in the reference\.•However, if the reference report does not have it, but the generated report fabricates or hallucinates it, you must severely deduct points\.\(3\)The output must be a valid JSON object where keys are the item numbers \(“1” to “17”\) and values are purely numerical scores from 0 to 100\.Do not include any other text\.For example:``` { "1": 95, "2": 80, ..., "17": 100 } ```

Figure 7\.Prompt for report evaluation inGeneral Summarytasks ofReport Generation\.
## Appendix CSupplemental Experimental Results

### C\.1\.Supplemental Generation Metrics

While the main manuscript primarily focuses on clinical accuracy, specialized semantic metrics likeF1\-Bioand automated LLM judgeScoreGPTto assess reasoning capabilities, standardBLEU\([44](https://arxiv.org/html/2608.19297#bib.bib34)\)n\-gram overlap metrics provide a useful supplementary perspective on textual generation quality\.

Table[9](https://arxiv.org/html/2608.19297#A3.T9)presentsBLEU\-1andBLEU\-4scores achieved by the evaluated baseline and fine\-tuned models onOpen\-QAtasks, specifically evaluatingEvent Timing,Diagnosis, andEvidence Reasoningsub\-tasks\.

Table 9\.Supplementary performance comparison of evaluated models onOpen\-QAtasks from Holtercare\-Bench, evaluated by additionalBLEUn\-gram overlap metrics\.ModelModalityEvent TimingDiagnosisEvidence ReasoningBLEU\-1↑\\uparrowBLEU\-4↑\\uparrowBLEU\-1↑\\uparrowBLEU\-4↑\\uparrowBLEU\-1↑\\uparrowBLEU\-4↑\\uparrowGeneralist ModelsGPT\-5\-mini37\.5010\.838\.701\.9696\.6738\.48Claude\-4\.5\-Haiku11\.681\.256\.980\.5557\.538\.97Phi\-4\-mini\-3\.8B53\.5713\.2253\.8515\.7396\.4955\.56Phi\-4\-mini\-3\.8BFT95\.0077\.3998\.4687\.0593\.7072\.64InternVL\-3\.5\-8B41\.468\.6535\.296\.0296\.3030\.39MiniCPM\-V4\.5\-8B16\.332\.0615\.791\.2591\.4935\.50Gemini\-3\.0\-Flash66\.6722\.9317\.652\.8796\.8845\.06Qwen3\-VL\-8B56\.6712\.6850\.000\.0095\.3552\.34Qwen3\-VL\-8BFT60\.9531\.9130\.7712\.3697\.1450\.30Medical ModelsLLaVA\-Med\-V1\.5\-7B37\.257\.5037\.505\.4586\.0542\.02MedGemma\-1\.5\-4B\-IT22\.222\.9016\.002\.1397\.6265\.55HealthGPT\-M3\-3\.8B34\.216\.1912\.361\.2490\.2443\.46Lingshu\-7B33\.335\.6125\.003\.9897\.6747\.90MedVLM\-R1\-2B40\.546\.6033\.334\.5085\.9623\.23HuatuoGPT\-Vision\-7B53\.3321\.0233\.338\.0598\.328\.16Table[10](https://arxiv.org/html/2608.19297#A3.T10)details the correspondingBLEU\-1andBLEU\-4metrics forReport Generationtasks, covering bothStatistical OverviewandGeneral Summary\. These supplemental metrics highlight the lexical and structural alignment between the models’ generated responses and the expert\-crafted ground truth\.

Table 10\.Supplementary performance comparison of evaluated models onReport Generationtasks from Holtercare\-Bench, evaluated by additionalBLEUn\-gram overlap metrics\.ModelModalityStatistical OverviewGeneral SummaryBLEU\-1↑\\uparrowBLEU\-4↑\\uparrowBLEU\-1↑\\uparrowBLEU\-4↑\\uparrowGeneralist ModelsGPT\-5\-mini44\.643\.6560\.476\.95Claude\-4\.5\-Haiku15\.921\.1527\.601\.00Phi\-4\-mini\-3\.8B56\.7610\.9952\.436\.03Phi\-4\-mini\-3\.8BFT97\.6768\.0394\.8371\.94InternVL\-3\.5\-8B81\.6342\.2071\.4314\.88MiniCPM\-V4\.5\-8B42\.556\.9561\.9612\.88Gemini\-3\.0\-Flash51\.115\.4867\.079\.86Qwen3\-VL\-8B34\.758\.5341\.133\.22Qwen3\-VL\-8BFT47\.314\.8850\.564\.73Medical ModelsLLaVA\-Med\-V1\.5\-7B45\.763\.5358\.738\.09MedGemma\-1\.5\-4B\-IT59\.629\.4246\.813\.94HealthGPT\-M3\-3\.8B10\.791\.4752\.6310\.27Lingshu\-7B38\.827\.5659\.8210\.40MedVLM\-R1\-2B68\.2918\.4260\.007\.18HuatuoGPT\-Vision\-7B37\.634\.7352\.147\.05
### C\.2\.Modality Alignment and Representation Analysis

To accommodate the diverse architectural constraints of contemporary MLLMs, our data engine, HolterAgent, constructs a format compatibility pipeline that transforms a single, unified\.edfsignal source into three distinct representation formats: raw signal, video stream, and clinical text\. This design is fundamentally intended to maximize MLLM compatibility rather than to introduce three independent data sources\. Specifically, the video modality is prioritized for its ability to preserve the continuous temporal dynamics and morphological evolution of streaming electrophysiological data\. Conversely, the text modality serves as a robust fallback for models lacking native video processing capabilities, representing the signal through discretized numerical sequences\.

To empirically verify that this multi\-representation design does not introduce modality\-specific artifacts or bias the evaluation, we conduct a comprehensive comparative analysis\. Table[11](https://arxiv.org/html/2608.19297#A3.T11)presents theClosed\-QAperformance of all video\-capable models when evaluated exclusively under the text modality\. The results demonstrate that neither modality consistently dominates across all models and tasks\. For instance, while certain models exhibit marginal improvements in specific tasks under the text modality, others show a clear preference for video inputs\. This non\-uniform performance distribution confirms that the observed performance gaps among different MLLMs reflect their intrinsic architectural capabilities and inductive biases in processing long\-term physiological sequences, rather than artifacts induced by our data formatting pipeline\.

Table 11\.Supplementary performance comparison of video\-capable models evaluated under the text modality onClosed\-QAtasks from Holtercare\-Bench, evaluated by accuracy\.ModelModalityPresenceEventCountingEventTimingHR ExtremumTimingDiagnosisGeneralist ModelsInternVL\-3\.5\-8B58\.0129\.6647\.2631\.8539\.37MiniCPM\-V4\.5\-8B54\.9051\.3828\.0535\.9926\.38Gemini\-3\.0\-Flash67\.3214\.9862\.5053\.1841\.73Qwen3\-VL\-8B60\.9531\.8028\.6634\.7132\.28Qwen3\-VL\-8BFT95\.1083\.7997\.2371\.9794\.88Medical ModelsLingshu\-7B47\.5535\.1724\.7029\.9425\.98MedVLM\-R1\-2B40\.2041\.2826\.5232\.1713\.39HuatuoGPT\-Vision\-7B52\.2925\.3826\.2232\.1723\.23
### C\.3\.Statistical Significance Analysis

To confirm whether the substantial performance improvements observed after fine\-tuning are statistically robust and not attributable to random sampling variance or favorable test\-set splits, we perform McNemar’s tests on all fiveClosed\-QAtasks for both fine\-tuned representative models, Phi\-4\-mini\-3\.8B\([1](https://arxiv.org/html/2608.19297#bib.bib13)\)and Qwen3\-VL\-8B\([3](https://arxiv.org/html/2608.19297#bib.bib1)\)\. As detailed in Table[12](https://arxiv.org/html/2608.19297#A3.T12), all pairwise comparisons between the zero\-shot baselines and fine\-tuned models yieldpp\-values≪0\.001\\ll 0\.001\. These statistical significances across all tasks and both models provide evidence that the fine\-tuning process induces genuine capability acquisition and knowledge internalization in long\-context ECG reasoning, rather than mere fluctuations in sampling variance\.

Table 12\.Statistical significance analysis of improvements from zero\-shot baselines to fine\-tuned models onClosed\-QAtasks from Holtercare\-Bench\. The reportedpp\-values are derived from McNemar tests\.ModelModalityPresenceEventCountingEventTimingHR ExtremumTimingDiagnosisPhi\-4\-mini\-3\.8B1\.95×10−51\.95\\times 10^\{\-5\}3\.99×10−163\.99\\times 10^\{\-16\}6\.33×10−206\.33\\times 10^\{\-20\}3\.09×10−83\.09\\times 10^\{\-8\}1\.76×10−111\.76\\times 10^\{\-11\}Qwen3\-VL\-8B6\.17×10−416\.17\\times 10^\{\-41\}5\.26×10−325\.26\\times 10^\{\-32\}2\.18×10−522\.18\\times 10^\{\-52\}6\.57×10−186\.57\\times 10^\{\-18\}3\.55×10−353\.55\\times 10^\{\-35\}Qwen3\-VL\-8B7\.08×10−447\.08\\times 10^\{\-44\}3\.14×10−363\.14\\times 10^\{\-36\}6\.14×10−476\.14\\times 10^\{\-47\}1\.08×10−171\.08\\times 10^\{\-17\}1\.30×10−281\.30\\times 10^\{\-28\}