ECG-Expert-QA: A Benchmark for Evaluating

Medical Large Language Models in Heart Disease

Diagnosis
Xu Wang1,2 , Jiaju Kang2,3,* , Puyu Han2,4

1
Shandong Jianzhu University 2 FUXI AI Lab 3 Beijing Normal University
4
Southern University of Science and Technology
arXiv:2502.17475v2 [eess.SP] 26 Feb 2025

Abstract—We present ECG-Expert-QA, a comprehensive mul-

1357
timodal dataset designed for evaluating diagnostic capabilities 1164
in ECG interpretation, integrating real clinical data with sys-
tematically generated synthetic cases. The dataset encompasses 2931

six fundamental diagnostic tasks, comprising 47,211 meticulously 2931

2623
curated question-answer pairs that span a spectrum of clinical
o 1466
scenarios, from basic rhythm analysis to complex case interpreta-
2931
tion. By simulating challenging clinical cases through a rigorous
PRTK
2931
medical knowledge-guided process, ECG-Expert-QA notPP only 2247
enhances the availability of annotated diagnostic data but MC also
20032

significantly increases the complexity and diversity of MROD

clinical 1931

presentations, including rare cardiac conditions and temporal

MEE
2592
progression patterns. This design enables comprehensive MCevalu-
ation of medical language models across multiple dimensions,
LTD 10000 7500 5000 2500 0 0 5000 10000 15000 20000
including diagnostic accuracy, clinical reasoning, and knowledge
GRA PRTK : Patient's right to know MEE : Medical entity extraction EBK : ECG Basic Knowledge
PP : Patient prognosis MCo : Medical counterfactual CMD : Cross modal diagnosis
integration. To facilitate global research collaboration, EBKECG- MROD : Multiple rounds of dialogue LTD : Long text diagnosis CD : Complex diagnosis
Expert-QA is available in both Chinese and English versions,
CMD MC : Memory correction.json GRA : Generate risk assessment CK : Cardiology Knowledge

with rigorous quality control ensuring linguistic and clinical

CD

consistency. The dataset’s challenging diagnostic tasks, CKwhich Fig. 1. ECG-Expert-QA data distribution
include interpretation of complex arrhythmias, identification of
subtle ischemic changes, and integration of clinical context,
establish it as an effective benchmark for advancing AI-assisted
To address these challenges, the ECG-Expert-QA dataset
ECG interpretation and pushing the boundaries of current
diagnostic models. Our dataset is open-source and available at developed in this study achieves multidimensional innovative
https://github.com/Zaozzz/ECG-Expert-QA. breakthroughs. Through a hierarchical architecture design,
Index Terms—Benchmark Dataset, ECG Medical Diagnostics, the dataset systematically integrates three core evaluation
Evaluation of Medical LLMs , Automated Generation Pipeline modules: the basic medical knowledge verification module
contains 4,839 samples, covering fundamental cardiovascular
I NTRODUCTION theory (2,592 items) and specialized ECG principles (2,247
This study systematically investigates the technological ad- items); the clinical reasoning evaluation module includes
vancements and evaluation challenges of multimodal large lan- 20,032 cross-modal diagnostic cases, focusing on the model’s
guage models (LLMs) in the field of intelligent electrocardio- comprehensive decision-making capabilities in complex clin-
gram (ECG) diagnosis, proposing innovative solutions. Current ical scenarios, including patient prognosis prediction (2,931
research faces three systemic challenges: first, traditional man- cases), multimodal information integration (1,931 cases), and
ual evaluation methods suffer from efficiency bottlenecks, with dynamic medical context reasoning (2,931 cases); the risk
their resource-intensive nature, subjective dependency, and control module innovatively introduces medical ethics dimen-
lack of scalability severely constraining large-scale model vali- sions, establishing a decision safety assessment benchmark
dation; second, existing datasets exhibit significant deficiencies for medical AI through 2,931 ethical conflict cases and 1,357
in case complexity, clinical scenario coverage, and multimodal informed consent scenarios.
reasoning capability assessment; third, linguistic homogeneity The innovation of this study is primarily manifested in
hinders the validation of applicability in cross-cultural medical three aspects: First, it proposes a standardized paradigm of
settings, particularly lacking in-depth characterization of rare ”evaluation as a service,” enabling quantitative assessment
pathological conditions and ethical risks. of model capabilities through structured task design, with
complex diagnostic tasks accounting for 15.3% of total sam-
Corresponding author: [email protected] ples, significantly higher than traditional medical datasets.
 Reports in MIMIC-IV-ECG Twelve lead ECG
June 19, 2199: Sinus rhythm, prolonged QT interval...
June 26, 2199: Atrial fibrillation with rapid ventricular response...
June 30, 2199: Atrial fibrillation, possible anteroseptal infarct...
August 5, 2199: Atrial fibrillation, possible septal infarct...
March 25, 2200: Sinus bradycardia, possible inferior infarct...

DeepSeek-R1

Medical Description Text

Patient 15296609 underwent five electrocardiograms (ECGs) between June
19, 2199, and March 25, 2200, showing various cardiac abnormalities. The
first examination (June 19, 2199) revealed sinus rhythm, prolonged QT
interval, left atrial abnormality, possible anteroseptal infarct , and low QRS
voltages. The second examination (June 26, 2199) showed atrial
fibrillation with rapid ventricular response, possible septal infarct, and low
QRS voltages. The third examination (June 30, 2199) also showed atrial
fibrillation, possible anteroseptal infarct, and low QRS voltages. The fourth
examination (August 5, 2199) revealed atrial fibrillation, possible septal
infarct, low QRS voltages, and lateral ST-T changes. The final
examination (March 25, 2200) showed sinus bradycardia, possible inferior
infarct, low QRS voltages, and possible left ventricular hypertrophy (LVH).
All examinations showed abnormalities, and further evaluation and
monitoring are recommended.

ChatGPT - 4o

Fig. 2. ECG-Expert-QA generate route

Second, the construction of a bilingual corpus breaks language important methodological references for building trustworthy
barriers, providing a comparable benchmark for cross-cultural medical AI evaluation systems, and its evaluation framework
medical AI research, though the distribution characteristics of design principles can be extended to other specialized medical
specific language pairs require further clarification in subse- fields.
quent studies. Third, the introduction of innovative evaluation
dimensions such as counterfactual reasoning (1,466 cases) and R ELATED W ORK
memory correction mechanisms (2,931 cases) effectively tests The MIT-BIH [1] Arrhythmia Database is an early represen-
the clinical rationality and robustness of model decision logic. tative dataset, containing 48 half-hour dual-channel dynamic
Notably, the dataset exhibits significant interdisciplinary ECG recordings from 47 patients, covering various types
characteristics at the technical implementation level. By in- of arrhythmias. The MIT-BIH database is not only the first
tegrating the multimodal representation capabilities of natural publicly released standardized ECG test dataset, but also
language processing with the specialized knowledge system of specifically used for evaluating the performance of arrhythmia
cardiovascular medicine, it not only covers 20,032 cross-modal detection systems. All records were independently annotated
diagnostic samples but also innovatively designs medical entity by multiple cardiologists. Although the MIT-BIH database has
extraction tasks (2,931 cases), providing new insights for only 48 records and 47 patients, there are certain limitations in
the development of explainable medical AI. In the ethical data scale, time span, testing equipment, and sample diversity.
dimension, the dataset specifically allocates 2.7% of samples However, the MIT-BIH database laid an important foundation
for evaluating model decision safety in medical risk scenarios, for the development of subsequent ECG datasets, promoting
representing a nearly 300% increase compared to conventional the release of ECG datasets for various application scenarios
medical evaluation standards. and expanding needs.
Future research directions should focus on enhancing the The MIMIC-IV [2], [3] dataset compensates for the limi-
dynamic evaluation capabilities of the dataset, particularly tations of MIT-BIH in terms of data scale, time span, testing
strengthening validation in temporal data analysis and inte- equipment, and sample diversity. This dataset includes a large
gration with real clinical workflows. Subsequent studies are amount of physiological data and diagnostic information from
recommended to establish quality control protocols for case critically ill patients, although it lacks normal samples, it holds
collection, clarify specific implementation plans for multi- extremely high application value in clinical research. The
lingual support, and explore domain adaptation mechanisms PTB-XL [4], [5] dataset was constructed by Patrick Wagner
within continuous learning frameworks. This study provides and others, containing 21,837 12-lead ECG records, covering
various cardiovascular pathological states, and supplementing depth of model applications.
the normal samples missing in MIMIC-IV. Furthermore, PTB-
M ETHODS
XL provides demographic information and signal quality met-
rics of patients, greatly increasing the quantity and diversity Data Screening
of ECG data. The text descriptions in the dataset come from the MIMIC-
In MIMIC-IV-V3 [2], [3], the dataset includes 364,627 IV-ECG [26] dataset and have been filtered based on the
patients and 94,458 ICU admission records, marking a break- dataset’s CSV annotations. The specific steps are as follows:
through in the volume of ECG datasets. However, despite • Similarity Calculation: Cosine similarity from
these datasets providing valuable resources for automated di- sklearn.metrics.pairwise [27], [28] is used to compute
agnostic models like CNNs [6]–[8] and Transformers [9]–[14], text similarity, and data with a similarity greater than
supervised learning methods, which rely on large amounts of 0.8 are excluded to enhance the dataset’s diversity.
annotated data, still face many challenges, especially in terms • Report Quantity Screening: Reports with fewer than four
of generalization across different patient populations and rare records are excluded to ensure the completeness of the
ECG patterns, which are significant limitations in real-world information, thereby enhancing the model’s judgment and
clinical applications. training capabilities.
In recent years, multimodal learning has emerged as a pow- • Patient Case Summary: The dataset is organized by
erful paradigm in the medical field, successfully combining patient, with all cases for each patient being aggregated.
medical imaging and textual information. To evaluate and train Only patients with more than four cases are selected, cre-
large multimodal models, datasets like MEDQA [15], created ating comprehensive patient profiles. This method allows
by Di Jin and others, have been constructed based on medical the model to analyze data more thoroughly, considering
exam question banks to test the model’s medical reasoning ECG variations at different time points, thus improving
ability in multiple-choice formats, enriching the diversity of its analytical capability [29].
medical question-answering tasks. However, the application of In the end, 2,931 records were selected from 20,000 for subse-
existing ECG datasets in natural language question answering quent generation tasks. After processing, the dataset’s diversity
and multimodal detection remains limited. Although Jungwoo and completeness have been enhanced. This allows the model
Oh and others constructed the ECG-QA [16] dataset based on to train on more representative samples, considering ECG
PTB-XL [17], [18] and introduced initial question-answering variations and patient background information, thus improving
features, its question-answering system still largely relies on its judgment and diagnostic accuracy while strengthening its
framework filling, and the number of question settings is application potential in complex clinical scenarios.
relatively limited. While its design meets the needs of large
models for question-answering formats, it still falls short in DATASET G ENERATION
handling complex diagnostic problems and simulating real This study proposes three main dataset generation methods:
interactive scenarios. expert knowledge-guided professional knowledge assessment,
To fill this gap, the ECG-Chat [19] and CardioGPT [20], cross-modal diagnosis in complex medical environments, and
which uses a large scale 12 lead electrocardiogram database medical risk assessment. Specifically, the expert knowledge-
[21], [22] for arrhythmia study dataset proposed by Guohua guided method aims to generate professional question-answer
Fu and others focuses on ECG diagnostic tasks, offering large- pairs by guiding large models with specialized knowledge;
scale, diversity, and high-quality expert annotations, as well as cross-modal diagnosis transforms semi-structured data such
the use of wavelet scattering networks for feature extraction as electrocardiograms (ECG) into textual descriptions to build
and the ability to support multi-label classification and com- diagnostic-related question-answer pairs; medical risk assess-
plex ECG pattern detection. However, it still lacks multi-round ment utilizes medical diagnostic results to generate question-
dialog-based question-answering features. To address this, the answer pairs involving medical counterfactuals, ethical harms,
ECG-Expert-QA dataset innovatively integrates multi-round and patient informed rights. Each method employs different
dialogue structures and patient background information, en- input forms and interaction strategies with the model to
abling long-text reasoning in complex clinical scenarios. ECG- ensure that the generated datasets effectively support tasks
Expert-QA not only introduces the first question-answering such as knowledge learning in the medical field, pathologi-
scenario for simulating ECG diagnostic text generation but cal reasoning, and multi-turn dialogues. These methods not
also fills the gap in traditional ECG datasets in terms of only significantly enhance the application capabilities of large
question-answer interaction functionality. It provides unprece- models in specialized fields but also improve the system’s
dented support for intelligent diagnosis and dynamic inter- adaptability and accuracy in complex medical environments.
action tasks. This dataset, centered around question-answer
interactions, deeply integrates multi-round dialogues, patient Expert Knowledge-Guided Professional Knowledge
background, memory correction, and patient rights [23]–[25], This method builds a corpus of medical question-answer
breaking through the limitations of single-diagnosis labels pairs by conducting continuous dialogues with multiple exist-
and pioneering a new type of intelligent ECG dataset in the ing high-quality large language models (LLMs), such as GPT-
question-answering field, greatly expanding the breadth and 4. The specific process includes deploying local large models
or utilizing relevant large model APIs, combining professional • Medical Consultation Simulation: Simulates multi-round
expert knowledge information, and using the model’s pro- interactions between doctors and patients where patients
grammability and role-based prompting mechanisms to create ask ECG-related questions and doctors provide profes-
simulated doctor-patient dialogue scenarios, thereby forming sional explanations, forming dialogue datasets beneficial
a professional medical knowledge dataset. for medical consultations.
• Basic Cardiology: Covers fundamental medical knowl- • Memory Adjustment: Tests the model’s ability to modify
edge of heart diseases, including cardiac anatomy and early diagnoses or explanations based on new infor-
function, common types of heart diseases (such as coro- mation, resolving ambiguities and enhancing memory
nary artery disease, heart failure), etiology, symptoms, adjustment in dialogues.
clinical diagnosis, and treatment methods. • Cardiac Entity Extraction: Uses natural language process-
• Electrocardiogram (ECG) Expertise: Focuses on ECG ing techniques to extract key information from textual
analysis, including basic principles, waveform analysis, descriptions, such as patient history, ECG abnormalities,
identification of common abnormalities (such as ven- and symptoms, creating structured datasets.
tricular premature beats, atrial fibrillation), and clinical This method generates cross-modal datasets in the medical
applications of ECG. field through image descriptions, with a particular focus on
• Complex Disease Analysis: Generates detailed analysis ECG image description and analysis. By converting processed
reports for complex cases by simulating multi-turn dia- semi-structured ECG image data into textual descriptions and
logues between doctors and patients or medical profes- combining the model’s programmability, role-based prompt-
sionals to formulate diagnostic and treatment plans. These ing, and text generation strategies, datasets covering various
dialogues not only accurately analyze symptoms but application scenarios are formed. These datasets support tasks
also provide in-depth decision support for personalized such as AI-assisted diagnosis, reasoning analysis, and medical
treatment. consultations.
By simulating diverse interactions between doctors and pa-
tients, professional medical question-answer dialogues are Medical Risk Assessment
generated, emphasizing the reduction of repetitive outputs Addressing existing rights issues (such as patient informed
and the enhancement of dataset diversity to ensure a rich consent rights) and medical counterfactuals and ethical issues
corpus with clinical application value. The final generated in diagnostic reports, different dialogue methods are used for
dataset can serve as high-quality training data for medical dataset generation (multi-turn dialogues for patient informed
question-answering and intelligent diagnostic systems, as well consent rights; single-turn dialogues for medical counterfac-
as support medical and patient education. tuals and medical ethics). The specific steps include using
Cross-Modal Diagnosis in Complex Medical Environments local models or API inputs with specified role-based prompts
First, preprocess semi-structured data of ECG images (such (such as doctor, patient) and leveraging the model’s language
as 12-lead ECG) and convert them into detailed textual processing capabilities to generate relevant outputs. The main
descriptions that explain waveform characteristics, abnormal application datasets include:
signals, related clinical symptoms, and potential diagnostic • Patient Informed Consent Rights: Emphasizes patients’
opinions. This step ensures a close association between images fundamental rights in medical decision-making, including
and text, laying the foundation for subsequent dataset gener- the concept of informed consent, legal and ethical issues
ation. Next, use local models or API inputs with specified surrounding patient autonomy, and communication norms
role-based prompts (such as doctor, patient) and leverage the between doctors and patients.
model’s powerful language processing capabilities to generate • Medical Counterfactuals: Hypothetically changes medical
high-quality diagnostic information, forming diverse datasets. decisions (such as treatment plans, medication dosages)
Specific applications include: to analyze their potential impacts on patient health, pro-
• Cross-Modal Diagnosis: Combines ECG images with tex- viding insights for clinical decision-making.
tual descriptions to explore the relationships between var- • Medical Ethics: Ensures that generated content complies

ious abnormal ECG waveforms and potential diagnostic with ethical standards, assesses risks, and avoids mislead-
opinions, covering common ECG pathological changes ing or providing inaccurate medical advice.
such as arrhythmias and myocardial infarction. The medical risk assessment module generates high-
• Patient Prognosis Analysis: Generates data for prognosis quality question-answer pairs covering patient rights, medical
analysis by describing the current cardiac health status decision-making, and ethical standards by simulating various
and predicting disease progression or recovery periods dialogue scenarios. These datasets not only help train and
based on abnormal signals in the images. optimize intelligent medical systems, ensuring higher accuracy
• Complex Patient Background Reasoning: Analyzes the and reliability when handling complex ethical and legal issues,
impact of multiple factors on patient health through long but also provide rich educational resources for medical pro-
textual descriptions, such as how chronic diseases like fessionals, enhancing their decision-making capabilities and
hypertension or diabetes affect ECG interpretation. communication skills in real medical environments.
 This study constructs high-quality datasets covering ex- (usually the closest length to the candidate transla-
tensive medical knowledge and practical application scenar- tion). The penalty ensures that overly short transla-
ios through three main dataset generation methods: expert tions are penalized.
knowledge-guided professional knowledge assessment, cross- – pn is the precision for n-grams, i.e., the proportion
of n-grams in the generated text that also appear in
modal diagnosis in complex medical environments, and med- the reference translations.
ical risk assessment. These datasets not only support medi- Number of matching n-grams
pn =
cal knowledge learning and pathological reasoning but also Total number of n-grams in the generated translation
enhance training for multi-turn dialogues and ethical risk as- – wn is the weight for each n-gram, typically set to
sessments, significantly improving the application capabilities wn = N1 , where N is the highest n-gram order being
of large models in specialized medical fields. Additionally, considered (e.g., if using 1-gram to 4-gram precision,
these methods enhance the system’s adaptability and accuracy N = 4).
in complex medical environments. Ultimately, the generated
Variants of BLEU:
datasets provide a solid data foundation for training and – BLEU@1 (1-gram): Measures precision for single
optimizing medical AI systems, promoting the development of words (1-grams). This metric is useful for evaluat-
intelligent medical technologies, and enhancing the precision ing the accuracy of word choices in the generated
translation. The precision is calculated as:
of medical services and the effectiveness of patient education.
Number of matching 1-grams
Role-based and Process-oriented Prompt: In preparing the p1 =
Number of 1-grams in the generated translation
dataset for large model invocation, we used a Role-based
and Process-oriented Prompt design. This approach improves – BLEU@5 (5-gram): Measures precision for 5-
grams, which are sequences of 5 consecutive words.
reasoning by guiding models through intermediate steps [30], BLEU@5 places more emphasis on evaluating the
enhances adaptability and generalizability [31], and enables fluency and grammatical structure of the generated
realistic role simulation by aligning responses with specific translation. It ensures that not only individual words
but also longer word sequences match the reference
knowledge bases [32]. This method balances reasoning, flex- translations.
ibility, and role fidelity, supporting model optimization. Number of matching 5-grams
p5 =
Number of 5-grams in the generated translation
E XPERIMENTS
• METEOR (Metric for Evaluation of Translation with
Evaluation Metrics: Evaluation metrics provide objective Explicit ORdering): METEOR was developed to address
means to assess the quality of generated text by comparing some of the limitations of BLEU, such as its inability to
it to reference texts. They evaluate aspects such as fluency, handle synonyms and word order variations. METEOR
semantic accuracy, and content coverage. Below is a concise evaluates translation quality by considering word overlap,
description of key metrics like BLEU, METEOR, NIST, and synonym matching, stemming (reduction to base forms
ROUGE, highlighting their unique strengths in text evaluation. of words), and word order. METEOR has been shown
• BLEU (Bilingual Evaluation Understudy): BLEU is to correlate better with human judgment than BLEU,
an automated metric for evaluating machine translation especially when it comes to evaluating readability and
output by comparing the generated translation with a semantic correctness.
set of reference translations. BLEU focuses on precision
(how many n-grams in the generated text match n-grams METEOR = F · (1 − α · penalty)
in the reference texts) and applies a brevity penalty to where:
discourage excessively short translations. The idea is that
– F is the harmonic mean of precision and recall,
a good translation should contain most of the n-grams
similar to the F1 score in classification tasks, which
from the reference translations without overly repeating
combines how many words from the generated trans-
words.
lation are correct (precision) and how many relevant
Formula:
! words are retrieved (recall).
N
BLEU = BP · exp
X
wn · log pn 10 · precision · recall
F=
n=1 9 · precision + recall
where: – α is a weight factor that controls the influence of the
penalty on the final METEOR score. It is typically
– BP is the Brevity Penalty, which adjusts for transla-
set to 0.5.
tions that are too short. It is computed as:
– penalty is a factor that penalizes the translation for
(
1 if c > r word order differences. The more the order of words
BP = r

in the generated text differs from the reference text,
exp 1 − c if c ≤ r
the higher the penalty.
where c is the length of the candidate (generated) • NIST (National Institute of Standards and Technol-
translation, and r is the effective reference length ogy): NIST is a variant of BLEU that is designed to
 TABLE I
R ESULTS OF ECG P ROFESSIONAL K NOWLEDG

Model BLEU@1 BLEU@5 METEOR NIST ROUGE-1 ROUGE-2

ChatGPT-4 0.39929 0.16525 0.43025 1.91089 0.54256 0.34027
ChatGPT-3.5-turbo 0.35547 0.13232 0.39739 1.72209 0.48996 0.28754
Qwen-plus 0.35193 0.11995 0.40641 1.71437 0.48758 0.27560
Claude 3 0.32296 0.10980 0.42207 1.57859 0.47026 0.26265
Llama 3.1 8B 0.31164 0.10625 0.38012 1.58009 0.45439 0.24375
GLM-4-Air 0.36863 0.13040 0.41310 1.77895 0.50769 0.29582
Gemini-Pro 0.35170 0.12934 0.36990 1.66379 0.48297 0.28759
NVIDIA Nemotron 0.27132 0.08588 0.41381 1.32598 0.42868 0.23479
360GPT 0.22092 0.08598 0.25399 1.02528 0.30523 0.18459

handle longer texts better and to provide more infor- semantic coherence and content coverage. The combined use
mative evaluation by incorporating the ”informativeness” of these metrics helps to identify the strengths and weaknesses
of the n-grams. It reduces the weight of common n- of models, providing valuable information for performance
grams, allowing the evaluation to focus more on rare and optimization.
informative n-grams, which is important for domains like Model Selection: Given the computational and financial
technical documents or academic articles. limitations of this study, we focused our practical tests on a
Formula: carefully chosen subset of the dataset, specifically targeting
N   professional knowledge related to ECG analysis. This ap-
X log cn
NIST = · pn proach allowed us to evaluate the performance of various mod-
n=1
log N els under realistic constraints while ensuring relevance to criti-
where: cal medical applications. We selected a diverse range of state-
of-the-art market models for evaluation, encompassing both
– cn is the number of unique n-grams in the reference
widely recognized and emerging systems, including ChatGPT-
translation.
4, GLM-4-Air, ChatGPT-3.5-turbo, Qwen-plus, Gemini-Pro,
– pn is the precision for n-grams.
Claude 3, Llama 3.1 8B, NVIDIA Nemotron, 360GPT, and
– The logarithmic scaling of cn adjusts the weight
DeepSeek-Chat. This selection reflects a balance of established
based on the frequency of the n-grams, allowing
benchmarks and innovative solutions, providing a comprehen-
more informative n-grams to have a greater impact
sive perspective on their capabilities and limitations within the
on the score.
domain of ECG analysis.
• ROUGE (Recall-Oriented Understudy for Gisting
Experiments: We implemented a consistent prompt design
Evaluation): ROUGE is a set of metrics used pri- across all large models, ensuring uniformity in the input
marily for evaluating automatic summarization systems, structure. This approach allows for a fair and systematic
though it is also used for machine translation and other evaluation by minimizing the variability introduced by prompt
NLP tasks. Unlike BLEU, which emphasizes precision, differences. The outputs generated by the models were then
ROUGE focuses on recall, evaluating how much of the assessed against our established ground truth data, allowing
important content in the reference summary is captured a precise and objective comparison. This method not only
by the generated summary. standardizes the evaluation process, but also ensures that
– ROUGE-1: Measures the recall of individual words performance differences among models can be attributed to
(1-grams). This metric evaluates how much of the their inherent capabilities rather than inconsistencies in input
vocabulary from the reference summary is captured design. By maintaining a consistent input format, we ensure re-
in the generated summary. liable and reproducible results, providing meaningful insights
Pn
Recall of 1-grams into the strengths and weaknesses of each model.
ROUGE-1 = i=1
n
R ESULT
– ROUGE-2: Measures the recall of 2-grams, which
are pairs of consecutive words. ROUGE-2 is more In this study, we constructed the ECG-Expert-QA dataset
sensitive to the coverage of content and is useful and evaluated it using various modern medical language mod-
for assessing how well the summary covers the els. The experimental results indicate significant differences in
important content of the reference. the performance of different models on ECG diagnostic tasks.
Pn In terms of evaluation metrics such as BLEU@1, METEOR,
Recall of 2-grams NIST, and ROUGE-1, ChatGPT-4 performed exceptionally
ROUGE-2 = i=1
n well, particularly in generating medically relevant content,
Through these metrics, we can comprehensively evaluate with noticeably higher semantic accuracy and content coverage
the quality of generated text, ranging from lexical accuracy to compared to other models. Additionally, after multiple rounds
of dialogue and clinical reasoning tests, ChatGPT-4 demon- [12] Thapa S, Howlader K, Bhattacharjee S, le W. MoRE: Multi-Modal
strated strong clinical reasoning capabilities, effectively inte- Contrastive Pre-training with Transformers on X-Rays, ECGs, and Di-
agnostic Report; 2024. Available from: https://arxiv.org/abs/2410.16239.
grating multimodal information for diagnosis. However, some [13] Vaid A, Jiang J, Sawant A, Lerakis S, Argulian E, Ahuja Y, et al..
models (e.g., 360GPT and NVIDIA Nemotron) performed HeartBEiT: Vision Transformer for Electrocardiogram Data Improves
poorly in terms of accuracy and content coverage, especially Diagnostic Performance at Low Sample Sizes; 2022. Available from:
https://arxiv.org/abs/2212.14040.
when dealing with complex ECG patterns. Overall, although [14] Zhao Z. Transforming ECG Diagnosis:An In-depth Review of
current medical AI systems have made significant progress, Transformer-based DeepLearning Models in Cardiovascular Disease
challenges still remain in diagnosing complex pathologies and Detection; 2023. Available from: https://arxiv.org/abs/2306.01249.
[15] Jin D, Pan E, Oufattole N, Weng WH, Fang H, Szolovits P. What
rare cases. Disease does this Patient Have? A Large-scale Open Domain Ques-
tion Answering Dataset from Medical Exams; 2020. Available from:
C ONCLUSION https://arxiv.org/abs/2009.13081.
The ECG-Expert-QA dataset presented in this study not [16] Oh J, Lee G, Bae S, myoung Kwon J, Choi E. ECG-QA: A Comprehen-
sive Question Answering Dataset Combined With Electrocardiogram;
only provides a diversified and challenging evaluation platform 2023. Available from: https://arxiv.org/abs/2306.15681.
but also holds significant importance in testing the diagnostic [17] Wagner P, Strodthoff N, Bousseljot R, Samek W, Schaeffter T. PTB-XL,
capabilities, clinical reasoning, and cross-modal information a large publicly available electrocardiography dataset (version 1.0.1).
PhysioNet; 2020. Available from: https://doi.org/10.13026/x4td-x982.
integration of medical language models. When compared with [18] Wagner P, Strodthoff N, Bousseljot RD, Kreiseler D, Lunze FI, Samek
existing medical datasets, this dataset demonstrates unique W, et al. PTB-XL: A Large Publicly Available ECG Dataset. Scientific
value through multidimensional testing in pathology analysis, Data. 2020. Available from: https://doi.org/10.1038/s41597-020-0495-6.
[19] Zhao Y, Zhang T, Wang X, Han P, Chen T, Huang L, et al.. ECG-Chat:
case complexity, and ethical decision-making. Nevertheless, as A Large ECG-Language Model for Cardiac Disease Diagnosis; 2024.
AI technology advances, future research should focus on en- Available from: https://arxiv.org/abs/2408.08849.
hancing dynamic evaluation capabilities in clinical workflows, [20] Fu G, Zheng J, Abudayyeh I, Ani C, Rakovski C, Ehwerhemuepha L,
et al. CardioGPT: An ECG Interpretation Generation Model. IEEE
optimizing multilingual support, and verifying its applicability Access. 2024;12:50254-64.
in real-world clinical environments. In conclusion, the ECG- [21] Zheng J, Guo H, Chu H. A large scale 12-lead electrocardiogram
Expert-QA dataset provides a crucial reference for building database for arrhythmia study; 2022. PhysioNet. Available from:
https://doi.org/10.13026/wgex-er52.
medical AI evaluation systems and contributes to the ongoing [22] Zheng J, Chu H, Struppa D, Zhang J, Yacoub SM, El-Askary H, et al.
development of intelligent medical diagnostic technologies. Optimal Multi-Stage Arrhythmia Classification Approach. Scientific
Reports. 2020 2;10(1):2898. Available from: https://doi.org/10.1038/
R EFERENCES s41598-020-59821-7.
[1] Moody GB, Mark RG. The impact of the MIT-BIH Arrhythmia [23] Cai Y, Wang L, Wang Y, de Melo G, Zhang Y, Wang YF, et al. MedE-
Database. IEEE Engineering in Medicine and Biology Magazine. valHub: A Large-Scale Chinese Benchmark for Evaluating Medical
2001;20(3):45-50. Large Language Models. In: Proceedings of the Thirty-Eighth AAAI
[2] Johnson A, Bulgarelli L, Pollard T, Gow B, Moody B, Horng S, et al.. Conference on Artificial Intelligence (AAAI-2024). AAAI; 2024. .
MIMIC-IV (version 3.1). PhysioNet; 2024. Available from: https://doi. [24] Liao Y, Meng Y, Liu H, Wang Y, Wang Y. An Automatic Evaluation
org/10.13026/kpb9-mt58. Framework for Multi-turn Medical Consultations Capabilities of Large
[3] Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng Language Models. arXiv preprint. 2023.
S, et al. MIMIC-IV, a freely accessible electronic health record dataset. [25] Yang Y, Liao Y, Wang Y, Wang L, He L, Zhang Y, et al. GenMedi-
Scientific Data. 2023 jan;10(1):1. Available from: https://doi.org/10. calEval: A Unified Medical Evaluation Benchmark for Chinese LLMs.
1038/s41597-022-01899-x. arXiv preprint. 2023.
[4] Wagner P, Strodthoff N, Bousseljot R, Samek W, Schaeffter T. PTB-XL, [26] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PC, Mark
a large publicly available electrocardiography dataset (version 1.0.1). RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: Components of
PhysioNet; 2020. Available from: https://doi.org/10.13026/x4td-x982. a new research resource for complex physiologic signals. Circulation.
[5] Wagner P, Strodthoff N, Bousseljot RD, Kreiseler D, Lunze FI, Samek 2000;101(23):e215-20.
W, et al. PTB-XL, a large publicly available electrocardiography dataset. [27] Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O,
Scientific Data. 2020 may;7(1):154. Available from: https://doi.org/10. et al. Scikit-learn: Machine Learning in Python. Journal of Machine
1038/s41597-020-0495-6. Learning Research. 2011;12:2825-30. Available from: https://jmlr.csail.
[6] Chandra BS, Sastry CS, Jana S. Robust Heartbeat Detection from mit.edu/papers/v12/pedregosa11a.html.
Multimodal Data via CNN-based Generalizable Information Fusion; [28] Zhang J, Marszalek M, Lazebnik S, Schmid C. Local Features
2018. Available from: https://arxiv.org/abs/1807.03232. and Kernels for Classification of Texture and Object Categories: A
[7] Jun TJ, Nguyen HM, Kang D, Kim D, Kim D, Kim YH. ECG arrhythmia Comprehensive Study. International Journal of Computer Vision.
classification using a 2-D convolutional neural network; 2018. Available 2007;73(2):213-38. Available from: https://hal.archives-ouvertes.fr/
from: https://arxiv.org/abs/1804.06812. hal-00171412/document.
[8] Reasat T, Shahnaz C. Detection of inferior myocardial infarction [29] Bianchi FM, Livi L, Ferrante A, Milosevic J, Malek M. Time series
using shallow convolutional neural networks. In: 2017 IEEE Region kernel similarities for predicting Paroxysmal Atrial Fibrillation from
10 Humanitarian Technology Conference (R10-HTC). IEEE; 2017. p. ECGs; 2018. Available from: https://arxiv.org/abs/1801.06845.
718–721. Available from: http://dx.doi.org/10.1109/R10-HTC.2017. [30] Wei J, Wang X, Schuurmans D, Bosma M, Ichter B, Xia F, et al.. Chain-
8289058. of-Thought Prompting Elicits Reasoning in Large Language Models;
[9] Chen J, Valehi A, Razi A. Predictive Modeling of Biomedical Signals 2023. Available from: https://arxiv.org/abs/2201.11903.
Using Controlled Spatial Transformation; 2018. Available from: https: [31] Zhang L, Ergen T, Logeswaran L, Lee M, Jurgens D. SPRIG: Improving
//arxiv.org/abs/1811.00079. Large Language Model Performance by System Prompt Optimization;
[10] Cheng R, Zhuang Z, Zhuang S, Xie L, Guo J. MSW-Transformer: 2024. Available from: https://arxiv.org/abs/2410.14826.
Multi-Scale Shifted Windows Transformer Networks for 12-Lead ECG [32] Zhang B, Huang Y, Cui W, Zhang H. Thinking Before Speaking: A
Classification; 2023. Available from: https://arxiv.org/abs/2306.12098. Role-playing Model with Mindset; 2024. Available from: https://arxiv.
[11] Samanta A, Karlov M, Ravikumar M, Clarke CM, Rajadas J, Hassani org/abs/2409.13752.
K. MVMTnet: A Multi-variate Multi-modal Transformer for Multi-
class Classification of Cardiac Irregularities Using ECG Waveforms and
Clinical Notes; 2023. Available from: https://arxiv.org/abs/2302.11021.
 What is the significance of the PR interval in an ECG?
Am I allowed to ask my doctor to perform an ECG at a specific
The PR interval represents the time it takes for the electrical time that suits me?
impulse to travel from the atria to the ventricles. It indicates the
Yes, you can request a specific time for your ECG, and the
health of the AV node and conduction system."
medical facility will do its best to accommodate your schedule,
depending on availability."

How is the heart rate calculated from an ECG?

Am I allowed to ask my doctor to not include certain information
The heart rate can be calculated by measuring the distance in my medical records?
between two consecutive R waves (RR interval) on the ECG.
The rate in beats per minute (bpm) is calculated as 1500 divided No, you cannot request the omission of certain information from
by the number of small squares between the R waves. your medical records if it is medically relevant. Medical records
must be complete and accurate."

What is the significance of the ST segment in an ECG? Can I request that my ECG be scheduled at a different facility?

The ST segment represents the period between ventricular Yes, you can request that your ECG be scheduled at a different
depolarization and repolarization. Its elevation or depression is facility, depending on availability and medical necessity.
crucial in diagnosing myocardial ischemia or infarction.

Fig. 4. Data display of Patient’s right to know

Fig. 3. Data display of ECG Basic Knowledge

Typhoid Fever-Induced Myocarditis

What are the typical ECG changes observed in patients with
myocarditis secondary to typhoid fever?
Will I have to stay on these medications for life? Can atrial
Typical ECG changes include low voltage QRS complexes, ST- fibrillation be cured?
segment depression, and T-wave inversions, indicative of
myocardial inflammation and injury." In most cases, medications are used long-term to manage atrial
fibrillation and prevent complications like stroke. While it’s not
always curable, procedures such as catheter ablation can
sometimes restore normal rhythm.
Diphtheritic Myocarditis
What ECG findings might be indicative of significant myocardial
involvement in diphtheritic myocarditis?

Significant myocardial involvement may be indicated by marked How can I avoid angina attacks? Are there activities I should stop
low voltage QRS complexes, diffuse ST-segment changes, and doing?
T-wave inversions. Avoiding physical or emotional stress can help prevent angina
attacks. Regular exercise, a healthy diet, and taking your
medications as prescribed will also reduce the risk. We’ll
Short QT Syndrome (SQTS) monitor your condition closely to adjust your treatment.

How does the presence of a family history of sudden cardiac

death impact the diagnosis and treatment of Short QT Syndrome?

A family history of sudden cardiac death is a significant risk

factor for Short QT Syndrome and suggests a higher likelihood What caused my heart failure, and will it get better with treatment?
of inheriting a severe form of the condition. This history
Heart failure can result from conditions like high blood pressure
necessitates a more aggressive approach to management,
or previous heart damage. Treatment can help manage symptoms
including early screening of family members, close monitoring,
and improve your quality of life, but it requires ongoing care,
and consideration of preventive measures such as ICD
including medications and lifestyle changes.
implantation.

Fig. 6. Data display of Patient prognosis

Fig. 5. Data display of Complex diagnosis
 Reports in MIMIC-IV-ECG
Is there any abnormality in the electrocardiogram? Can you
explain it in detail? What are the long-term outcomes of atrial septal defect (ASD)
closure in adults?
Your electrocardiogram (ECG) shows a sinus rhythm, which is
ASD closure in adults typically results in improved functional
a normal heart rhythm. However, there is a possible right atrial
capacity and a reduced risk of atrial arrhythmias, stroke, and
abnormality, which suggests that the right atrium of your heart
right heart failure. However, in older adults or those with
may not be functioning normally. The ECG is also borderline,
pulmonary hypertension or atrial fibrillation prior to closure,
meaning that some results are not clearly normal or abnormal.
long-term outcomes may be less favorable, and arrhythmias may
This could indicate a potential issue that needs further persist
evaluation, such as an atrial enlargement or other cardiac
condition.
How does the management of hypertrophic obstructive
cardiomyopathy (HOCM) differ from non-obstructive
Reports in MIMIC-IV-ECG hypertrophic cardiomyopathy?",

Could you explain what my ECG results suggest and what In HOCM, the goal is to relieve left ventricular outflow tract
actions I should consider next? obstruction, often with medications like beta-blockers, calcium
channel blockers, or disopyramide. In refractory cases, surgical
Your ECG shows a sinus rhythm with an rSr' pattern in lead V1, myectomy or alcohol septal ablation may be required. Non-
which is likely a normal variant. The QRS voltages are low in obstructive hypertrophic cardiomyopathy focuses more on
the precordial leads, and the overall assessment is borderline. symptom relief and preventing sudden cardiac death with the
This ECG does not indicate any specific disease but suggests use of medications and, in some cases, implantable
close monitoring due to the borderline nature. cardioverter-defibrillators (ICDs)."

What is the role of fractional flow reserve (FFR) in the

Reports in MIMIC-IV-ECG management of coronary artery disease?
What's the heart's activity travel itinerary that this ECG is FFR is an invasive measurement used to assess the
mapping? physiological significance of coronary artery stenoses. FFR-
guided decision-making helps to determine whether a coronary
The ECG shows a sinus rhythm with extensive T wave changes, lesion should be revascularized, with revascularization typically
which could indicate myocardial ischemia or myocardial recommended if FFR is below 0.80, indicating reduced blood
infarction. The rhythm is regular, but the abnormalities in the T flow to the myocardium."
waves suggest significant cardiac distress. This pattern is
consistent with myocardial infarction.
Fig. 8. Data display of Cardiology Knowledge
Fig. 7. Data display of Cross modal diagnosis
Patient 15296609 underwent five electrocardiograms (ECGs) between June 19, 2199, and March 25, 2200, showing various cardiac abnormalities. The first examination (June 19, 2199) revealed sinus rhythm, prolonged QT interval, left atrial abnormality, possible
anteroseptal infarct , and low QRS voltages. The second examination (June 26, 2199) showed atrial fibrillation with rapid ventricular response, possible septal infarct, and low QRS voltages. The third examination (June 30, 2199) also showed atrial fibrillation,
possible anteroseptal infarct, and low QRS voltages. The fourth examination (August 5, 2199) revealed atrial fibrillation, possible septal infarct, low QRS voltages, and lateral ST-T changes. The final examination (March 25, 2200) showed sinus bradycardia,
possible inferior infarct, low QRS voltages, and possible left ventricular hypertrophy (LVH). All examinations showed abnormalities, and further evaluation and monitoring are recommended.

Comprehensive Diagnostic QA: "Q": "What are the potential long-term complications and prognosis for Patient 15296609 "Q": "What is the recommended long-term treatment plan for Patient 15296609 based on the
"Q": "What are the key cardiac abnormalities observed in Patient 15296609 over the five ECGs based on the observed cardiac abnormalities?", observed cardiac abnormalities?",
conducted between June 19, 2199, and March 25, 2200?", "A": "The potential long-term complications and prognosis for Patient 15296609 include: "A": "The recommended long-term treatment plan for Patient 15296609 includes:
"A": "The key cardiac abnormalities observed in Patient 15296609 include: Increased Risk of Arrhythmias: The prolonged QT interval and atrial fibrillation observed Rate Control in Atrial Fibrillation: Given the presence of atrial fibrillation with a rapid
Sinus Rhythm and Prolonged QT Interval: Observed in the first ECG on June 19, 2199. indicating in multiple ECGs increase the risk of potentially life-threatening arrhythmias such as ventricular response in multiple ECGs, medications such as beta-blockers (e.g., metoprolol) or
a normal sinus rhythm but with an abnormally prolonged QT interval, which can be associated ventricular tachycardia or fibrillation. Heart Failure: The presence of low QRS voltages, calcium channel blockers (e.g., diltiazem) should be considered to control the heart rate and
with an increased risk of arrhythmias. Left Atrial Abnormality: Also noted in the first ECG. infarcts, and possible LVH suggests a higher risk of developing heart failure, which can prevent tachycardia-induced cardiomyopathy. Anticoagulation: Due to the presence of atrial
suggesting possible left atrial enlargement or dysfunction. Anteroseptal Infarct and Inferior T lead to symptoms such as shortness of breath, fatigue, and fluid retention. Stroke: Atrial fibrillation, anticoagulation therapy with warfarin or a direct oral anticoagulant (DOAC) like
Wave Changes: The first ECG also showed an anteroseptal infarct (age undetermined) and fibrillation significantly increases the risk of stroke due to the formation of blood clots in apixaban or rivaroxaban is recommended to reduce the risk of stroke. Antiarrhythmic Therapy:
nonspecific inferior T wave changes. which could indicate previous myocardial injury or ischemia. the atria, which can embolize to the brain. Recurrent Myocardial Infarction: Given the If atrial fibrillation persists and is symptomatic, antiarrhythmic drugs such as amiodarone or
Generalized Low QRS Voltages: Present in all ECGs. suggesting potential myocardial damage or history of infarcts and ischemic changes, there is a risk of recurrent myocardial infarction, sotalol may be considered to maintain sinus rhythm. Management of Ischemic Heart Disease:
infiltrative cardiomyopathy. Atrial Fibrillation with Rapid Ventricular Response: Observed in the which can further damage the heart muscle and worsen heart function. Cardiomyopathy: Given the presence of infarcts and ST-T changes, lifestyle modifications (e.g., smoking
second, third, and fourth ECGs. indicating a rapid and irregular heart rate due to atrial fibrillation, The rapid ventricular response in atrial fibrillation can lead to tachycardia-induced cessation, diet, exercise) and medications such as aspirin, statins, and ACE inhibitors or ARBs
which can lead to poor cardiac output and increased risk of stroke Possible Septal Infarct: Noted cardiomyopathy, which may improve with rate control but could persist if not managed should be considered to manage ischemic heart disease and reduce the risk of further cardiac
in the second, third, and fourth ECGs. Lateral ST-T Changes: Seen in the fourth ECG. which may effectively. Prognosis: The overall prognosis is guarded due to the multiple cardiac events. Monitoring and Follow-Up: Regular follow-up with ECGs, echocardiograms, and
indicate myocardial ischemia or injury in the lateral wall of the heart. Sinus Bradycardia and abnormalities and the risk of serious complications. However, with appropriate stress tests to monitor for changes in heart function, arrhythmias, and ischemia. Additionally,
Possible Inferior Infarct: Observed in the final ECG on March 25, 2200. indicating a slow heart management and regular follow-up, the patient may experience improved quality of life regular blood tests to monitor anticoagulation levels and renal function. Symptomatic
rate and possible infarct in the inferior wall of the heart. QRS Changes in V3/V4: Noted in the and reduced risk of adverse events. Close monitoring and timely intervention are crucial Management: If the patient experiences symptoms such as chest pain, shortness of breath, or
final ECG. which may be due to left ventricular hypertrophy (LVH) or other structural changes." to managing the patient's condition and preventing further deterioration. " syncope, immediate medical evaluation and management should be pursued."

"Q": "What are the main cardiac abnormalities observed in my ECGs?", "initial conversation between doctor and patient": "Doctor: Good morning, Mr. Patient Overview: Patient 15296609 has undergone five electrocardiograms (ECGs) between
"A": "The main cardiac abnormalities observed in your ECGs include sinus rhythm, prolonged QT 15296609. I have reviewed your ECG records from the past year. Can you tell me if you June 19, 2199, and March 25, 2200, each revealing various cardiac abnormalities. The
interval, left atrial abnormality, anteroseptal and inferior infarcts (age undetermined), inferior T have experienced any chest pain, shortness of breath, or palpitations recently? Patient: patient's ECG findings include sinus rhythm, atrial fibrillation, prolonged QT interval, left
wave changes, generalized low QRS voltages, atrial fibrillation with rapid ventricular response, Yes, I have had some chest pain, especially when I exert myself. It feels like a pressure on atrial abnormality, infarcts (anteroseptal, septal, inferior), inferior T wave changes,
lateral ST-T changes, and sinus bradycardia. These findings suggest a complex cardiac condition my chest, and it usually goes away after I rest. Doctor: How long have you been generalized low QRS voltages, and lateral ST-T changes.
that may involve multiple areas of the heart." experiencing this chest pain? Patient: It started about six months ago, but it has been Risk Factor Analysis:
"Q": "What does a prolonged QT interval indicate?", getting more frequent lately. Doctor: Have you noticed any changes in your heart 1. Atrial Fibrillation (AF): - Present in multiple ECGs, indicating a significant arrhythmia. -
"A": "A prolonged QT interval on an ECG can indicate an increased risk of abnormal heart rhythm? For example, do you feel your heart skipping beats or racing? Patient: Yes, AF increases the risk of thromboembolism, particularly stroke. - Rapid ventricular response
rhythms, particularly torsades de pointes, which is a type of ventricular tachycardia. This condition sometimes my heart feels like it's racing, and I can feel it pounding in my chest. It's quite in AF can lead to hemodynamic instability.
can be life-threatening and may require medical intervention to correct." unsettling...... 2. Infarcts: - Possible anteroseptal, septal, and inferior infarcts, with age undetermined. -
"Q": "What are the implications of atrial fibrillation with a rapid ventricular response?", "patient memory correction": "Doctor: Mr. 15296609, I noticed that your first ECG Infarcts suggest previous myocardial damage, which can predispose to further ischemic events.
"A": "Atrial fibrillation with a rapid ventricular response means that the heart's upper chambers showed a prolonged QT interval and left atrial abnormality. Can you recall if you had any ......
(atria) are beating irregularly and rapidly, which can lead to a fast and irregular heartbeat in the specific symptoms around that time? Patient: Oh, I remember now. I had some Suggested Clinical Actions:
lower chambers (ventricles). This condition can increase the risk of stroke, heart failure, and other palpitations around that time, and I felt a bit dizzy. I didn't think much of it at the time. 1. Anticoagulation Therapy: - Given the atrial fibrillation, anticoagulation therapy (e.g.,
complications. It may require medication or other interventions to control the heart rate and Doctor: Thank you for the clarification. That information is very helpful.", warfarin, direct oral anticoagulants) is recommended to reduce the risk of stroke.
rhythm." "doctor adjusts diagnosis based on new information": "Doctor: Based on your recent 2. Cardiac Monitoring: - Continuous cardiac monitoring to detect and manage any
"Q": "What are the potential causes of the infarcts observed in my ECGs?", symptoms and the ECG findings, it appears that you may have a history of myocardial arrhythmias, especially given the prolonged QT interval.
"A": "The infarcts observed in your ECGs, both anteroseptal and inferior, suggest areas of the heart infarction, particularly in the anteroseptal and inferior regions. The ECGs also show atrial ......
muscle that have been damaged due to reduced blood flow. This could be due to blockages in the fibrillation and low QRS voltages, which could indicate underlying heart disease. Patient: Conclusion:Patient 15296609 presents with multiple significant cardiac abnormalities,
coronary arteries, which supply blood to the heart muscle. The exact cause would need to be Is that serious? What does that mean for me? Doctor: It does suggest that you have some including atrial fibrillation, infarcts, and a prolonged QT interval, placing them at high risk for
determined through further diagnostic tests, such as coronary angiography." significant cardiac issues. We will need to conduct further tests to confirm the diagnosis thromboembolism, arrhythmias, and cardiac decompensation. Immediate clinical intervention,
"Q": "What are the possible treatments for my condition?", and determine the best course of treatment. Doctor: Have you ever been diagnosed with including anticoagulation therapy, cardiac monitoring......
"A": "The treatment for your condition..... high blood pressure or diabetes? Patient: Yes, I have......

Fig. 9. Data display of patient 15296609

Patient 15772179 underwent eight ECGs between April 2, 2151, and February 11, 2153, all showing cardiac abnormalities. The first test revealed sinus rhythm with bigeminal PVCs, left axis deviation, an undetermined-age inferior infarct, and lateral ST-T changes.
Subsequent tests showed atrial fibrillation, sinus tachycardia, IV conduction defects, demand pacing, atrial pacing, borderline first-degree A-V block, and extensive ST-T changes. The final ECG on February 11, 2153, indicated a possible ectopic atrial rhythm with
PVCs, poor R wave progression, and an abnormal ECG.

"Q": "What are the key cardiac abnormalities observed in Patient 15772179 across the eight ECGs "Q": "Based on the comprehensive diagnosis, what should be the primary components of "Q": "What are the potential long-term complications and prognosis for Patient 15772179
conducted between April 2, 2151, and February 11, 2153?", the long-term treatment plan for Patient 15772179?", based on the observed cardiac abnormalities?",
"A": "The key cardiac abnormalities observed in Patient 15772179 include: Sinus rhythm with "A": "The long-term treatment plan for Patient 15772179 should include: Antiarrhythmic "A": "The potential long-term complications and prognosis for Patient 15772179 include:
bigeminal PVCs and left axis deviation in the first ECG. Atrial fibrillation with rapid ventricular therapy to manage atrial fibrillation and premature ventricular contractions (PVCs). Beta- Increased risk of arrhythmias, including atrial fibrillation and PVCs, which can lead to
response, leftward axis, poor R wave progression, inferior lateral ST-T changes and possible left blockers or calcium channel blockers to control heart rate and rhythm, especially in cases palpitations, syncope, and in severe cases, cardiac arrest. Progression of left ventricular
ventricular hypertrophy in the second ECG. Sinus tachycardia, left axis deviation, IV conduction of sinus tachycardia and rapid ventricular response. ACE inhibitors or ARBs to manage hypertrophy and heart failure, particularly if the inferior and lateral ST-T changes worsen or if
defect, QRS changes V3/V4, and lateral ST-T changes in the third ECG. Undetermined rhythm, possible left ventricular hypertrophy and improve cardiac function. Coronary artery coronary artery disease is not adequately managed. Conduction system abnormalities, such as
demand pacing, demand atrial pacing, and sinus rhythm with borderline 1st degree A-V block in disease management to address inferior and lateral ST-T changes, including lifestyle the IV conduction defect and borderline 1st degree A-V block, which may necessitate
subsequent ECGs. Possible ectopic atrial rhythm with PVC(s), IV conduction defect, poor R wave modifications, statins, and possibly revascularization if indicated. Pacemaker pacemaker implantation and increase the risk of bradycardia-related symptoms.
progression, extensive ST-T changes and abnormal ECG in the final ECG." implantation if the patient continues to experience conduction defects or bradycardia, as Cardiovascular events, including myocardial infarction and stroke, particularly if atrial
suggested by the borderline 1st degree A-V block and demand pacing. Regular follow-up fibrillation persists and leads to thromboembolic events. Overall prognosis is guarded, with a
ECGs and cardiac monitoring to track the progression of abnormalities and adjust need for close monitoring and aggressive management of risk factors to prevent complications
treatment as necessary." and improve quality of life."
"Q": "What does the term sinus rhythm mean in my ECG results?",
"A": "Sinus rhythm indicates that your heart is beating in a normal, regular pattern, which
originates from the sinoatrial (SA) node, the natural pacemaker of the heart......"
"Q": "What are bigeminal PVCs and how do they affect my heart?", "initial conversation between doctor and patient": "Doctor: Good morning, Mr. 15772179. Patient History and ECG Findings: Multiple ECG Examinations: The patient underwent eight
"A": "Bigeminal PVCs refer to premature ventricular contractions (PVCs) that occur in pairs. This I have reviewed your ECG records from the past few years. Can you tell me if you have ECGs over a period of approximately two years, each showing various cardiac abnormalities.
means that every other heartbeat is a PVC, which can cause irregularities in your heart experienced any chest pain, shortness of breath, or palpitations recently? Patient: Yes, I've Initial ECG (April 2, 2151, ID: 40737557): Sinus rhythm with bigeminal PVCs, left axis
rhythm......" been having some chest pain, especially when I exert myself. It feels like a tightness in my deviation, inferior infarct (age undetermined), and lateral ST-T ECG.Key Risk Factors
"Q": "What does left axis deviation mean, and is it serious?", chest. Doctor: How long have you been experiencing this chest pain?", Identified: Atrial Fibrillation: Present in multiple ECGs,
"A": "Left axis deviation means that the electrical axis of your heart is tilted more to the left than "patient memory correction": "Patient: Wait, I just remembered something. I actually had Potential Risk Assessment:High Risk of Cardiovascular Events: Atrial Fibrillation: Increases
normal. This can be caused by various factors, including left ventricular hypertrophy, left anterior a really bad episode of chest pain around that time. It was so severe that I had to stop what the risk of stroke and heart failure.- Left Ventricular Hypertrophy: Indicates increased cardiac
fascicular block, or other structural heart abnormalities......" I was doing and sit down. Doctor: Thank you for sharing that. That's important load and risk of heart failure.- Inferior Infarct: Suggests a history of myocardial injury,
"Q": "What are inferior infarct - age undetermined and lateral ST-T changes?", information. Can you describe the severity of the chest pain during that episode? Patient: increasing the risk of future infarctions.- Lateral ST-T Changes: Indicate ongoing ischemia or
"A": "Inferior infarct - age undetermined suggests that there may have been a previous heart attack It was very intense, like an elephant was sitting on my chest. It lasted for about 10 minutes myocardial injury
(myocardial infarction) affecting the inferior (bottom) part of your heart, but it is unclear when this before it started to ease up.Doctor: Did you take any medication or seek medical attention Suggested Clinical Actions: Immediate Actions: Cardiac Ultrasound (Echocardiogram): To
occurred......" during that episode?Patient: No, I didn't take anything. I just waited for it to pass." evaluate left ventricular size, function, and wall thickness.- Transesophageal
Echocardiography (TEE): To rule out atrial thrombus in the setting of atrial fibrillation.-
Holter Monitoring: To assess the frequency and pattern of PVCs and atrial fibrillation......

Fig. 10. Data display of patient 15772179

Patient 12294203 underwent eight ECGs between October 16, 2164, and July 9, 2166, all showing cardiac abnormalities. The first test indicated sinus bradycardia, prolonged QT interval, left anterior fascicular block, IV conduction defect, and left ventricular hypertrophy. Subsequent tests revealed
sinus arrhythmia, ectopic atrial bradycardia, left axis deviation, QRS changes due to left ventricular hypertrophy, and various ST-T abnormalities. The final ECG on July 9, 2166, showed possible ectopic atrial rhythm, left axis deviation, QRS changes, left ventricular hypertrophy, and lateral ST-T
changes associated with hypertrophy.

"Q": "What are the key cardiac abnormalities observed in Patient 12294203 over the course of "Q": "What is the recommended long-term treatment plan for Patient 12294203 based on "Q": "What are the potential long-term complications and prognosis for Patient 12294203
their eight ECGs?", their cardiac abnormalities?", given their cardiac abnormalities?",
"A": "The key cardiac abnormalities observed in Patient 12294203 include: Sinus Bradycardia: "A": "The recommended long-term treatment plan for Patient 12294203 includes:Beta- "A": "The potential long-term complications and prognosis for Patient 12294203 include:
Present in multiple ECGs, indicating a slow heart rate. Prolonged QT Interval: Observed in the Blockers: To manage sinus bradycardia and reduce the risk of arrhythmias. ACE Increased Risk of Arrhythmias: Due to prolonged QT interval and ectopic atrial rhythms, there
first ECG. Suggesting a risk of arrhythmias. Left Axis Deviation: Present in several ECGs. Inhibitors or ARBs: To manage left ventricular hypertrophy and improve heart function. is a higher risk of developing more serious arrhythmias such as atrial fibrillation or ventricular
Indicating a possible left ventricular hypertrophy or conduction defect.Left Ventricular Antiarrhythmics: To control ectopic atrial rhythms and reduce the frequency of PACs and tachycardia. Progression of Left Ventricular Hypertrophy: If not managed effectively, LVH
Hypertrophy (LVH): Consistently noted, with QRS changes in V3/V4 and lateral ST-T changes PVCs. Diuretics: To manage any associated fluid overload due to left ventricular can lead to heart failure, reduced ejection fraction, and decreased quality of life. Conduction
due to hypertrophy. Conduction Defect: Seen in the first and second ECGs, suggesting a block in hypertrophy. Regular Monitoring: Frequent ECGs and echocardiograms to monitor the System Disorders: The IV conduction defect and left axis deviation could progress to more
the intraventricular conduction system.Ectopic Atrial Rhythm: Observed in multiple ECGs, progression of LVH and any changes in conduction defects. Lifestyle Modifications: severe conduction blocks, potentially requiring pacemaker implantation. Cardiovascular
indicating irregular atrial activity." Encourage a heart-healthy diet, regular exercise, and avoidance of stimulants like caffeine Events: Increased risk of cardiovascular events such as myocardial infarction and stroke,
and nicotine." particularly if atrial fibrillation persists and leads to thromboembolic events.Overall prognosis
is guarded, with a need for close monitoring and aggressive management of risk factors to
prevent complications and improve quality of life."
"Q": "What do the ECG results indicate about my heart condition?", "initial conversation between doctor and patient": "Doctor: Good morning, Mr.
"A": "The ECG results indicate a variety of cardiac abnormalities, including sinus bradycardia (a 12294203. I have your ECG results here. Can you tell me how you've been feeling lately?
slow heart rate), prolonged QT interval (which can be associated with arrhythmias), left axis Any chest pain, shortness of breath, or palpitations? Patient: Good morning, Doctor. I've Key findings include:
deviation (a shift in the electrical axis of the heart), and left ventricular hypertrophy (enlargement been feeling a bit tired lately, especially after physical activities. I do get some chest 1. Sinus Bradycardia: Present in multiple examinations, indicating a slow heart rate.
and thickening of the left ventricle). These findings suggest that your heart is working under discomfort, but it's not severe. I also feel my heart racing sometimes, but it goes away 2. Prolonged QT Interval: Seen in the first examination, which can increase the risk of
significant strain and may be at risk for more serious arrhythmias or heart failure." quickly. arrhythmias.
"Q": "What are the potential causes of these abnormalities?", "patient memory correction": "Doctor: Just to clarify, you mentioned that the chest 3. Left Anterior Fascicular Block (LAFB): Indicated in the first examination, suggesting a
"A": "The causes of these abnormalities can be multifactorial. Left ventricular hypertrophy, for discomfort started around the time of the second examination on October 28, 2164. Is conduction problem in the heart.
example, can be due to high blood pressure, aortic valve disease, or genetic factors. Sinus that correct? Patient: Oh, I think I might have mixed up the dates. The chest discomfort Potential Risk Assessment:
bradycardia and prolonged QT interval can be related to electrolyte imbalances, certain actually started around the time of the third examination on August 27, 2165. I remember Given the persistent and varied cardiac abnormalities, the patient is at high risk for:
medications, or underlying heart conditions. Left axis deviation can be due to structural heart because it was right after a particularly stressful project at work. Doctor: Thank you for 1. Progressive Heart Disease: The presence of LVH and conduction defects suggests ongoing
disease or conduction system abnormalities." correcting that. So, the chest discomfort started around August 27, 2165, and you noticed cardiac stress and potential for further deterioration.
"Q": "Do I need further diagnostic tests to confirm these findings?", it more during stressful periods. Is that right? Patient: Yes, that's correct.", 2. Arrhythmias: The prolonged QT interval and atrial arrhythmias increase the risk of serious
"A": "Yes, further diagnostic tests are necessary to confirm these findings and to understand the "doctor adjusts diagnosis based on new information": "Doctor: Based on the new arrhythmias, including ventricular tachycardia or fibrillation.
underlying causes. These may include echocardiograms to assess the structure and function of your information, it seems that your symptoms, particularly the chest discomfort and Suggested Clinical Actions:
heart, stress tests to evaluate how your heart performs under physical stress, and possibly a cardiac palpitations, started around the time of the third examination on August 27, 2165. This 1. Cardiac Ultrasound (Echocardiography): To evaluate the structure and function of the heart,
MRI or CT scan for more detailed imaging." aligns with the findings of possible ectopic atrial bradycardia with PAC(s), left axis particularly the left ventricle.
"Q": "What treatment options are available for my condition?", deviation, QRS changes V3/V4 due to LVH, left ventricular hypertrophy, and 2. Holter Monitoring: To capture any intermittent arrhythmias that may not be evident on
"A": "Treatment options will depend on the underlying causes and the severity of your condition. inferior/lateral STT changes due to hypertrophy. Given this, we should focus on routine ECG.
This may include medications to manage blood pressure, reduce the risk of arrhythmias, or managing these symptoms and monitoring your condition more closely. Patient: Okay, 3. Electrophysiological Study (EPS): If arrhythmias are suspected to be causing significant
improve heart function. In some cases, lifestyle changes such as dietary modifications, exercise, what do you recommend? Doctor: I recommend starting you on a betablocker to help symptoms or risk.
and stress management may be recommended. If the condition is severe, surgical interventions manage the palpitations and reduce the strain on your heart.
such as valve repair or replacement, or even heart transplantation, may be considered."

Fig. 11. Data display of patient 12294203

Patient 17684034 underwent eight ECGs between June 22, 2135, and October 23, 2137, all showing cardiac abnormalities. The first test suggested possible atrial flutter, left axis deviation, and repolarization changes. Subsequent tests revealed sinus rhythm with
PACs, 1st-degree A-V block, prolonged QT interval, ectopic atrial bradycardia, left anterior fascicular block, and left ventricular hypertrophy, along with various ST-T abnormalities. The final ECG on October 23, 2137, showed sinus bradycardia with 1st-degree A-V
block, left anterior fascicular block, left ventricular hypertrophy, and lateral ST-T changes.

"Q": "What are the primary cardiac abnormalities observed in Patient 17684034 over the course of "Q": "What treatment strategies should be considered for Patient 17684034 to manage their "Q": "What are the potential long-term complications and prognosis for Patient 17684034
their eight ECGs?", cardiac abnormalities?", given their cardiac abnormalities?",
"A": "The primary cardiac abnormalities observed in Patient 17684034 include: Atrial "A": "The treatment strategies for Patient 17684034 should include:Rate Control: "A": "The potential long-term complications and prognosis for Patient 17684034
Arrhythmias: Possible atrial flutter, premature atrial contractions (PACs), ectopic atrial Medications such as beta-blockers or calcium channel blockers to manage bradycardia and include:Progression of Arrhythmias: There is a risk of the atrial arrhythmias progressing to
bradycardia, and sinus bradycardia with sinus arrhythmia. Conduction Disturbances: First-degree control heart rate. Antiarrhythmic Therapy: Medications like amiodarone or sotalol to more severe forms, such as atrial fibrillation, which could increase the risk of stroke. Heart
atrioventricular (A-V) block and left anterior fascicular block. Electrical Axis Deviations: Left manage atrial arrhythmias and prevent atrial flutter or PACs. Conduction System Support: Failure: Left ventricular hypertrophy and conduction disturbances could lead to heart failure
axis deviation in multiple ECGs. Repolarization Changes: Lateral ST-T changes and prolonged Monitoring and possibly medications to manage A-V block and left anterior fascicular block. over time. Overall prognosis is guarded, with a need for close monitoring and aggressive
QT interval. Structural Changes: Left ventricular hypertrophy." Repolarization Management: Consideration of medications such as potassium supplements management of risk factors to prevent complications and improve quality of life."
or magnesium to address prolonged QT intervals.Regular Monitoring: Frequent ECGs and
echocardiograms to monitor the progression of LVH, conduction defects.....

"Q": "What are the main cardiac abnormalities observed in my ECGs over the past few years?", Risk Assessment Report for Patient 17684034 Risk Factor Analysis:The patient has
"A": "The main cardiac abnormalities observed in your ECGs include possible atrial flutter, sinus "initial conversation between doctor and patient": "Doctor: Good morning, Mr. 17684034. undergone eight electrocardiograms (ECGs) over a period of two years, each showing various
rhythm with premature atrial contractions (PACs), first-degree atrioventricular (A-V) block, left axis I have reviewed your ECG records from the past few years. Can you tell me if you have cardiac abnormalities.
deviation, left anterior fascicular block, lateral ST-T changes, left ventricular hypertrophy, and experienced any symptoms related to your heart, such as chest pain, palpitations, or Key findings include:
prolonged QT interval. These findings suggest a complex cardiac condition that requires careful shortness of breath? Patient: Yes, I have had some chest discomfort, especially after 1. Atrial Flutter/Ectopic Atrial Bradycardia: The first ECG showed possible atrial flutter,
monitoring and management." physical activity. I also feel my heart racing sometimes, but it usually goes away after a while subsequent ECGs showed irregular ectopic atrial bradycardia. These arrhythmias can
"Q": "What does left ventricular hypertrophy mean, and how does it affect my heart?", few minutes. Doctor: I see. Have you noticed any pattern or triggers for these symptoms? increase the risk of thromboembolism and stroke.
"A": "Left ventricular hypertrophy (LVH) means that the muscle wall of your left ventricle, the main For example, do they occur more frequently at certain times of the day or after specific 2. Left Axis Deviation: Persistent left axis deviation is noted in multiple ECGs, indicating
pumping chamber of your heart, has thickened. This can occur in response to increased workload on activities?Patient: It seems to happen more in the evening, especially after Ive been sitting possible left anterior fascicular block (LAFB). LAFB is associated with left ventricular
the heart, such as high blood pressure or aortic valve disease. LVH can lead to decreased heart for a long time. I also notice it more when Im stressed......", hypertrophy (LVH) and can lead to conduction abnormalities.
function over time and increase the risk of heart rhythm disturbances and heart failure." "patient memory correction": "Doctor: Just to clarify, you mentioned that your symptoms Potential Risk Assessment:Given the multiple cardiac abnormalities observed in the ECGs,
"Q": "What are the potential causes of these abnormalities?", seem to occur more in the evening and after sitting for a long time. Did you say that you the patient is at high risk for the following:
"A": "The potential causes of these abnormalities could include underlying heart disease, such as also experience these symptoms after physical activity? Patient: Oh, I think I misspoke 1. Progressive Heart Disease: The consistent findings of LVH and LAFB suggest ongoing
hypertension, coronary artery disease, or valvular heart disease. Additionally, genetic factors, earlier. Its actually more common after physical activity, not just sitting. I feel the chest cardiac remodeling and potential progression to heart failure.
lifestyle choices (such as diet and exercise), and other medical conditions (like diabetes) could discomfort and racing heart more after I exercise......", 2. Arrhythmias: The presence of atrial flutter, ectopic atrial bradycardia, and prolonged QT
contribute to these findings." "conclusion": "Doctor: Youre welcome, Mr. 17684034. Well schedule your next interval increases the risk of life-threatening arrhythmias.
"Q": "Do I need further tests to confirm the diagnosis and assess the severity of my condition?", appointment in a month to review your progress. In the meantime, please keep track of Suggested Clinical Actions:
"A": "Yes, further tests are necessary to confirm the diagnosis and assess the severity of your your symptoms and any changes in how you feel. If you experience any severe symptoms 1. Cardiac Imaging: Echocardiography: To assess the structure and function of the heart,
condition. These may include echocardiograms to visualize the heart structure and function, stress or have concerns, dont hesitate to contact us immediately. Take care, and well see you particularly the left ventricle. Transesophageal Echocardiography (TEE): To rule out atrial
tests to evaluate how your heart performs under physical activity, and possibly cardiac soon.Patient: Thank you, Doctor. Ill keep an eye on my symptoms and follow up as thrombus, especially given the atrial arrhythmias.
catheterization if there is suspicion of coronary artery disease." needed.Doctor: Great. Have a good day, and take care.Patient: You too. Goodbye." 2. Electrophysiological Studies: - Electrophysiological Study (EPS): To evaluate the
conduction system and risk of arrhythmias.

Fig. 12. Data display of patient 17684034

Fig. 13. Data display of Medical entity extraction

Fig. 14. Data display of Medical counterfactual