ElCardioCC: Clinical Benchmark & Cardiac Models
- ElCardioCC is a disambiguated label that primarily denotes a Greek clinical coding benchmark featuring 1,000 cardiology discharge letters annotated with ICD-10 codes.
- The benchmark supports modular tasks like named entity recognition, entity linking, and explainable multi-label classification via transformer-based models and domain-specific lexicons.
- Outside clinical NLP, the label also applies to distinct cardiac systems including an STM32-based ECG monitor, a 3D electrophysiological model, and a multimodal screening pipeline.
ElCardioCC is an overloaded label in the recent arXiv literature rather than a single, stable technical referent. In its clearest and most explicit usage, ELCardioCC denotes a Greek clinical coding benchmark for cardiology discharge letters, introduced as a BioASQ 2025 shared task and used for multilingual clinical entity linking to ICD-10 (Nentidis et al., 28 Aug 2025, Vassileva et al., 5 Sep 2025). In other supplied descriptions, the same or closely similar label is attached to a portable STM32-based ECG monitoring system, a five-dipole three-dimensional electrophysiological model, and a pilot multimodal ECG–fundus triage pipeline (Guan, 2024, Rueda et al., 2022, Muthukumar et al., 31 Mar 2025). The term therefore requires immediate disambiguation: in current clinical NLP usage it refers primarily to a Greek ICD-10 coding benchmark, whereas in cardiac engineering and modeling contexts it has been used for substantially different artifacts.
1. Nomenclature and referential scope
The orthography varies between ELCardioCC and ElCardioCC. Across the supplied sources, the label is not semantically uniform. That ambiguity is not incidental; it affects how the term should be interpreted in bibliographic search, benchmark comparison, and citation practice.
| Usage in supplied literature | Domain | Definition in the source |
|---|---|---|
| ELCardioCC | Clinical NLP benchmark | Greek discharge letters annotated with mention positions and ICD-10 codes (Nentidis et al., 28 Aug 2025) |
| ElCardioCC | Clinical entity linking dataset | 1,000 de-identified Greek cardiology discharge letters for ICD-10 linking (Vassileva et al., 5 Sep 2025) |
| ElCardioCC | Embedded ECG monitoring system | Portable STM32-based ECG monitor with cloud connectivity (Guan, 2024) |
| ElCardioCC / 3DFMM | Cardiac electrophysiology model | Five-dipole 3D model for ECG and VCG reconstruction (Rueda et al., 2022) |
| ElCardioCC | Multimodal screening pipeline | ECG + fundus FFT/EMD pipeline for four-class CVD triage (Muthukumar et al., 31 Mar 2025) |
This multiplicity is sharpened by a negative case. The cloud ECG analysis system "CardioLearn" explicitly does not mention the query term “ElCardioCC”; the system name used throughout that paper is CardioLearn (Hong et al., 2020). A practical implication is that the label should not be treated as a canonical synonym for cloud ECG analytics or for ECG deep learning more broadly.
2. ELCardioCC as a Greek clinical coding corpus
In the BioASQ-oriented interpretation, ELCardioCC is a specialized corpus of Greek discharge letters from hospitals, created for cardiology-focused clinical coding and related information extraction (Nentidis et al., 28 Aug 2025). The BioASQ overview specifies a training set of 1,000 discharge letters and a test set of 500 discharge letters, with annotation of the positions of mentions and their corresponding ICD-10 codes (Nentidis et al., 28 Aug 2025). The multilingual ICD-10 linking paper describes ELCardioCC as consisting of 1,000 de-identified hospital discharge letters, written in Greek, produced by cardiology doctors, and labeled by medical professionals (Vassileva et al., 5 Sep 2025).
The annotation scope is clinically oriented rather than generic. In the entity-linking paper, labeled spans are related to chief complaint, diagnosis, prior medical history, and findings, with each span assigned an ICD-10 code based on the term’s meaning in context (Vassileva et al., 5 Sep 2025). The BioASQ overview names chief complaint, diagnosis, prior medical history, drugs, and cardiac echo among the annotated mention categories (Nentidis et al., 28 Aug 2025). Taken together, these descriptions indicate a document collection centered on cardiology discharge summaries but designed to support multiple downstream formulations, from span extraction to code normalization and document-level coding.
The ICD-10 granularity reported for the Greek entity-linking experiments is the category level, i.e. the 3-character ICD-10 code (Vassileva et al., 5 Sep 2025). The same paper frames the broader problem as an extreme multi-class classification task because the full label space can involve about 2K ICD-10 categories or about 14K ICD-10 subcategories, although ELCardioCC itself is evaluated at the Greek category level only (Vassileva et al., 5 Sep 2025). A common misconception is that ELCardioCC is an end-to-end coding benchmark in which systems discover mentions and assign codes jointly. The entity-linking study explicitly states that the pipeline addresses entity linking only, while mention detection is left for future work (Vassileva et al., 5 Sep 2025).
3. BioASQ 2025 task structure and official benchmarking
Within BioASQ 2025, ELCardioCC was introduced as a new shared task on clinical coding in cardiology for Greek discharge letters (Nentidis et al., 28 Aug 2025). The overview situates it at the intersection of named entity recognition (NER), entity linking (EL), and multi-label classification with explainability (MLC-X). The three subtasks were defined as: NER, which identifies cardiology-related mention spans; EL, which links extracted mentions to ICD-10 codes; and MLC-X, which performs document-level ICD-10 multi-label prediction together with justification (Nentidis et al., 28 Aug 2025).
The official evaluation metric for ELCardioCC in the overview is micro-F1, with micro-averaged precision, recall, and F1 reported in the results tables (Nentidis et al., 28 Aug 2025). The baseline family is also explicit: a cased multilingual BERT (mBERT) model fine-tuned for BIO2 tagging for NER, a context-aware hierarchical classifier built on mBERT for EL, and a Greek-BERT multi-label model over the 40 most frequent ICD-10 codes for MLC-X, with rule-based justification variants (Nentidis et al., 28 Aug 2025).
| Subtask | Best system | Micro-F1 |
|---|---|---|
| NER | droidlyx system1 | 0.7328 |
| EL | droidlyx system1 | 0.6778 |
| MLC (Subtask 3a) | droidlyx system1 | 0.8472 |
| Explainability (Subtask 3b) | ELCardioCC_baseline MLCX2_baseline | 0.5122 |
The participating-system profile is equally informative. The task attracted five teams, and the overview highlights droidlyx, enigma, bhuang, pjmathematician, and ELCardioCC_baseline (Nentidis et al., 28 Aug 2025). The dominant methodological pattern was transformer-based modeling, including fine-tuning of Greek BERT and XLM-RoBERTa for NER, embedding-based semantic similarity for EL, and LLM-based classification and explanation for MLC-X (Nentidis et al., 28 Aug 2025). The top-line conclusion reported in the overview is that droidlyx was strongest overall, while the baseline remained notably competitive, especially in document-level coding (Nentidis et al., 28 Aug 2025).
This benchmark structure is significant because it makes ELCardioCC more than a single dataset release. It is a coordinated evaluation setting in which span detection, concept normalization, and explainable document-level coding are separated but still clinically connected. This suggests that the benchmark is intended to support modular system design rather than only monolithic end-to-end modeling.
4. Dictionary–LLM clinical entity linking on ElCardioCC
The paper "Using LLMs for Multilingual Clinical Entity Linking to ICD-10" uses ElCardioCC as its Greek benchmark for clinical entity linking (Vassileva et al., 5 Sep 2025). The task input is a mention and its surrounding discharge-summary context, and the output is the most appropriate ICD-10 code (Vassileva et al., 5 Sep 2025). The proposed system is a multistage hybrid pipeline: first, a dictionary exact match stage searches a language-specific ICD-10 dictionary; if the mention matches unambiguously to one code, that code is returned immediately; otherwise the mention is passed to GPT-4.1 for in-context prediction (Vassileva et al., 5 Sep 2025).
For Greek, the dictionary is built from Greek ICD-10 specifications and supplemented with ElCardioCC train-set terms; it contains about 11,500 terms mapped to 3-character ICD-10 codes (Vassileva et al., 5 Sep 2025). The LLM stage uses GPT-4.1, with prompts asking for a JSON array containing the medical term, ICD-10 code, and explanation, and instructing the model to choose the most specific code available, use context, state assumptions, provide multiple ICD-10 codes if needed, and return a code for all medical terms (Vassileva et al., 5 Sep 2025). The prompting setup uses one example discharge summary in the same language, and that example itself was generated using GPT-4o (Vassileva et al., 5 Sep 2025).
The evaluation on ElCardioCC reports Precision, Recall, and F1, under settings including 0-shot, 1-shot, and with or without dictionary combination, using temperature = 0.5, 6K max tokens, and a private Azure OpenAI deployment for privacy (Vassileva et al., 5 Sep 2025). The best Greek result is Dict + GPT-4.1 1-shot, with Precision = 0.856, Recall = 0.856, and F1 = 0.856 (Vassileva et al., 5 Sep 2025). The study also reports that GPT-4.1 is substantially stronger than GPT-4o, especially in recall, and explicitly concludes that adding the dictionary improves overall F1 (Vassileva et al., 5 Sep 2025).
Several limitations are stated. Prompt length / context complexity can reduce the model’s ability to assign codes for all mentions; GPT-4o especially struggled with recall; ICD-10 symptoms and “not otherwise classified” cases, especially in the R00–R99 chapter, are hard; performance may depend on the quality of the Greek ICD-10 dictionary; and the study addresses linking, not mention detection (Vassileva et al., 5 Sep 2025). These constraints matter because they delimit what the reported F1 actually measures: high-quality normalization given gold mentions, not full end-to-end coding from raw discharge text.
5. Alternative engineering and modeling uses of the label
Outside the Greek clinical NLP setting, the supplied literature associates ElCardioCC with technically unrelated cardiology artifacts. In one description, ElCardioCC is a portable, STM32-based ECG monitoring system intended for continuous and remote cardiac monitoring (Guan, 2024). That system combines an analog ECG front end, embedded processing on an STM32F429, 4G cloud connectivity through an L610-4G module, remote access through Tencent Cloud, and alerting when heart-rate thresholds are exceeded (Guan, 2024). Its analog chain includes instrumentation amplifier, voltage amplifier, high-pass filter, low-pass filter, 50 Hz notch filter, voltage lifting circuit, and right-leg drive circuit, with an overall gain of approximately 1500× and a bandwidth constrained to approximately 0.05 Hz to 70 Hz (Guan, 2024).
In another usage, ElCardioCC / 3DFMM is presented as a unique 3D cardiac electrical model (Rueda et al., 2022). The cardiac electric source is modeled as the sum of five dipole components, one for each major ECG wave, and , and each lead projection is represented by a frequency-modulated morphology (FMM) wave (Rueda et al., 2022). The model aims to reconstruct both standard 12-lead ECG and vectorcardiogram (VCG) signals with physiologically interpretable parameters such as amplitude, timing, asymmetry, and sharpness, and it is posed as a solution to both the forward and inverse problems of electrocardiographic modeling (Rueda et al., 2022).
A further use applies the label to a pilot multimodal cardiovascular screening system that integrates ECG and retinal fundus images for early detection and triaging of cardiovascular disease (Muthukumar et al., 31 Mar 2025). In that pipeline, ECG is preprocessed with bandpass filtering from 0.5 Hz to 50 Hz and R-peak detection, while fundus images are used without preprocessing; both modalities are transformed with FFT, Earth Mover’s Distance (EMD) is computed on the FFT-derived feature distributions, the resulting modality-specific EMD values are concatenated, and a CNN predicts one of four classes (Muthukumar et al., 31 Mar 2025). The pilot study uses 112 paired ECGs and fundus images and reports 84% overall accuracy (Muthukumar et al., 31 Mar 2025).
These uses are not minor variants of a common benchmark. They refer to a hardware system, a mathematical model, and a multimodal classifier with different inputs, outputs, and evaluation criteria. The terminological collision is therefore substantive rather than stylistic.
6. Interpretation, misconceptions, and research significance
The most important interpretive point is that ElCardioCC is not a universally fixed cardiology benchmark name across the supplied literature. In arXiv-facing clinical NLP discussion, the dominant explicit meaning is the Greek cardiology clinical coding benchmark associated with BioASQ 2025 and multilingual ICD-10 entity linking (Nentidis et al., 28 Aug 2025, Vassileva et al., 5 Sep 2025). In other supplied descriptions, the same label is attached to artifacts in embedded sensing, electrophysiological modeling, and multimodal screening (Guan, 2024, Rueda et al., 2022, Muthukumar et al., 31 Mar 2025). Any technical discussion that omits this disambiguation risks category errors, especially when comparing metrics across papers.
A second misconception is that ELCardioCC is simply an ICD-10 classification dataset. The BioASQ formulation is broader: it includes NER, EL, and MLC-X, and therefore spans mention extraction, terminology normalization, document-level coding, and justification (Nentidis et al., 28 Aug 2025). Conversely, the multilingual LLM study evaluates only entity linking, explicitly leaving mention detection unresolved (Vassileva et al., 5 Sep 2025). This distinction is methodologically important because a system can score strongly on linking with gold spans while remaining untested on full clinical-text coding.
A third misconception is that modern performance on the benchmark is purely a function of LLMs. The available results argue for a hybrid picture. In the BioASQ shared task, transformer-based fine-tuning and hierarchy-aware baselines are already strong, with droidlyx only modestly ahead of the strongest baseline in some subtasks (Nentidis et al., 28 Aug 2025). In the multilingual ICD-10 linking study, clinical lexicons are central: the dictionary exact match stage improves overall F1, and the best ElCardioCC result is achieved by Dict + GPT-4.1 1-shot, not by GPT-4.1 alone (Vassileva et al., 5 Sep 2025). This suggests that, at least for Greek cardiology discharge letters, language-adapted resources and controlled normalization remain structurally important even in an LLM-mediated pipeline.
The research significance of ELCardioCC in its benchmark sense is therefore twofold. First, it extends clinical coding research beyond English into Greek, an underrepresented clinical language, while retaining fine-grained ICD-10 grounding (Nentidis et al., 28 Aug 2025, Vassileva et al., 5 Sep 2025). Second, it exposes the full stack of unresolved problems in clinical coding: span detection, code assignment, hierarchical label structure, explanation, and low-resource terminology coverage. In that respect, ELCardioCC functions less as a single task than as a compact testbed for multilingual clinical NLP in cardiology.