Cardio-NLP: Integrating Language & Cardiology

• Cardio-NLP is the application of natural language processing techniques to extract, analyze, and interpret cardiovascular data from clinical narratives and signals.
• It leverages a range of methods from rule-based pipelines to deep learning, successfully addressing tasks like extraction, risk prediction, and report generation with notable metrics.
• The field faces challenges in portability, reproducibility, and standardization across institutions, driving the need for hybrid, multimodal, and interpretable system designs.

Searching arXiv for recent and foundational CARDIO-NLP papers to ground the article. Cardio-NLP is the application of natural language processing to cardiovascular data, narratives, and reporting workflows, spanning clinical text mining, concept-value extraction from cardiology reports, language-grounded modeling of cardiac signals, multimodal risk prediction, and report generation. Across the literature, the term encompasses at least four distinct but related problem families: extraction of cardiovascular findings from unstructured clinical narratives; prediction or phenotyping from cardiology-related text; alignment of cardiac signals such as electrocardiograms with language representations; and generation of clinically styled reports or explanations from cardiovascular signals or multimodal inputs. The field is methodologically heterogeneous, ranging from rule-based pipelines and conventional supervised classifiers to Transformer encoders, multimodal deep learning, and LLMs, with substantial variation in task definition, supervision, reproducibility, and portability across institutions (Adekkanattu et al., 2019, Qiu et al., 2023, Yang et al., 19 Oct 2025).

1. Domain definition and scope

Cardio-NLP is not a single task. In the reviewed literature, NLP in cardiology spans identification and classification, prediction, information extraction, automation and model evaluation, and text-guided generation (Yang et al., 19 Oct 2025). The largest data source category is electronic health records, which account for 82.6% of reviewed studies, with physician progress notes, discharge summaries, cardiac imaging reports, and ECG reports representing major subtypes (Yang et al., 19 Oct 2025). This concentration reflects a basic property of cardiology practice: clinically consequential information is often dispersed across consultation reports, echocardiography reports, ECG interpretations, radiology narratives, and longitudinal notes rather than confined to structured fields.

A central distinction within Cardio-NLP is between text-centric and language-grounded multimodal work. Text-centric studies operate directly on physician notes, discharge letters, or echocardiography reports to extract findings or predict disease states. Examples include rule-based concept-value extraction from echocardiogram reports across institutions (Adekkanattu et al., 2019), machine-learning detection of cardiac failure from French physician narratives (Le et al., 2021), and multimodal risk prediction from chest X-ray reports plus structured predictors (Bagheri et al., 2020). Language-grounded multimodal studies instead use text as supervision or as an embedding space for non-text cardiac modalities. Examples include transferring language-model representations to ECGs for diagnosis report generation and zero-shot-style disease detection (Qiu et al., 2023), multilingual ECG captioning (Kiyasseh et al., 2021), and LLM-mediated symptom analysis for cardiovascular risk prediction from free text (Yang et al., 15 Jul 2025).

The reviewed cardiology-NLP landscape is also disease-diverse. A narrative review of the field reports that 35.3% of studies address general cardiovascular disease, 27.1% heart failure, 12.0% coronary artery disease, 10.9% arrhythmias, and 7.0% structural heart disease (Yang et al., 19 Oct 2025). This suggests that Cardio-NLP is best understood as an umbrella field organized around data types and tasks rather than a single disease program.

2. Information extraction from cardiology narratives

Information extraction is one of the most mature Cardio-NLP subareas, particularly for structured or semi-structured reports. A prominent example is the evaluation of EchoExtractor, a Department of Veterans Affairs rule-based system built within Leo, a UIMA-based framework, for extracting concept-value pairs from echocardiogram reports (Adekkanattu et al., 2019). The task was not simple mention detection. Reviewers identified all mentions of a concept and the associated quantitative or qualitative values, and the evaluation required matching a concept together with its value, such as “no,” “mild,” “moderate,” “dilated,” or a numeric measurement (Adekkanattu et al., 2019). This formulation is clinically important because, in cardiology, “aortic stenosis” is incomplete without its severity, polarity, or associated quantitative finding.

The multi-site echocardiography study demonstrates both the promise and fragility of this extraction paradigm. The original system targeted 27 concepts, but 24 were formally evaluated across Weill Cornell Medicine, Mayo Clinic, Northwestern Medicine, and MIMIC-III because three concepts were initially absent at one site (Adekkanattu et al., 2019). The 24 evaluated targets included left ventricular ejection fraction, valvular regurgitation and stenosis, atrial and ventricular dimensions, pulmonary artery pressure, and valve orifice area (Adekkanattu et al., 2019). The methodology was explicitly rule-based: concept detection used regular expression matching and a custom term lookup dictionary; value detection used regular expressions for quantitative and qualitative values; and concept-value relation extraction used pattern matching over echocardiography-domain patterns (Adekkanattu et al., 2019).

The main finding was limited portability. Each site installed the existing VA EchoExtractor “without any system modifications whatsoever,” and only a small subset of concepts transferred robustly (Adekkanattu et al., 2019). Across all four datasets, only aortic valve regurgitation and left atrium size at end systole achieved $F$-score $> 0.90$ (Adekkanattu et al., 2019). Across Weill Cornell Medicine, Mayo, and Northwestern, the concepts with $F$-score $> 0.90$ were aortic valve regurgitation, left atrium size at end systole, mitral valve regurgitation, and tricuspid valve regurgitation (Adekkanattu et al., 2019). Left ventricular ejection fraction illustrates site sensitivity especially clearly: $F$-score 0.99 at Weill Cornell Medicine, 0.99 at Northwestern, 0.95 at MIMIC, but at Mayo precision was 99% and recall only 55%, largely because semicolons rather than colons disrupted concept-value linking in phrases like “Calculated left ventricular ejection fraction; 65 %” (Adekkanattu et al., 2019).

The error analysis in that study defines several enduring extraction challenges in Cardio-NLP. These include term mismapping, institution-specific synonyms, semi-structured table parsing failure, confusion between measured values and reference ranges, punctuation and line-break interference, and relation extraction failure when multiple concepts occur in close proximity (Adekkanattu et al., 2019). The paper quantifies one especially consequential pathology: for aortic valve orifice area, 23% of reports were false positives because reference ranges were extracted as measurements; for mitral valve orifice area, the corresponding false-positive rate was 94% (Adekkanattu et al., 2019). This supports a broader conclusion: in cardiology report NLP, semantics may be standardized, but documentation practices are not (Adekkanattu et al., 2019).

3. Predictive modeling from clinical text and multimodal EHRs

Another major strand of Cardio-NLP uses unstructured narratives as predictors for disease detection, prognosis, or risk stratification. A focused example is the detection of cardiac failure from French-language physician notes in a pediatric intensive care setting (Le et al., 2021). This study used admission notes and evaluation notes from CHU Sainte-Justine Hospital, manually labeled by two physicians as positive or negative for cardiac failure based on clinical indicators including inotrope exposure, diagnoses such as dilated cardiomyopathy or cardiogenic shock, and physiologic surrogates such as $EF < 50\%$, $FR < 25\%$, or pro-BNP $> 1000$ ng/L (Le et al., 2021). The pipeline combined lowercasing, stop-word removal, tokenization, one of three document representations—BoW, TF-IDF, or word2vec-style embeddings—and one of three classifiers—logistic regression, Gaussian Naive Bayes, or multilayer perceptron neural network (Le et al., 2021). The best-performing configuration was TF-IDF + MLP-NN, achieving Accuracy = 0.87, Precision = 0.86, Recall = 0.88, and F1 = 0.87 with feature selection (Le et al., 2021). The paper’s own interpretation is that in this relatively small French PICU corpus, sparse lexical features outperformed learned dense embeddings (Le et al., 2021).

A different use case is multimodal cardiovascular risk prediction from radiology narratives and structured predictors (Bagheri et al., 2020). In a cohort of 5603 patients and their reports from the SMART study at UMC Utrecht, the task was binary prediction of major adverse cardiovascular events during follow-up using chest X-ray radiology report text and structured cardiovascular risk variables (Bagheri et al., 2020). The model family included a structured-only neural network, a text-only BiLSTM, and three multimodal variants: MI-CNN, MI-LSTM, and MI-BiLSTM (Bagheri et al., 2020). The unimodal text model performed weakly, with AUC 0.570, while the structured-only neural model reached AUC 0.651. The main result was that text became useful when fused with structured variables: MI-CNN reached 0.730, MI-LSTM 0.794, and MI-BiLSTM 0.847 AUC (Bagheri et al., 2020). This supports a recurring CARDIO-NLP pattern: clinical text often contributes complementary signal rather than dominating prediction alone (Bagheri et al., 2020).

More recent LLM-oriented work pushes risk prediction toward symptom reasoning. “LLM-Augmented Symptom Analysis for Cardiovascular Disease Risk Prediction” describes a pipeline using domain-adapted LLMs for symptom extraction, contextual reasoning, and correlation from free-text reports, with Bio_ClinicalBERT embeddings + lightweight classifier as the most concrete implementation (Yang et al., 15 Jul 2025). The paper reports accuracy: 85.7%, recall: 83.3%, F1-score: 85.3% on a “small balanced test set,” while also stating “F1-score = 86.55% and more” in the conclusion (Yang et al., 15 Jul 2025). Cardiologist evaluation involved three board-certified cardiologists, 20 anonymized test cases, mean clinical relevance 4.3/5, and Cohen’s kappa = 0.82 (Yang et al., 15 Jul 2025). However, the same paper explicitly states that experiments used synthetic examples, and it does not provide enough detail on MIMIC-III or the named CARDIO-NLP dataset to reconstruct the cohort or labels (Yang et al., 15 Jul 2025). This suggests an application-driven proof of concept rather than a benchmark-grade study.

CardioAI extends the predictive and monitoring logic into cardio-oncology. It combines Garmin wearable data, an LLM-powered voice assistant built on GPT-4 with retrieval-augmented generation, daily summarization, and an explainable cardiotoxicity risk prediction module in a system for cancer treatment-induced cardiotoxicity monitoring (Wu et al., 2024). The evaluation is workflow-oriented rather than predictive: SUS 72.33 ± 1.89, NASA-TLX effort 1.13 ± 0.70, mental demand 1.13 ± 0.23, physical demand 0.73 ± 0.42 (Wu et al., 2024). The system is important in Cardio-NLP because it treats symptom language as an essential remote-monitoring modality, but the paper does not provide enough mathematical or modeling detail for the Transformer-Weibull risk module or for the GPT-4 prompting policy (Wu et al., 2024).

4. Language-grounded modeling of cardiac signals

A major expansion of Cardio-NLP reframes cardiac signal understanding as a language-grounded representation-learning problem. One early example is “Transfer Knowledge from Natural Language to Electrocardiography,” which trains a hierarchical Transformer ECG encoder on PTB-XL and aligns ECG-derived embeddings with expert report embeddings using a pretrained LLM and an Optimal Transport loss (Qiu et al., 2023). The processed ECG input dimension is 864, consisting of 600 ECG signals plus 264 time/frequency-domain features (Qiu et al., 2023). The architecture was evaluated on two downstream tasks: automatic ECG diagnosis report generation and zero-shot cardiovascular disease detection (Qiu et al., 2023). With Transformer + BERT, the report-generation metrics were BLEU-1 26.93, ROUGE-1 F 28.08, METEOR 21.23, BERTScore 88.90; the zero-shot-style disease detection results were Accuracy 0.77, AUROC 0.92, and F1 0.68 (Qiu et al., 2023). Among tested LLMs, BART was slightly best for report generation with BLEU-1 27.21, ROUGE-1 F 29.56, METEOR 24.51, BERTScore 89.61, while BioClinical BERT was best for disease detection with Accuracy 0.78 and F1 0.71 (Qiu et al., 2023). The paper is conceptually influential because it treats expert ECG reports as semantic supervision, but several critical components—such as the exact zero-shot matching rule and the decoding mechanism for non-generative encoders—are underdefined (Qiu et al., 2023).

Multilingual report generation from ECGs develops this line further. “Let Your Heart Speak in its Mother Tongue” formulates ECG-to-text generation as multilingual captioning, using PTB-XL ECGs paired with reports in seven languages: German, Greek, English, Spanish, French, Italian, and Portuguese (Kiyasseh et al., 2021). Each ECG recording is segmented into 5-second non-overlapping segments of 2500 samples, and patient-level splits yield 22,670 train instances, 3,284 validation instances, and 3,304 test instances (Kiyasseh et al., 2021). The architecture combines a CNN-based ECG encoder with a Transformer decoder, while multilinguality is handled via a shared encoder-decoder backbone plus language-specific output heads (Kiyasseh et al., 2021). The paper’s distinctive pretraining method, Replaced Token Language Prediction (RTLP), performs on par with MLM and MARGE: average validation BLEU-1 29.3, METEOR 34.3, ROUGE-L 33.4 for RTLP-uniform, versus 29.4, 34.6, 33.5 for MLM (Kiyasseh et al., 2021). The same work also reports a “blessing of multilinguality,” where multilingual fine-tuning improves German report generation under RTLP relative to monolingual fine-tuning (Kiyasseh et al., 2021). The generated reports were evaluated with BLEU-1, METEOR, ROUGE-L, and Self-BLEU, but no cardiologist factuality assessment was reported (Kiyasseh et al., 2021).

Signal-to-language inspiration also appears in papers that are not full report-generation systems. “Enhancing Electrocardiogram Signal Analysis Using NLP-Inspired Techniques” treats ECG signals as language-like sequences and introduces a Temporal Autoencoder producing 6-dimensional latent embeddings for each temporal token, followed by a CNN-LSTM-Self Attention classifier (Ganguly et al., 2024). On PTB-XL, after SMOTE balancing to 45415 samples, the model achieved Accuracy = 91%, with per-class F1 scores ranging from 0.85 to 0.95 across CD, HYP, MI, NORM, and STTC (Ganguly et al., 2024). The ablation shows that adding the TAE compression/embedding improves the best CLSA configuration from 87.0% to 91.0% accuracy (Ganguly et al., 2024). This suggests that language-inspired embedding and context mechanisms can improve ECG classification even without explicit text supervision.

A different kind of language-grounded signal work is LLM-guided source separation. “LLM-based Nonnegative Matrix Factorization For Cardiorespiratory Sound Separation” introduces LingoNMF, in which a LLaMA-2-7b-chat model both interprets extracted acoustic features and updates the target fundamental-frequency vector in an NMF cost function (Torabi et al., 9 Feb 2025). The method was tested on 100 synthesized mixtures of real measurements and 210 recordings from a clinical manikin (Torabi et al., 9 Feb 2025). The penalty weight was tuned to $\lambda = 0.01$, with an initial $f = 50\,\text{Hz}$ for both sources (Torabi et al., 9 Feb 2025). The work is relevant to Cardio-NLP because the LLM is part of the optimization loop, not merely a post hoc reporter, but the interpretive disease-prediction component remains qualitative rather than benchmarked (Torabi et al., 9 Feb 2025).

5. LLMs, multilinguality, long context, and multimodal reasoning

The rise of LLMs has expanded Cardio-NLP beyond extraction and classification into prompting, zero-shot reasoning, summarization, explanation, and multilingual generation (Yang et al., 19 Oct 2025). Yet the field’s recent work repeatedly shows that generative flexibility does not eliminate domain-specific modeling requirements.

A clear example is large-context clinical classification. “Automatic Cardiac Risk Management Classification using large-context Electronic Patients Health Records” addresses binary cardiovascular risk management eligibility in a Dutch geriatric outpatient cohort of 3,482 patients, using one year of longitudinal consultation reports, medication embeddings, and anthropometric variables (Vitale et al., 10 Mar 2026). The study benchmarks classical ML, a custom hierarchical Transformer, and zero-shot general-purpose LLMs. The custom Hierarchical Transformer 1D with CLS-token pooling and late fusion achieved the best reported performance, with F1 = 92.48% (0.67) and MCC = 0.758 (0.024) (Vitale et al., 10 Mar 2026). The sequence length was padded or truncated to 8192 tokens, the encoder used 3 Transformer layers, 4 attention heads, RoPE, and hierarchical attention (Vitale et al., 10 Mar 2026). The paper’s central conclusion is that a task-specific long-context encoder outperforms zero-shot general-purpose LLMs on this non-English, longitudinal, clinically nuanced classification problem (Vitale et al., 10 Mar 2026). This suggests that, in Cardio-NLP, large context alone does not guarantee robust inference; architecture-task alignment remains critical.

Multimodal prompting under sparse supervision is explored in “Few-Label Multimodal Modeling of SNP Variants and ECG Phenotypes Using LLMs for Cardiovascular Risk Stratification” (Menon et al., 18 Oct 2025). The cohort is reported as 8,856 participants, though the tier counts sum inconsistently, and only 350 participants have confirmed cardiac diagnoses (Menon et al., 18 Oct 2025). The pipeline uses TF-IDF encoding of SNP rsIDs, a transformer-based ECG encoder over 12 ECG features (QC), k-means clustering with $> 0.90$0 for pseudo-labeling, LoRA with rank = 8 and alpha = 16, and Chain-of-Thought prompting over serialized multimodal evidence (Menon et al., 18 Oct 2025). In ablations, removing ECG harms recall while removing SNP harms precision, and the full multimodal DeepSeek 1.3B configuration reports Accuracy 0.920, Precision 0.831, Recall 0.810, F1 0.820 under the baseline setting (Menon et al., 18 Oct 2025). The work is notable for its few-label and prompt-serialization strategy, but its label ontology, pseudo-label refinement rule, and evaluation consistency are insufficiently specified (Menon et al., 18 Oct 2025).

A broader narrative review situates such work historically. It reports that text-guided generation emerged in 2022, and that LLMs since then have expanded cardiology NLP into prediction, information extraction, text generation, AI assistants, and retrieval-augmented generation (Yang et al., 19 Oct 2025). At the same time, the review emphasizes unresolved issues of interpretability, trustworthiness, and privacy, and identifies interpretable LLMs, multimodal methods, and open-source ecosystems as principal future directions (Yang et al., 19 Oct 2025).

6. Portability, evaluation problems, and future directions

Portability and reproducibility remain fundamental challenges in Cardio-NLP. The echocardiography portability study is explicit that software portability is not linguistic portability: all sites installed EchoExtractor within about a day, but true extraction generalization was limited (Adekkanattu et al., 2019). Local differences in vocabulary, punctuation, table structure, abbreviations, section order, and reporting style materially degraded performance (Adekkanattu et al., 2019). This is a caution against assuming that high within-site performance on cardiology narratives will transfer “as is” to other institutions.

Benchmark design is another recurring problem. Several papers omit inter-annotator agreement, confidence intervals, or full cohort definitions. The pediatric cardiac failure study reports no inter-annotator agreement and does not specify the number of cross-validation folds (Le et al., 2021). The symptom-risk LLM paper mentions MIMIC-III and a CARDIO-NLP dataset but does not provide cohort construction, note types, split sizes, or label derivation, while also acknowledging synthetic examples (Yang et al., 15 Jul 2025). The ECG-language alignment paper does not fully specify its zero-shot disease scoring rule or generation pipeline for BERT-like encoders (Qiu et al., 2023). The few-label SNP+ECG study contains internal count inconsistencies and leaves pseudo-label refinement underdefined (Menon et al., 18 Oct 2025). These patterns suggest that Cardio-NLP still contains many proof-of-concept systems whose translational significance exceeds their reproducibility.

The field also faces a persistent tension between performance and interpretability. Rule-based systems remain competitive on narrow extraction tasks precisely because they are transparent and auditable (Adekkanattu et al., 2019, Yang et al., 19 Oct 2025). Deep and multimodal models often improve coverage or end-to-end capability, but many provide limited mechanistic explanation. Even when models output rationales or use Chain-of-Thought prompting, this does not by itself establish faithful reasoning. This suggests that future progress may depend on hybrid designs: rule-based extraction for stable structured targets, neural encoders for long-range context and multimodal fusion, and carefully controlled LLM interfaces for explanation or generation.

Another likely direction is deeper multimodal grounding. CardioLab shows that ECG waveforms plus demographics, biometrics, and vital signs can estimate or monitor many laboratory abnormalities, with AUROC > 0.70 in a statistically significant manner for 23 laboratory values in the estimation setting and up to 26 values in the monitoring setting (Alcaraz et al., 2024). Although not a text-based study, it points toward a broader CARDIO-NLP future in which language systems interface with physiologic and structured clinical signals rather than operating on text alone. PULSE makes a similar argument from the imaging side, proposing a unified architecture for segmentation, classification, and “clinically grounded text output,” though its language component remains largely aspirational and unevaluated (Ghouse et al., 3 Dec 2025). These examples suggest that the next phase of Cardio-NLP may be less about isolated note mining and more about language as the representational, explanatory, or supervisory layer connecting text, imaging, waveforms, labs, symptoms, and longitudinal EHR structure.

In sum, Cardio-NLP has evolved from rule-based extraction of cardiology report content into a broader family of methods for cardiovascular information extraction, risk prediction, signal-language alignment, and clinical text generation. Its strongest mature use cases remain structured finding extraction and clinically grounded classification from narratives. Its most ambitious frontier lies in language-grounded multimodal reasoning over ECGs, echocardiograms, imaging, symptoms, and longitudinal records. The literature suggests that future robust systems will require site-aware adaptation, clearer benchmarks, stronger external validation, and careful integration of LLMs with domain-specific representations rather than reliance on prompting alone (Adekkanattu et al., 2019, Qiu et al., 2023, Yang et al., 19 Oct 2025).