Heart failure (HF) represents a major public health challenge, with its prevalence rising as a result of population aging and improved survival among individuals with cardiovascular conditions. Among its phenotypes, heart failure with preserved ejection fraction (HFpEF) accounts for nearly 50% of cases and is characterized by congestion and dyspnea in the absence of marked systolic dysfunction. Despite its clinical relevance, HFpEF remains diagnostically challenging due to its pathophysiological and etiological heterogeneity [1].

A key underlying cause of HFpEF is cardiac amyloidosis, an infiltrative disorder characterized by extracellular deposits of misfolded transthyretin protein that impair myocardial structure and function. The most prevalent variant is the wild type form of transthyretin cardiac amyloidosis (ATTR-CM), which predominantly affects older adults and leads to progressive cardiac dysfunction and worsening HF symptoms. Unlike other HFpEF etiologies, ATTR-CM is amenable to disease-modifying therapies that can slow progression, reduce hospitalizations, and improve survival outcomes. Nonetheless, early diagnosis remains difficult due to low clinical suspicion, nonspecific echocardiographic findings, and historically limited awareness within the medical community. Therapeutic efficacy is maximized when treatment is initiated early, underscoring the importance of timely and optimized detection strategies in HF populations [2].

In this context, artificial intelligence (AI) has emerged as a promising approach for the early identification of disease within large-scale clinical datasets [3]. Specifically, natural language processing (NLP) and deep learning models can analyze electronic health records (EHRs) to identify and extract mentions of HFpEF and associated terminology, facilitating the development of high-quality annotated corpora [4]. Such corpora are essential for training AI models capable of detecting textual patterns indicative of ATTR-CM, enabling the prioritization of high-risk patients. Applying these technologies to the structured identification of HF mentions in clinical narratives can enhance diagnostic accuracy, support data-driven decision-making, and enable timely interventions that improve prognosis and quality of life.

Despite the growing interest in clinical NLP, the development of resources in Spanish remains limited. The scarcity of expert-annotated clinical corpora in Spanish hinders the creation and evaluation of language models tailored to this language. The lack of pretrained clinical language models in Spanish, compared to the resources available for English, further restricts the development of effective applications in this domain. Recently, García Subies et al [5] conducted a study on the efficiency of encoder-based Transformer models in named entity recognition and classification tasks, aiming to identify the most effective resources in this context. The authors highlighted the significant gap in NLP resources for the Spanish language, particularly in the clinical sector. As described in their work, annotated corpora in Spanish are extremely scarce, and most are designed for named entity recognition tasks, with very few available for text classification. The authors noted that, when working with clinical narratives, encoder-based models currently outperform emerging generative language models. Their findings emphasize the urgent need to develop encoder-based models specialized in Spanish that can process clinical data with high precision. One of the most significant challenges in clinical NLP, both in general and particularly in Spanish, is the handling of negation and speculation in clinical texts. Although some initiatives such as NUBes [6] have emerged, available resources to properly address this linguistic feature remain extremely limited. As a result, this remains one of the most relevant and complex issues currently faced by the scientific community in the field.

Our study aims to develop and validate an AI-based system for the automated detection of HFpEF mentions in Spanish EHRs, with the goals of supporting early detection of potential ATTR-CM cases, enhancing diagnostic efficiency, and mitigating underdiagnosis. As outlined in [7], several research challenges remain in clinical NLP. This work addresses three of them: applying deep learning to clinical text classification, overcoming language-related barriers, and leveraging transfer learning for narrative clinical report classification.

To guide this work, we pose the following research questions (RQs):

RQ1. Can a manually annotated corpus of Spanish EHRs be effectively constructed and validated for the detection of HFpEF-related symptoms?

RQ2. What is the performance of general-purpose, biomedical-pretrained, and long-document, encoder-based Transformer models when applied to clinical text classification in Spanish?

RQ3. What is the impact of different optimization strategies (area under the curve [AUC], F₁-score, and sensitivity) on model performance for symptom detection?

RQ4. To what extent can encoder-based Transformer models support the early identification of HFpEF symptomatology indicative of cardiac amyloidosis in Spanish-language clinical narratives?

To address these questions, this study makes the following contributions: (1) the construction of a manually annotated corpus of Spanish EHRs for HFpEF symptom detection, validated by cardiology specialists and intended for public release; (2) a comparative evaluation of general, biomedical, and long-document Transformer models for clinical text classification; (3) an analysis of optimization strategies tailored to AUC, F₁-score, and sensitivity; and (4) a demonstration of the feasibility of using NLP models to support the early detection of HFpEF symptoms relevant to cardiac amyloidosis.

Transformer-based models have shown increasing effectiveness in extracting clinically relevant information from unstructured EHRs, particularly within the context of HF. Adejumo et al [8] used ClinicalBERT to extract New York Heart Association classifications and HF symptoms from clinical notes, achieving area under the receiver operating characteristic curves (AUROCs) greater than 0.98 and identifying functional status in 83% more cases than explicit documentation alone. Similarly, Liu et al [9] used ClinicalBERT embeddings within a predictive framework to identify early HF onset, outperforming traditional methods in a large-scale real-world dataset.

In the context of clinical trials, Marti-Castellote et al [10] developed a hybrid NLP approach integrating Clinical Longformer and GPT-4o to adjudicate HF-related hospitalizations. Their system reproduced expert adjudications with 83% concordance, substantially reducing manual effort and demonstrating the scalability and reliability of Transformer-based models in high-stakes research. Ahmad et al [11] provided a comprehensive overview of machine learning applications in HFpEF, emphasizing the potential of unsupervised learning techniques to identify novel patient subgroups and improve phenotypic classification. Their work underscores the importance of leveraging diverse data modalities, including clinical notes, to enhance diagnostic precision in HFpEF. Complementarily, Houssein et al [12] developed an advanced NLP framework using stacked BERT and character embeddings to detect heart disease risk factors from clinical narratives. Their model achieved an F₁-score of 93.66% on the i2b2 dataset, demonstrating the efficacy of combining deep learning techniques for extracting critical clinical information. Houssein et al [13] benchmarked 5 Transformer architectures (BERT, BioBERT, RoBERTa, XLNet, and BioClinicalBERT) on the i2b2 dataset for risk factor extraction in heart disease, achieving state-of-the-art performance with a micro F₁-score of 94.26%. These results reinforce the value of domain-specific Transformer models in clinical NLP applications. Notably, Fan et al [14] implemented a Transformer-based clustering framework to identify 7 HF subtypes in a cohort of more than 379,000 patients. These subgroups, some of which were independent of left ventricular ejection fraction, highlight the capacity of deep learning approaches to uncover novel and clinically meaningful HF phenotypes, which is particularly relevant given the heterogeneity of HFpEF. These studies highlight the growing role of machine learning and NLP in advancing HF research and patient care. In the context of the Spanish language, recent work such as that of García Subies et al [5] has explored the use of encoder-based Transformer models for named entity recognition and classification tasks in Spanish clinical texts. However, their efforts have focused on the evaluation and application of language models on existing corpora and have not addressed the detection of HFpEF-specific symptoms.

Despite these advances, most of the existing research focuses on English-language EHRs, frequently relying on structured data or datasets with broad cardiovascular end points. To date, only a limited number of studies have addressed the automated detection of HFpEF-specific symptoms and none, to our knowledge, have done so in Spanish-language clinical narratives. Moreover, existing models are rarely optimized for clinically critical metrics such as sensitivity, nor do they account for the long and complex structure of real-world clinical narratives. This study addresses these gaps by introducing a manually annotated Spanish-language corpus for HFpEF symptom detection, evaluating multiple Transformer architectures, including models adapted for long documents, and implementing task-specific optimization strategies aligned with diagnostic priorities. This represents the first demonstration of the feasibility and clinical applicability of Transformer-based NLP for detecting HFpEF in Spanish EHRs, with direct implications for improving early recognition of cardiac amyloidosis.

This study was approved by the Andalusian Biomedical Research Ethics Coordinating Committee under protocol version 1 dated October 2, 2020 (internal protocol code 2382-N-20). The corpus documents were anonymized using a rigorous, tailor-made protocol specifically designed to protect patient data privacy.

The Transformer architecture has become a cornerstone in NLP, offering high efficiency and versatility across a wide range of tasks. Its ability to model long-range dependencies through attention mechanisms has revolutionized the field, driving significant advancements in tasks such as text classification [17-19]. This architecture is distinguished not only by its accuracy but also by its scalability and adaptability across diverse domains and datasets of varying sizes. Transfer learning allows for domain-specific adaptation through fine-tuning, as applied here to the health care domain [20]. By leveraging pretrained models that have already learned general language representations from large, diverse corpora, transfer learning facilitates their specialization for more domain-specific datasets through additional training. This adaptability has established Transformers as a powerful tool for advancing NLP in specialized fields.

A key limitation of Transformer models is their inherent constraint in processing large texts. Their architecture is typically designed to handle a maximum of 512 tokens, posing challenges for tasks that require longer sequences, such as document-level analysis or summarization of extensive texts. To address this limitation, the Longformer architecture [21], specifically designed to efficiently handle longer text sequences, was introduced. Longformer incorporates sparse attention mechanisms, allowing it to extend the context window while preserving computational efficiency, thus mitigating a core limitation of traditional Transformer models. This innovation broadens the applicability of Transformer models to tasks involving lengthy documents, such as EHRs. These considerations motivated the inclusion of 3 model categories in this study: general-purpose Transformers, domain-adapted clinical models, and Longformers. This categorization allows for a comprehensive comparison of performance across different levels of domain specialization and sequence-length flexibility. General-purpose models serve as a baseline, while fine-tuned clinical models assess the added value of domain-specific adaptation. Longformers, by contrast, address the limitations of standard sequence lengths and offer insight into the role of extended context modeling in clinical NLP.

In clinical contexts, maximizing recall for the positive class (Recall-1) is often prioritized to capture as many at-risk cases as possible. However, this strategy can lead to an increased number of false positives, elevating alert rates and potentially burdening health care systems. To address this trade-off, we incorporated additional evaluation metrics, specifically the AUC and the F₁-score, to support a more balanced classification approach. Each metric provides a distinct perspective: AUC assesses the model’s overall ability to discriminate between positive and negative cases, while F₁-score balances precision and sensitivity, helping to reduce both false positives and false negatives.

This study shows that the performance of Transformer-based models in HFpEF symptom detection can be substantially influenced by the chosen optimization strategy. Each approach targeting (AUC, F₁-score, or sensitivity) serves a distinct clinical and operational objective. Notably, the sensitivity-driven strategy proved particularly effective in minimizing false negatives, a critical consideration in early disease detection scenarios. This highlights the value of aligning model training objectives with specific clinical priorities when developing NLP tools for medical applications.

It should be noted that direct comparisons with related studies are not feasible, as each uses different datasets. Nevertheless, the strength of the results confirms the validity of the experimental design and the robustness of the trained models under the proposed optimization strategies, as reflected in the evaluation metrics. Additionally, the involvement of cardiology specialists provides expert validation and qualitative support for these findings.

Beyond retrospective evaluation, two real-world integration scenarios are envisioned: (1) batch-mode screening, where the model is periodically applied to existing EHR data to prioritize patients for further testing, and (2) real-time alerting, where symptom mentions in clinical notes trigger nonbinding alerts to the clinician. In either case, the system would function as a clinical decision support tool. To mitigate false positives, adjustable probability thresholds and explanatory interfaces that highlight the relevant text segments are recommended. These strategies align with current frameworks for trustworthy and transparent AI in health care.

A qualitative error analysis was performed from a clinical perspective, focusing on both false negative and false positive cases. Special attention was given to false negatives to identify the reasons why the models predicted the absence of the symptom in a document (label “0”) when, in reality, it had been classified as positive, meaning that the document contained information indicating the presence of the symptom. We selected the misclassified documents from the 3 models with the lowest false negative rates (see Figure 5).

Textbox 1 presents a false negative example, in which all models predicted the case as negative, despite human annotation indicating the presence of the symptom. However, a review of the annotation guidelines revealed that the annotator’s interpretation relied on clinical reasoning that, while valid, extended beyond the defined annotation criteria. The text includes terms such as “cardiomegalia, signos de hipertensión pulmonar precapilar y poscapilar con infiltrados de edema intersticial alveolar bilaterales” (“cardiomegaly, signs of pre- and post-capillary pulmonary hypertension with bilateral interstitial-alveolar edema infiltrates”), which can be associated with HFpEF. However, the guidelines did not explicitly state that these findings alone were sufficient for a positive classification without additional diagnostic information or clinical context. In this regard, the model’s prediction strictly adhered to the established criteria, whereas the human annotation reflected a broader interpretation based on implicit clinical knowledge. This example highlights the inherent complexity of symptom detection and the subtle distinction between rule-based annotation and clinically contextual judgment. Despite these occasional discrepancies, the models have demonstrated a high level of consistency in applying the predefined criteria. Future revisions of the annotation process could explore potential adjustments to better align guideline application with the expected clinical reasoning for this task.

Conversely, Textbox 2 shows an example of a false positive error. In this case, the models incorrectly predicted the presence of HFpEF symptoms in a document explicitly indicating their absence. Although the key phrase “no existe daño renal ni signos de ICC” (“no kidney damage or signs of CHF”) clearly indicated the negative class, the models failed to interpret the negation correctly in this context. Such errors are a well-known challenge in clinical NLP, where negation is often expressed using diverse and context-dependent linguistic patterns. Notably, all 3 models produced the same misclassification, indicating that the difficulty in interpreting negation structures is model-independent.

Despite these specific instances, the overall performance of the models has been consistent across most evaluated cases. However, future improvements could focus on refining negation handling through domain-adapted preprocessing pipelines or model adaptations that explicitly account for linguistic cues. In particular, the integration of rule-based algorithms, enhanced tokenization schemes, or syntactic features could support more accurate interpretation of negated or speculative statements. These enhancements represent promising avenues for improving model robustness in real-world clinical applications.

From a clinical standpoint, accurately identifying patients with suspected HFpEF is critical for ensuring appropriate management. Classification errors may result in the omission of key diagnostic procedures or delays in initiating disease-modifying therapies. In particular, false negatives represent a major concern, as they may leave early-stage patients undiagnosed and untreated. Conversely, false positives may prompt unnecessary diagnostic procedures, increasing both clinical workload and health care costs. Therefore, tuning strategies that prioritize sensitivity while preserving acceptable specificity are essential for the effective deployment of these models in clinical settings.

This study has several limitations that should be acknowledged. First, all clinical data were obtained from a single hospital in Spain, using one specific EHR system. As a result, the trained models reflect documentation practices, linguistic patterns, and institutional conventions specific to that context. This may limit the generalizability of the findings to other health care settings, regions, or EHR platforms. Second, no external validation was performed on datasets from other institutions or countries. Although the models demonstrated strong internal performance, further evaluation on external corpora will be essential to assess their robustness and adaptability to different clinical environments and documentation styles. As highlighted in recent studies, external validation is a critical step to ensure the reliability and applicability of NLP models across diverse health care systems [29]. Despite these limitations, this work provides a valuable foundation for developing NLP systems tailored to Spanish-language clinical narratives and highlights the feasibility of symptom detection in the context of cardiac amyloidosis. Future research should focus on validating and refining these models in multicenter, cross-national settings to enhance their clinical relevance.

This study was conducted using anonymized clinical data in accordance with the General Data Protection Regulation (regulation [EU] 2016/679) and Spanish data protection laws (Ley Orgánica 3/2018 de Protección de Datos Personales y garantía de los derechos digitales). Data use was approved by the institutional ethics committee. Looking forward, any clinical deployment of the proposed AI system would need to comply with the EU Artificial Intelligence Act and, depending on its intended function, may also fall under the scope of the Medical Device Regulation (regulation [EU] 2017/745). Consequently, future implementations will require robust validation, human oversight mechanisms, and explainability features to ensure safe and ethical integration into clinical practice.

This study proposed an experimental framework for optimizing Transformer-based language models to automatically detect clinical documents suggesting the presence of HFpEF. Our findings underscore the importance of hyperparameter tuning, as optimal configurations directly influence model performance according to the prioritized evaluation metric. Specifically, to minimize false negatives, optimization should prioritize model sensitivity. This approach allows model performance to be adapted to varying clinical and operational requirements, maintaining a suitable balance between sensitivity and specificity depending on the context of use.

It has been demonstrated that using of domain-specific pretrained language models significantly improves adaptability and transferability for specialized clinical tasks. This advantage lies in the ability of context-aware models to better capture linguistic patterns, domain-specific terminology, and semantic relationships, thereby enhancing task-specific performance. Moreover, the capacity to process longer sequences during fine-tuning and inference is especially beneficial in the clinical domain. Medical documents are often lengthy and rich in interrelated concepts, which may be lost when contextual processing is constrained. Therefore, using models designed for extended input sequences is essential for comprehensive and accurate clinical text analysis.

Having validated our experimental approach and confirmed model reliability for HFpEF detection, future work will focus on developing and fine-tuning models capable of comprehensively identifying all symptoms associated with cardiac amyloidosis. To this end, the top-performing models from this study will be retrained on manually annotated corpora to incorporate the full range of clinically relevant manifestations. This process will improve the system’s ability to detect textual patterns indicative of the disease with increased precision and sensitivity. The ultimate aim is to implement an automated screening system based on EHR analysis to support the early identification of patients with suspected cardiac amyloidosis. This approach may optimize diagnostic workflows, improve clinical decision-making, and enable more personalized treatment strategies for a condition that remains difficult to detect at early stages.