Multicriteria Optimization of Language Models for Heart Failure With Preserved Ejection Fraction Symptom Detection in...

Area of Science:

• Natural Language Processing
• Artificial Intelligence in Medicine
• Clinical Informatics

Background:

• Heart failure with preserved ejection fraction (HFpEF) is a key indicator of cardiac amyloidosis, often underdiagnosed due to vague symptoms.
• Electronic health records (EHRs) offer potential for early symptom detection via NLP, but challenges exist in Spanish EHRs.
• Lack of annotated corpora and domain-specific models hinders NLP application in Spanish clinical narratives.

Purpose of the Study:

• To evaluate NLP feasibility for identifying HFpEF phenotypes in Spanish EHRs for early cardiac amyloidosis detection.
• To assess the impact of domain-specific language models and optimization strategies on symptom detection reliability, sensitivity, and generalizability.
• To improve early diagnosis of cardiac amyloidosis through advanced NLP techniques in Spanish-speaking healthcare settings.

Main Methods:

• Developed and validated a novel corpus of 15,304 Spanish clinical documents annotated by cardiology experts.
• Evaluated 8 Transformer-based language models, including general and biomedical-specialized variants (e.g., Longformer).
• Applied 3 clinically guided optimization strategies (AUC, F1-score, sensitivity) during fine-tuning on unseen patient data.

Main Results:

• All evaluated models demonstrated high performance, with AUC > 0.940.
• The top model, Longformer Biomedical-clinical, achieved AUC 0.987, F1-score 0.985, sensitivity 0.987, and specificity 0.987.
• Sensitivity optimization reduced the false-negative rate below 3%, crucial for clinical safety.

Conclusions:

• Transformer models reliably detect HFpEF symptoms in Spanish EHRs, overcoming class imbalance and linguistic complexity.
• Domain-specific pretraining, long-context architectures, and clinical optimization enhance classification performance, especially sensitivity.
• These models show potential as effective screening tools for prioritizing patients at risk for cardiac amyloidosis in Spanish-speaking regions.