A NOVEL deep learning-based language model has demonstrated high precision and accuracy in identifying heart failure with reduced ejection fraction (HFrEF) from hospital discharge summaries, paving the way for improved automated quality care assessments.
The lack of effective tools to automate quality care assessment in heart failure management has limited the implementation of national guidelines for HFrEF, a condition defined by a left ventricular ejection fraction (LVEF) of less than 40%. Researchers aimed to address this gap by developing a deep-learning natural language processing model to identify HFrEF patients from discharge summaries. By automating this process, the model could significantly enhance the evaluation of care quality and promote guideline-directed therapy for affected patients.
The model was trained and tested on 13,251 notes from 5,392 individuals hospitalized for heart failure at Yale New Haven Hospital between 2015 and 2019. Among the patients, 2,487 (46.1%) met the criteria for HFrEF. The model achieved outstanding performance, with an AUROC and AUPRC of 0.97 each in the held-out dataset. External validation on 19,242 notes from Northwestern Medicine yielded an AUROC of 0.94 and AUPRC of 0.91, while evaluations at Yale community hospitals and the MIMIC-III database also demonstrated excellent accuracy (AUROCs of 0.95 and 0.91, respectively). Importantly, the model achieved a net reclassification improvement of 60.2% compared to diagnosis codes, highlighting its ability to better identify HFrEF cases.
These findings underscore the potential of deep-learning tools in improving clinical practice by streamlining the identification of patients with HFrEF. This approach supports more comprehensive evaluations of care quality, potentially leading to increased adherence to guideline-directed medical therapies. Future efforts should focus on expanding these methods to other healthcare systems and conditions, as well as integrating them into real-time clinical workflows. Such advancements could contribute significantly to enhancing patient outcomes and standardising care delivery in HFrEF management.
Katrina Thornber, EMJ
Reference
Nargesi AA et al. Automated identification of heart failure with reduced ejection fraction using deep learning-based natural language processing. J Am Coll Cardiol HF. 2025;13(1):75-87.
