More than words: AI takes NLP to the next level to identify signs of heart failure
Advanced artificial intelligence (AI) algorithms can be trained to scan electronic health records for clinical evidence of heart failure with reduced ejection fraction (HFrEF), according to a new analysis published in JACC: Heart Failure. This suggests that clinicians will be able to identify high-risk patients earlier than ever, ensuring they receive the care they need to live long, healthy lives.
Of course, AI models have been able to flag signs of HFrEF in patient records for quite some time thanks to the power of natural language processing (NLP). What makes this team’s research stand out is the fact that its new NLP algorithm excels where others have struggled in the past.
“Previous NLP-based models have been limited to semantic algorithms that search for prespecified terms with unclear performance in identifying HFrEF,” wrote co-first authors Arash A. Nargesi MD, with Brigham and Women’s Hospital, and Philip Adejumo, BS, with Yale School of Medicine, and colleagues. “Misspelling of disease diagnosis, frequent use of abbreviations in clinical documentation, and variations in textual descriptions of clinical entities such as 'heart failure with recovered ejection fraction' are barriers to properly identifying the heterogeneous population of patients with HFrEF. In this study, we developed and externally validated a state-of-the-art deep-learning language model that identifies HFrEF through contextual analysis of clinical notes. Our model can effectively process long notes without being limited to predefined terms.”
The group used more than 13,000 clinical notes from the discharge summaries of more than 5,000 patients to develop this new-look NLP algorithm. All patients were hospitalized for heart failure-related symptoms from 2015 to 2019 at a single high-volume hospital. The mean patient age was 73 years old, and 52% were men. More than 46% of patients were diagnosed with HFrEF based on echocardiography-confirmed left ventricular ejection fraction data.
The AI model was trained on 70% of the data and then tested on the remaining 30%. Overall, it was associated with an area under the ROC curve (AUROC) of 0.97 when asked to scan patient records for signs of HFrEF. The group then externally validated their algorithm, finding that it had an AUROC of 0.94 when used on more than 19,000 notes from Northwestern Medicine and an AUROC of 0.91 when used on 146 manually reviewed notes from the Medical Information Mart for Intensive Care III database.
“The algorithm demonstrated high discrimination in identifying HFrEF across demographic subgroups of age, sex, race, cardiovascular comorbidities, and discharge summaries of varying lengths,” the authors wrote. “It also showed a robust performance across multiple regionally and temporally distinct academic and community clinical settings. This strategy improved the classification of patients with HFrEF compared with the chart diagnosis codes and identified discernible gaps in guideline-directed therapies in practice. Moreover, our language-based approach is explainable, because it detects clinically relevant terms through contextual analysis of clinical notes.”
Knowing a patient’s heart failure subtype is crucial, the group added, if clinicians want to provide the best treatments possible.
“Although physician-level inertia to intensify guideline-directed medical therapies remains a major barrier to improve the quality of care in HFrEF, systematic identification of individuals with HFrEF is the first step in assessment of care quality in this population,” they wrote.
Click here to read the full analysis in JACC: Heart Failure, an American College of Cardiology journal.
Michael has more than 18 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.