A metanalysis published in the Heart journal identified 37 studies developing and applying NLP in various areas of cardiology.

This systematic review found that the majority of NLP work in cardiology has been concentrated in a small number of clinical domains and NLP methods, with the most common applications being used for disease identification and classification purposes. Patient sample sizes ranged from 60 to 621,000. Data sources ranged from single centers to national healthcare system databases and registries, most of them taking place in the US.

NLP unlocking insights

Most of the analyzed studies focused on a particular specialty within cardiology:

• Electrophysiology: These studies have been used to identify patients with atrial fibrillation, characterize those most likely to receive thromboembolic prophylaxis, develop algorithms to evaluate the significance of atrial fibrillation alerts, extract family history information, and predict cardiac resynchronization therapy outcomes. Findings from these studies have reported high accuracies and F-scores, with machine learning models incorporating unstructured data having sensitivities of 91-95% and specificities of 90-98%.

• ‍Heart Failure (HF): Many studies found NLP to be successful in accurately identifying and classifying HF, with sensitivities of 60-100%, specificities of 96-99%, PPVs of 71-96%, NPVs of 87-100%, F-scores of 74-94%, and accuracies of 77-100%. Additionally, studies have demonstrated success in predicting hospital readmissions and mortality, with deep learning models outperforming other machine learning models. NLP has been used to characterize HF patients for quality care metrics and for clinical trials eligibility, as well as to identify patients with ineffective self-management, evaluate medication adherence, and identify HF symptoms.‍

• Coronary Artery Disease (CAD): Most of the studies conducted resulted in sensitivities, specificities, and positive and negative predictive values of over 80%, while two studies also reported an AUC of 72% for predicting major adverse cardiovascular events and in patient admissions following cardiac catheterization.

Challenges and opportunities

In the authors’ opinion, most of the studies are showing good data quality and predictive power. However, generalizability remains one of the biggest challenges to solve. At LynxCare we have seen how different clinics and hospitals rely on disparate EHR and data systems, utilize different vocabularies, expressions and languages, making it difficult for a single NLP pipeline to be suitable for any given healthcare provider without significant time investment from clinical experts and data analysts to adapt it to the local context. Thanks to subject matter expertise, and experience on the market, LynxCare is filling this gap as a fundamental step by developing clinically relevant data dictionaries to improve the quality of concept mapping and making it as broad as possible without sacrificing concept granularity.

Explainability and usability of NLP and other AI technologies, particularly predictive modelling and classification systems (i.e., risk calculators) also remain important challenges that could be at least partially solved by clearer communication and tighter collaboration between stakeholders, from clinicians to hospital managers to technology providers.

NLP has the potential to unlock an incredible wealth of information from unstructured notes in EHRs and other information systems, which can ultimately help clinicians and hospitals during routine care.

Source: https://heart.bmj.com/content/108/12/909