Abstract
Introduction
Heart failure (HF) is a common health concern and echocardiography is widely used to aid its diagnosis. Data from echocardiogram reports may become a powerful resource for HF research. We aimed to formulate and validate a natural language processing (NLP) algorithm that extracts keywords from echocardiogram reports to identify patients with impaired left ventricular systolic function (LVSF) and left ventricular hypertrophy (LVH).
Methods
Free text descriptions of LVSF were identified and processed from 150,000 echocardiograms performed in Ninewells Hospital since 1994. Code was developed to parse the free text and generate a lexicon from which to determine the reporter's impression of LVSF. Validation was performed using a selection of 1000 reports that were manually reviewed by an independent investigator and correlation between manual and code-based assessment were examined. One hundred and five individuals with impaired LVSF were identified and their case notes reviewed to test the reliability of the above HF definition and 42 scans with varying degrees of left ventricular impairment were blindly re-reported by British Society of Echocardiography (BSE) accredited echocardiographers. The algorithm was retested following an update to the echocardiography database in 2012.
Results
The algorithm used identified 19,758 individuals with impaired LVSF. Of the 1000 reports, 98% reviewed manually showed concordance with the algorithm. Of the 105 HF cases, 91% based on impaired LVSF and loop diuretic prescription were confirmed cases of heart failure on review of their case notes. The algorithm was also found to be applicable to the new NHS Tayside database as it had an overall recall of 94.2%, overall precision of 98.39%, and an overall F1 score of 96.25%.
Conclusion
This algorithm provides a robust method of identifying those with evidence of impaired LVSF on echocardiogram. This may be linked to prescribing data to accurately identify HF, allowing its use for clinical research.
Heart failure (HF) is a common health concern and echocardiography is widely used to aid its diagnosis. Data from echocardiogram reports may become a powerful resource for HF research. We aimed to formulate and validate a natural language processing (NLP) algorithm that extracts keywords from echocardiogram reports to identify patients with impaired left ventricular systolic function (LVSF) and left ventricular hypertrophy (LVH).
Methods
Free text descriptions of LVSF were identified and processed from 150,000 echocardiograms performed in Ninewells Hospital since 1994. Code was developed to parse the free text and generate a lexicon from which to determine the reporter's impression of LVSF. Validation was performed using a selection of 1000 reports that were manually reviewed by an independent investigator and correlation between manual and code-based assessment were examined. One hundred and five individuals with impaired LVSF were identified and their case notes reviewed to test the reliability of the above HF definition and 42 scans with varying degrees of left ventricular impairment were blindly re-reported by British Society of Echocardiography (BSE) accredited echocardiographers. The algorithm was retested following an update to the echocardiography database in 2012.
Results
The algorithm used identified 19,758 individuals with impaired LVSF. Of the 1000 reports, 98% reviewed manually showed concordance with the algorithm. Of the 105 HF cases, 91% based on impaired LVSF and loop diuretic prescription were confirmed cases of heart failure on review of their case notes. The algorithm was also found to be applicable to the new NHS Tayside database as it had an overall recall of 94.2%, overall precision of 98.39%, and an overall F1 score of 96.25%.
Conclusion
This algorithm provides a robust method of identifying those with evidence of impaired LVSF on echocardiogram. This may be linked to prescribing data to accurately identify HF, allowing its use for clinical research.
| Original language | English |
|---|---|
| Pages (from-to) | 51 |
| Number of pages | 1 |
| Journal | Scottish Medical Journal |
| Volume | 64 |
| Issue number | 3 |
| Early online date | 31 Jul 2019 |
| DOIs | |
| Publication status | Published - 1 Aug 2021 |