Using AI to identify chest x rays requiring follow up after pneumonia

Abstract

<bold>Background</bold> BTS guidelines suggest that all high risk patients should have a chest X-ray (CXR) repeated 6 weeks after community acquired pneumonia to exclude underlying pathologies. <bold>Aims</bold> To assess the use of artificial intelligence natural language processing (NLP) software to review CXR reports to ensure patients do not miss follow up. <bold>Method</bold> <fig><object-id>erj;66/suppl_69/PA4979/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig> <bold>Results</bold> Iteration 1: 126 reports were reviewed. Sensitivity and specificity for NLP to detect consolidation or pneumonia was 61% and 57% respectively. Sensitivity and specificity for the string search to detect follow up advice was 91% and 88% respectively. Iteration 2: 332 reports were reviewed. The sensitivity and specificity for NLP detecting abnormality was 54% and 38% respectively. The string search picked up follow up advice in 60% of reports. The sensitivity and specificity for string search detecting follow up advice was 96% and 89% respectively. <bold>Conclusion:</bold> NLP as a method of analysing CXR reports is limited, possibly due to the varied descriptive terms used. String search to identify follow up is better, although relies on reports including follow up recommendations, which will be variable between centres and radiologists.

Ähnliche Arbeiten