Leveraging natural language processing artificial intelligence for automated data extraction of ejection fraction and strain from cardiovascular imaging reports.

Study objective: To determine the feasibility of using natural language processing (NLP) to extract ejection fraction (EF) and related cardiac imaging parameters for cancer survivors, across multiple imaging modalities, over a 20-year period. Design: Retrospective observational study applying NLP to multi-modality cardiovascular imaging reports. Setting: Large academic medical center with electronic health records spanning 2000-2020. Participants: More than 4000 patients with multi-modality cardiac imaging reports, including echocardiography, nuclear cardiology, and cardiac magnetic resonance (CMR). Interventions: Application of NLP algorithms to extract EF and related measurements from free-text cardiac imaging reports. Manual review was performed to assess and resolve discrepancies. Main outcome measures: Applicability of NLP for extracting EF across echocardiography, nuclear cardiology, and CMR reports. Results: NLP successfully extracted EF values across all imaging modalities. The method demonstrated comparable performance in echocardiography, nuclear cardiology, and CMR reports. Discrepancies identified through manual review highlighted the importance of algorithm training to accommodate modality-specific terminology and improve sensitivity. Conclusion: NLP offers a scalable approach for extracting EF and related parameters from large volumes of unstructured imaging reports across multiple modalities. While NLP is modality-agnostic, algorithm training and manual review are essential. This approach can facilitate longitudinal analyses of cardiac function in cancer survivors and other patient populations.