Abstract
Rule-based natural language processing (NLP) tools can identify pulmonary embolism (PE) via radiology reports. However, their external validity remains uncertain.In this cross-sectional study, 1,712 hospitalized patients (with and without PE) at Mass General Brigham (MGB) hospitals (2016-2021) were analyzed. Two previously published NLP algorithms were applied to radiology reports to identify PE. Chart review by two physicians was the reference standard. We tested three approaches: (A) NLP applied to all patients; (B) NLP limited to radiology reports of patients with principal or secondary International Classification of Diseases 10th revision (ICD-10) PE discharge codes; and (C) NLP applied to patients with PE discharge codes or a Present-on-Admission (POA) indicator ("Y") for PE. All others were assumed PE-negative in Approaches B and C to minimize NLP false positives. Weighted estimates were derived from the MGB hospitalized cohort (n = 381,642) to calculate F1 scores (as the harmonic mean of sensitivity and positive predictive value [PPV]).In Approach A, both NLP tools showed high sensitivity (82.5%, 93.0%) and specificity (98.9%, 98.7%) but low PPV (60.3%, 59.6%). Approach B improved PPV (95.2%, 94.9%) but reduced sensitivity (74.1%, 76.2%), while Approach C preserved both high sensitivity (82.5%, 93.0%) and PPV (95.6%, 95.8%). Approach C demonstrated the best performance, yielding significantly higher F1 scores for both NLP tools (88.6%, 94.4%) compared with Approach A (69.7%, 72.6%) and Approach B (83.3%, 84.5%) (P < 0.001).The accuracy of PE detection improves when rule-based NLP algorithms are operationalized using administrative claims data in addition to radiology reports.