Background: Computed tomography imaging routinely detects incidental findings; most research focuses on malignant findings. However, benign diseases such as hiatal hernia also require identification and follow-up. Natural language algorithms can help identify these non-malignant findings.
Methods: Imaging of adult trauma patients from 2010 to 2020 who underwent CT chest/abdomen/pelvis was evaluated using an open-source natural language processor to query for hiatal hernias. Patients who underwent subsequent imaging, endoscopy, fluoroscopy, or operation were retrospectively reviewed.
Results: 1087(10.6%) of 10 299 patients had incidental hiatal hernias: 812 small (74.7%) and 275 moderate/large (25.3%). 224 (20.7%) had subsequent imaging or endoscopic evaluation. Compared to those with small hernias, patients with moderate/large hernias were older (66.3 ± 19.4 vs 79.6 ± 12.6 years, P < .001) and predominantly female (403[49.6%] vs 199[72.4%], P < .001). Moderate/large hernias were not more likely to grow (small vs moderate/large: 13[7.6%] vs 8[15.1%], P = .102). Patients with moderate/large hernias were more likely to have an intervention or referral (small vs moderate/large: 6[3.5%] vs 7[13.2%], P = .008). No patients underwent elective or emergent hernia repair. Three patients had surgical referral; however, only one was seen by a surgeon. One patient death was associated with a large hiatal hernia.
Conclusions: We demonstrate a novel utilization of natural language processing to identify patients with incidental hiatal hernia in a large population, and found a 10.6% incidence with only 1.2%. (13/1087) of these receiving a referral for follow-up. While most incidental hiatal hernias are small, moderate/large and symptomatic hernias have high risk of loss-to-follow-up and need referral pipelines to improve patient outcomes.
Keywords: hernia; hiatal hernia; natural language processing; quality improvement.