AUTOMATED IDENTIFICATION OF PANCREAS CYSTS FROM RADIOLOGICAL REPORTS YIELDS ACTIONABLE INSIGHTS

Background
Pancreas mucinous cystic lesions are known precursors of pancreas cancer. Traditionally, these cysts have been identified incidentally during radiological studies and then referred to specialists for evaluation. We automated detection via a natural language processing (NLP) based computational tool we developed that can classify radiological reports for evidence of pancreatic lesions. This study implemented this NLP approach to patients with pancreas cystic lesions at scale.

Methods
The NLP model was deployed on all abdominal CT and MRI radiological reports from a large integrated health system during January and July 2023. The following variables were manually extracted from classified reports: cyst size, cyst stability (when reported), pancreatic duct dilatation, and presence of Fukuoka criteria/high-risk features (mural nodules, enhancing cyst wall, downstream pancreatic atrophy). These features were also utilized to categorize the cysts into high and low risk cysts requiring clinical follow up. A Chi-square test was used to compare the two cyst cohorts. The patient charts were reviewed to determine if appropriate specialist follow-up was scheduled, and rates of follow-up were calculated.

Results
999 reports from January 2023 and 1,384 reports from July 2023 containing suspicious lesions classified by the model were reviewed. Known cancers, lesions described as masses, and non-cystic lesions were excluded to yield 412 and 650 reports, respectively. The composition of the two cohorts did not differ significantly based on cyst size, cyst history, or classification into low or high-risk cysts. 13.6% (56) of cysts demonstrated either "Worrisome” or “High-Risk” features according to Fukuoka guidelines, of which 6.8% (28) were at least 3 cm. Within the January cohort, 63.1% (260) cysts were determined to require follow-up based on high-risk criteria (GI appointment with imaging within 1 year or advanced GI appointment). 23.4% (60) patients underwent a GI follow-up appointment within our health system.

Discussion/Conclusion
The identification of preneoplastic pancreatic cystic lesions at scale from radiological reports represents a previously unrecognized opportunity to improve patient outcomes. Here we show the utility of an NLP based algorithm to classify radiological reports from a large integrated healthcare system and identify hundreds of suspicious cystic lesions from a month of data. We find that 63.1% of these lesions required gastroenterology follow-up based on current pancreatic cyst management guidelines, but that only 23.4% received a GI follow-up appointment in our system. We are now prospectively navigating the identified patients to clinics within the healthcare system to improve follow up metrics with an eventual goal of automated referral of patients with cystic lesions to appropriate follow-up.