Pancreatic cysts are fluid-filled lesions in the pancreas that are increasingly detected as a consequence of the widespread use of cross-sectional abdominal imaging for other clinical indications. The majority are benign, but a subset, particularly intraductal papillary mucinous neoplasms (IPMNs) and mucinous cystic neoplasms, carry a real risk of malignant transformation into pancreatic cancer.
Clinical guidelines recommend surveillance imaging for patients with pancreatic cysts meeting certain criteria, as serial monitoring can detect malignant change at an early and potentially curable stage. However, this surveillance is only effective if patients with cysts are systematically identified in the first place.
In large healthcare systems, the reports that describe pancreatic cysts are written in free-text narrative format by radiologists and pathologists. These reports are rarely coded or structured in ways that allow automated retrieval, meaning that many patients with cysts go unrecognized and do not enter surveillance programs.
This study developed and validated an automated natural language processing (NLP) system to scan large volumes of clinical narrative reports and accurately identify all patients with documented pancreatic cysts, enabling systematic surveillance to be initiated for those at risk.
The NLP system was built using the UIMA (Unstructured Information Management Architecture) framework, an open-source platform developed by IBM and widely used for clinical text processing. UIMA allows modular construction of text analysis pipelines from individual processing components that can be combined and customized.
A key challenge in clinical NLP is negation detection: medical reports frequently use phrases like "no evidence of cyst" or "cyst was not identified," which contain the word cyst but describe its absence. The system incorporated the DEEPEN negation algorithm, which identifies linguistic negation patterns and excludes negated mentions from the positive cyst classification.
The system was trained and validated on radiology and pathology reports from a large academic medical center. The corpus included reports from CT scans, MRI studies, endoscopic ultrasound, and surgical pathology, covering the full range of imaging and tissue-based modalities through which pancreatic cysts are detected.
Manual review by clinical experts was used to create a gold standard annotation set against which the NLP system's performance was measured. Annotators labeled each report as positive or negative for pancreatic cyst, providing the reference labels needed to calculate sensitivity, specificity, and other accuracy metrics.
The system was applied to a corpus of 566,233 clinical reports from 50,669 patients over a seven-month period. This scale represents routine clinical report volumes at a large academic medical center and demonstrates that the system can operate efficiently in a real-world healthcare informatics environment.
Against the manually annotated gold standard, the NLP system achieved a sensitivity of 99.9% and a specificity of 98.8%. These performance levels are exceptional for clinical NLP and represent meaningful improvements over manual chart review, which is prohibitively time-consuming at this scale.
The near-perfect sensitivity means that essentially no patient with a documented pancreatic cyst in a clinical report would be missed by the automated system. This is the critical requirement for a surveillance identification tool: missing cases is far more clinically dangerous than flagging an occasional false positive for manual review.
The high specificity of 98.8% means that the system generates very few false positives, minimizing the administrative burden of reviewing incorrectly flagged patients. This balance between sensitivity and specificity makes the system highly suitable for practical clinical deployment at institutional scale.
Across the 50,669 patients whose records were analyzed, the NLP system identified a substantial number with documented pancreatic cysts. This prevalence finding itself has clinical significance, as it illustrates that pancreatic cysts are common in the patient population undergoing abdominal imaging at a large medical center.
Many of the identified cysts had been documented in multiple reports over time, reflecting longitudinal imaging follow-up. The system was able to link mentions across reports for the same patient, supporting construction of a cyst surveillance registry with chronological documentation of each patient's cyst history.
A proportion of patients with cysts in the identified cohort had no prior enrollment in a formal cyst surveillance program, representing a population of patients who were at potential risk but had not been systematically flagged for follow-up. Identification of this group is a primary clinical benefit of automated NLP-based case finding.
The system also captured mentions of cyst type when specified in reports, including IPMNs and mucinous cysts, helping prioritize which patients require closer surveillance based on the known malignant potential of different cyst subtypes.
The practical application of the NLP system described in this study is the construction of a pancreatic cyst registry: a systematically maintained database of all patients with documented cysts, their imaging history, and their risk profile. Such a registry is the foundation for an organized surveillance program.
Without automated NLP, building such a registry would require manual review of hundreds of thousands of reports, an impractical undertaking for any clinical team. The NLP approach enables this registry to be created retrospectively for existing records and maintained prospectively as new reports are generated.
Registry-based surveillance ensures that no patient falls through the cracks between follow-up imaging appointments. Patients can be automatically scheduled for repeat imaging, and those whose cysts meet criteria for concern can be flagged for expedited endoscopic or surgical evaluation.
The study's institutional implementation showed that this approach is feasible in a large academic health system with complex, multi-format clinical records. The workflow is scalable to other institutions and compatible with different electronic health record platforms, provided the underlying NLP system is adapted for local terminology and report structures.
This study demonstrates that clinical NLP can serve as critical infrastructure for cancer prevention programs, enabling systematic identification of at-risk populations that cannot be captured through structured coding or database queries alone. Pancreatic cyst surveillance is one compelling use case, but the approach generalizes to many other cancer risk conditions documented in free text.
The UIMA and DEEPEN combination proved highly effective for the specific challenges of medical text: complex sentence structures, abundant negations, and domain-specific terminology. Future work can extend the system to extract additional structured information from cyst reports, including cyst size, morphology, and radiologist management recommendations.
As electronic health records accumulate decades of unstructured clinical narrative, NLP systems capable of extracting actionable clinical signals from this information represent a significant opportunity to improve preventive care. The pancreatic cyst case illustrates how automating surveillance eligibility identification can translate computational capabilities directly into patient benefit.
Widespread adoption of similar NLP-driven surveillance systems across healthcare networks could meaningfully increase the proportion of pancreatic cancer cases caught at an early, treatable stage, contributing to the broader goal of reducing mortality from this disease.