Family history is one of the strongest known risk factors for pancreatic cancer. A person with one first-degree relative (parent, sibling, or child) with pancreatic cancer has a 7- to 9-fold increased risk of developing the disease themselves. With three or more first-degree relatives affected, the risk increases 17- to 32-fold.
Despite this, family history of pancreatic cancer is rarely systematically identified or tracked. Doctors record family history information in free-text clinical notes — written in natural language, often abbreviated, and not easily searchable by computers. Mining this information at scale requires natural language processing (NLP) tools that can read and understand unstructured text.
This study built on prior work identifying pancreatic cysts from clinical notes, extending the approach to identify patients with a family history of pancreatic cancer. Crucially, it tested whether an algorithm developed at one hospital (Indiana University) could be transferred to a completely different institution (Mayo Clinic) without retraining — a property called portability.
The NLP system was built in the UIMA (Unstructured Information Management Architecture) framework, a modular pipeline where each component handles one step of the analysis. The pipeline included: (1) a report separator that identifies the boundaries of each clinical note in a batch of records; (2) a metadata annotator that tags patient ID, date, and note type; (3) a section header detector that identifies the family history section; and (4) a family member and diagnosis identifier that finds references to specific relatives and cancer diagnoses.
After identifying relevant sentences, the system performed relation discovery — linking each family member term (father, mother, brother, sister, cousin) to the specific cancer diagnosis mentioned in the same sentence. For complex sentences with multiple family members, a rule-based approach divided the sentence into sub-sentences based on punctuation and conjunctions to assign the correct relative to the correct diagnosis.
A negation detection algorithm called DEEPEN was applied to determine whether the cancer was affirmed or negated (for example, 'no family history of pancreatic cancer' versus 'mother had pancreatic cancer'). DEEPEN uses dependency parsing — analyzing grammatical relationships — rather than simple proximity of negation words to concepts, making it more accurate on complex sentences.
At Indiana University (IU), the system achieved a family history identification precision of 88.9% on the test set. When the same algorithm — without any modification — was applied to the Mayo Clinic dataset, the precision was 87.8%. This consistent performance across two different institutions with different clinical note formats demonstrates that the rule-based approach is genuinely portable.
After customizing the algorithm to the Mayo Clinic dataset (adding Mayo-specific section headers and terminology), the precision improved further to 88.1%. This suggests that a modest amount of customization can bridge institutional differences without requiring a full retraining from scratch.
In the relation discovery task — correctly linking family members to their diagnoses — the system achieved precision of 75.3%, recall of 91.6%, and F-measure of 82.6%. Negation detection precision was exceptionally high at 99.1%, meaning the system almost never confused a denied cancer history for a confirmed one. Among the 3,573 Mayo Clinic patients with pancreatic cancer, 268 (7.5%) had at least one family member with pancreatic cancer identified in their notes.
The portability of the algorithm is a critical finding. Machine learning approaches to NLP typically require large amounts of labeled training data from each new institution to perform well — an expensive and time-consuming requirement. Rule-based systems, while more laborious to initially build, transfer more readily because their logic is explicit and can be manually inspected and corrected.
The most common errors occurred in sentences with complex family relationships — for example, 'His mother's sister's granddaughter was diagnosed with pancreatic cancer' — where the system incorrectly assigned the cancer to each family member term separately rather than identifying the correct relationship chain. Handling these multi-step relations requires semantic inference that goes beyond pattern matching.
Another challenge was co-referencing — when a patient is referred to by a pronoun later in a note ('His son... the one who is living...') rather than a direct family member term. The current system does not handle co-referencing, which accounts for a subset of missed cases.
This study showed that a rule-based NLP system for identifying family history of pancreatic cancer can be built once and deployed at multiple institutions with only minor customization, achieving consistent precision around 88%. This portability makes large-scale deployment feasible without the institutional overhead that machine learning approaches typically require.
The practical implication is significant: a hospital network could systematically screen all its patients' clinical notes to identify those with a family history of pancreatic cancer, then offer these patients targeted surveillance — imaging, genetic counseling, and referral to specialized pancreatic cancer prevention programs.
Future work will focus on developing a risk stratification model based on family history findings — distinguishing between patients with one first-degree relative versus three or more — to guide surveillance intensity. Integration with the pancreatic cyst identification system described in companion work would create a comprehensive automated platform for pancreatic cancer early detection.