Identifying signs and symptoms of pancreatic cancer: a population-based study using electronic health records and natural language processing.

OBJECTIVES

Patient-reported symptoms are often the trigger for diagnosis of pancreatic cancer (PC). We aimed to report the prevalence and lead time of symptoms of PC from electronic health records in a population-based sample.

METHODS

In this nested case-control study, 2611 PDAC cases and 10,085 matched controls 50-84 years of age were identified. A total of 17 symptoms identified by literature review and other approaches were extracted from clinical notes and emails/messages between patients and care providers using natural language processing.

RESULTS

The most common symptoms in PDAC cases were abdominal pain (85.5 %), weight loss (71.4 %), fatigue (66.5 %), nausea/vomiting (63.2 %), anorexia (54.2 %), and back pain (54.2 %). After adjusting for patient characteristics, all gastrointestinal (GI) symptoms, all systemic symptoms, and mental/sleep disorders were significantly associated with PC (p < 0.05 for depression and p < 0.01 for all other symptoms). Back pain and fatigue had a median onset time of 11.5 and 7.1 months, respectively. All GI and other systemic symptoms had a late onset, with a median onset time of 4.1 months prior to cancer diagnosis. Anxiety, depression, and insomnia started 12.0, 16.5, and 8.2 months (median) prior to cancer diagnosis, earlier than GI or other systemic symptoms; however, their associations with PDAC appeared weaker compared to those of GI/systemic symptoms.

CONCLUSIONS

Our approach to capture symptoms from EHR has the potential to enhance the systematic identification of patients with symptoms for earlier detection.