Abstract
Identifying patient characteristics that influence the rate of colorectal polyp recurrence can provide important insights into which patients are at higher risk for recurrence. We used natural language processing to extract polyp morphological characteristics from 953 polyp-presenting patients' electronic medical records. We used subsequent colonoscopy reports to examine how the time to polyp recurrence (731 patients experienced recurrence) is influenced by these characteristics as well as anthropometric features using Kaplan-Meier curves, Cox proportional hazards modeling, and random survival forest models. We found that the rate of recurrence differed significantly by polyp size, number, and location and patient smoking status. Additionally, right-sided colon polyps increased recurrence risk by 30% compared to left-sided polyps. History of tobacco use increased polyp recurrence risk by 20% compared to never-users. A random survival forest model showed an AUC of 0.65 and identified several other predictive variables, which can inform development of personalized polyp surveillance plans.
Introduction
Colorectal cancer (CRC) is a leading cause of death globally, but it is largely preventable if precancerous polyps are identified and removed during colonoscopies. Currently, follow-up (surveillance) schedules are based on broad clinical guidelines. However, patients develop recurrent polyps at significantly different rates.
A major challenge in personalizing care is that critical data—such as the specific size, number, and exact location of polyps—is often buried in "free-text" clinical notes rather than structured databases. This study aims to bridge this gap by using natural language processing (NLP) to extract these features and build predictive models that estimate the specific risk of recurrence for individual patients.
Methods
The researchers developed a multi-stage analytical pipeline:
- NLP Extraction: Using Python and the NLTK library, they built a custom pipeline to scan thousands of colonoscopy reports. The system automatically identified polyp locations, quantities, and sizes (standardizing measurements like "mm" and "cm" into numerical values).
- Data Integration: This extracted data was merged with structured electronic medical record (EMR) data, including age, BMI, and smoking history.
- Time-to-Event Analysis:
- Kaplan-Meier Curves: Used for univariate analysis to visualize how single factors affect the time to recurrence.
- Cox Proportional Hazards: A standard statistical model used to determine the Hazard Ratios (HR) of various factors.
- Random Survival Forest (RSF): A machine learning approach capable of handling non-linear relationships and complex interactions between 37 different predictors.
Datasets
The study utilized a dataset from the Dartmouth-Hitchcock Medical Center (DHMC):
- Initial Cohort: 4,273 randomly selected patients who underwent colonoscopies between 2011 and 2017.
- Final Analytical Set: 953 patient record pairs who met strict inclusion criteria (e.g., had at least two colonoscopies separated by at least 6 months).
- Recurrence Data: Within this group, 731 patients (76.7%) experienced a polyp recurrence over a 6.25-year period.
Key Results and Insights
Anatomical and Behavioral Risk Factors
Polyps located in the right colon (the proximal side) were associated with a significantly faster recurrence rate. This aligns with medical theories that right-sided lesions may be more biologically aggressive or harder to detect initially. As shown in the Kaplan-Meier analysis, patients with right-sided polyps had a lower probability of remaining recurrence-free over time compared to those with left-sided polyps.
Additionally, a history of tobacco use increased the risk of recurrence by 20%. This suggests smoking creates a "field effect" of instability in the colon lining that persists even after initial polyp removal.
Quantitative Risk Assessment
The Cox proportional hazards model quantified the impact of these factors. Right-sided polyps increased recurrence risk by approximately 30% (RR = 1.3), while tobacco use increased it by 20% (RR = 1.2). Both the maximum size of the initial polyp and the total count of polyps were also highly significant predictors of how quickly a new polyp would appear.
Machine Learning Performance
The Random Survival Forest model achieved an AUC of 0.65 at a 1,500-day prediction horizon. Interestingly, BMI was a strong predictor in the RSF model but appeared insignificant in simpler statistical models. This suggests that BMI's influence on cancer risk is likely tied to complex interactions with other variables that traditional linear models miss.
Limitations and Future Work
While promising, the study is limited by its single-center design and the potential for "missed polyps" during initial screenings to be misclassified as recurrences. Future directions include collaborating with state-level registries to create a more comprehensive dataset and eventually integrating these risk scores directly into EMR systems to provide doctors with real-time, personalized surveillance recommendations.
Summary updated