Of 11,622 patients with a completed fecal test, 1,723 (14.8%) were abnormal, and 699 (40.6%) of those had a subsequent completed colonoscopy in their EHR record within 12 months (Fig. 1). However, only 597 (34.6%) of those patients had record of a completed a colonoscopy within 6 months of their abnormal FIT test. One small clinic system was excluded due to low numbers of patients with abnormal FIT results (n = 13). We also only included patients without non-missing data for all predictors (n = 1,596). Of patients included in the final model, 34.8% (n = 556) had recorded completed colonoscopies within 6 months.
Table 1 illustrates all baseline characteristics for the entire cohort and the subgroup that had a recorded completed colonoscopy within 6 months. Overall, patients were typically white (83.3%), aged 50–64 (81.5%) and had a low rate of preventive screenings: flu shots (14.3%); prior CRC screening (38.3%)). Only eight variables were retained for the final model as they contributed to the explained variation in risk.
Table 2 shows the eight characteristics that were retained for the final Cox regression model, and the hazard ratios, confidence intervals, and number of risk points assigned to each characteristic. No notable differences were determined when the model was run for men and women separately, so therefore we combined men and women to develop one model. Our final model included: age, race, insurance, GINI income inequality, long term anticoagulant use, receipt of a flu vaccine in the past year, frequency of missed clinic appointments, and clinic site. The hazard ratios and risk score points for the final prediction model indicated that health center, age, long term anti-coagulant use, and receipt of a flu vaccine in the past year were the variables with highest points assigned in the model. Race, insurance, GINI income inequality, and number of missed appointments were also predictive of likelihood of completing a colonoscopy.
The mean predicted risk of completion of colonoscopy was 34.8%, and the model was able to accurately predict the patients who were least likely to receive a follow-up colonoscopy (lowest two quintiles, 15.9% and 28.5% respectively). Likelihood of obtaining a follow-up colonoscopy within 6 months varied across quintiles: patients with the highest predicted risk of non-adherence (bottom quintile) had an estimated 16% chance of obtaining a colonoscopy; whereas, patients with the lowest predicted risk of non-adherence (top quintile) had a greater than 55% chance of obtaining a follow-up colonoscopy. Figure 2 shows the predicted probability of colonoscopy completion.
Risk score points can be assigned to a patient to determine their risk of completing a colonoscopy. For example, we can score a patient who is on Medicaid (15 points), white (34 points), 54 years old (83 points), receives his care at health center 3 (100 points), has not missed appointments (31 points), has received a flu shot (40 points), isn’t on anticoagulants (54 points) and lives in an area with low income inequality (21 points). His total point count is 378, which predicts that he has an 81% probability of completing a colonoscopy, compared to the 35% likelihood of the average patient (data not shown).
The model showed modest separation of patients across risk levels for non-adherence to follow-up colonoscopy (C-statistic > 0.66, bootstrap-corrected C-statistic > 0.63) and excellent calibration or high agreement between observed and predicted risk. The R2 statistic, derived from the D-statistic, showed only 14% of the variation in outcome was explained in this model (R2 (95% CI)=14.03 (10.17–18.18), D (95% CI) = 0.83 (0.69–0.96)). A logistic regression, predicting the completion of a colonoscopy, showed similar results for non-adherence to follow-up colonoscopy (C-statistic = 0.66, bootstrap-corrected C-statistic > 0.64).