Background
Despite being a preventable disease, cervical cancer continues to be a public health concern, affecting mainly lower and middle-income countries. Therefore, in Bogotá a home-visit based program was instituted to increase screening uptake. However, around 40% of the visited women fail to attend their Pap smear test appointments. Using this program as a case study, this paper presents a methodology that combines machine learning methods, using routinely collected administrative data, with Champion’s Health Belief Model to assess women’s beliefs about cervical cancer screening. The aim is to improve the cost-effectiveness of behavioural interventions aiming to increase attendance for screening. The results presented here relate specifically to the case study, but the methodology is generic and can be applied in all low-income settings.
Methods
This is a cross-sectional study using two different datasets from the same population and a sequential modelling approach. To assess beliefs, we used a 37-item questionnaire to measure the constructs of the CHBM towards cervical cancer screening. Data were collected through a face-to-face survey (N = 1699). We examined instrument reliability using Cronbach’s coefficient and performed a principal component analysis to assess construct validity. Then, Kruskal–Wallis and Dunn tests were conducted to analyse differences on the HBM scores, among patients with different poverty levels. Next, we used data retrieved from administrative health records (N = 23,370) to fit a LASSO regression model to predict individual no-show probabilities. Finally, we used the results of the CHBM in the LASSO model to improve its accuracy.
Results
Nine components were identified accounting for 57.7% of the variability of our data. Lower income patients were found to have a lower Health motivation score (p-value < 0.001), a higher Severity score (p-value < 0.001) and a higher Barriers score (p-value < 0.001). Additionally, patients between 25 and 30 years old and with higher poverty levels are less likely to attend their appointments (O.R 0.93 (CI: 0.83–0.98) and 0.74 (CI: 0.66–0.85), respectively). We also found a relationship between the CHBM scores and the patient attendance probability. Average AUROC score for our prediction model is 0.9.
Conclusion
In the case of Bogotá, our results highlight the need to develop education campaigns to address misconceptions about the disease mortality and treatment (aiming at decreasing perceived severity), particularly among younger patients living in extreme poverty. Additionally, it is important to conduct an economic evaluation of screening options to strengthen the cervical cancer screening program (to reduce perceived barriers). More widely, our prediction approach has the potential to improve the cost-effectiveness of behavioural interventions to increase attendance for screening in developing countries where funding is limited.