Predicting the uptake of free cancer screening services in rural India using machine learning tools.

Abstract

e22523 Background: 1,413,316 cancer cases and 916,827 cancer deaths were reported in India in 2022 (Globocan, 2022); the incidence is estimated to increase by 12.8% in 2025 and 57.5% by 2040 (Sathishkumar K, 2023). Poor uptake of cancer screening of less than 5% and late stage at diagnosis continue to be a cause for concern (NFHS – 5, 2022). “Sarva Rakshana”, a longitudinal, population-based, cancer prevention programme, was launched in a rural subdistrict of India in July 2022. Through this programme, more than 25% of the eligible adult population were screened for free for oral, breast and cervical cancers and received cancer prevention & risk reduction counselling. Predicting the likelihood of uptake and non-uptake of cancer screening programmes in rural India can guide healthcare workers to augment their efforts to improve acceptance among those less likely to attend the programmes. Methods: IEC approval was obtained for the programme. Machine learning tools were tested to predict the uptake of screening services. Participant data was collected by trained health workers using validated, interviewer administered questionnaire and uptake of screening services by participants were logged by referring to the screening database. Correlation between sociodemographic factors and personal history with screening uptake was performed using Fisher’s Exact Test. Factors showing significant correlation were incorporated to machine learning tools to perform predictive analysis using 3 combinations of training and test splits of 70:30, 80:20 and 90:10. Results: 57270 individuals between 25-80 years across 256 villages were invited to free cancer screening camps held in their respective villages between January to December 2023. 25.84% (14801) individuals accepted the screening invitation and attended the screening camps. 9 variables showed significant correlation with screening uptake. Results from predictive analysis using different models is shown in Table 1. All models had an accuracy of >70% and Brier Score of less than 0.2 except Neural Network. Conclusions: Machine learning tools can be used to predict uptake of cancer screening services with fairly good accuracy. Future implications: Integrating these algorithms in public health programmes may guide specialised efforts to increase uptake. Results from predictive analysis using different models. Model Splits Accuracy AUC Brier Score Logistic Regression 70:3080:2090:10 75.1075.0675.10 0.770.770.77 0.1590.1580.157 Decision Tree 70:3080:2090:10 75.5575.7176.06 0.710.710.71 0.1670.1660.164 eXtreme Gradient Boosting 70:3080:2090:10 76.0776.1775.9 0.770.780.78 0.1580.1570.157 Random Forest 70:3080:2090:10 75.7475.7575.68 0.770.770.77 0.1590.1590.158 Support Vector Machine 70:3080:2090:10 76.1176.1876.04 0.670.650.62 0.1760.1760.176 Neural Network 70:3080:2090:10 70.0173.2370.7 0.770.800.79 0.2130.1950.210