Submission ID 130133
| Issue/Objective | Cervical cancer is the fourth most common cancer among women globally, yet it is almost entirely preventable. Every two minutes, a woman dies of it, and over 90% of those deaths occur in low- and middle-income countries, where fragmented governance, chronic underfunding, and collapsed screening infrastructure trap women into late-stage diagnosis. Sub-Saharan Africa bears the sharpest end of this inequity: high HIV prevalence compounds risk, cancer registries remain underpowered, and the political will to invest in early detection competes with a fragmented global health architecture under mounting fiscal strain. Within this regional crisis, Rwanda stands as both a cautionary case and a beacon of ambition: the first African country to launch a national HPV vaccination programme, yet one where cervical cancer remains the second most prevalent cancer among women in 2023, with the majority still presenting at advanced, life-limiting stages. The quantified survival penalty of late detection, and the potential of machine learning to generate the actionable intelligence needed to reclaim equity in outcomes, remains insufficiently documented. This study addresses that evidence gap directly, from within Rwanda's own health system, using seven years of national cancer registry data. |
|---|---|
| Methodology/Approach | A retrospective cohort study analysed data from 5,529 women tested for cervical cancer in Rwanda's National Cancer Registry (2016-2023), retaining 2,476 confirmed cases with complete clinical records. Detection stage was classified using FIGO criteria: early (Stages I-IIA) versus late (Stages IIB-IV). Kaplan-Meier survival curves and log-rank tests compared survival distributions between groups. Cox proportional hazards models (C-index: 0.74) estimated mortality risk, adjusting for age, HIV status, and detection stage. Three machine learning (ML) models -Logistic Regression, Gradient Boosting for Survival Analysis (GBSA), and Survival Support Vector Machines (SSVM) were trained on 70% of data and validated on 30%, evaluated by accuracy, sensitivity, ROC-AUC, Brier score, and calibration plots. |
| Results | Early detection was associated with a 39% reduction in the hazard of death (HR=0.61; 95% CI: 0.37-0.98; p=0.046). Kaplan-Meier analysis confirmed significantly better survival in the early detection group throughout the 11-year observation window (log-rank p<0.001; χ²=17.30). HIV positivity more than doubled mortality risk (HR=2.12; 95% CI: 1.50-3.00; p<0.001) and each additional year of age at diagnosis increased risk by 3%. Gradient Boosting demonstrated superior predictive performance: accuracy 88.2%, sensitivity 99.7%, ROC-AUC 0.67, and the lowest Brier score (0.14), with calibration plots confirming best alignment between predicted and observed outcomes. |
| Discussion/Conclusion | Stage at detection, HIV status, and age at diagnosis are upstream determinants of survival inequity in cervical cancer in Rwanda. The 39% survival penalty imposed by late detection reflects structural governance failures - barriers to screening access, incomplete registry data, and the invisibility of vulnerable women to health planners. This Global South-led study demonstrates that national cancer registry data, combined with accessible ML tools, can generate the evidence base needed for bold policy action. Scaling early detection infrastructure and integrating ML-assisted triage into cancer programmes are not technical luxuries, they are equity imperatives for LMICs committed to reclaiming health justice. |
| Presenters and Affiliations | Piero Mazimpaka Irakiza Ministry of Health_National Health Intelligence Centre Mazimpaka Irakiza Piero Ministry of Health_National Health Intelligence Centre Francine Uwamahoro University of Rwanda Prof Belson Rugwizangoga University Teaching Hospital of Kigali (CHUK) Emmanuel Christian Nyabyenda University of Rwanda, Center of Excellence in Data Science Dr Eric Remera NHIC/MOH |
