Development of a Bayesian Hierarchical Model for Medical Resource-Limited Settings: Prediction of Treatment Efficacy in Breast Cancer Patients in Kenya

TL;DR

This study develops a Bayesian model to predict breast cancer treatment outcomes in Kenya, accounting for both patient biology and healthcare variability.

Contribution

An enhanced Bayesian hierarchical model is introduced to capture patient-tumor-institution interactions in resource-limited settings.

Findings

The model predicted the highest pCR in HR-/HER2+ patients and lowest in HR+/HER2− patients.

Higher clinical stage was associated with larger odds ratios for the modeled outcome in HR- patients.

The model demonstrated 97.0% composite robustness with significant center-level variation in adjusted pCR.

Abstract

Background: Breast cancer remains a leading public health challenge in Kenya, where treatment outcomes are shaped by the complex interplay between institutional variability and patient-specific biological factors. Studies have demonstrated that Bayesian hierarchical models can effectively capture such multi-level interactions, improving prediction accuracy and clinical applicability. Building on this framework, an enhanced Bayesian hierarchical model incorporating biologically plausible interaction offers additional insight into how patient biology and healthcare delivery context jointly influence outcomes. In resource-limited settings, such a model is essential for guiding targeted treatment allocation, reducing outcome disparities, and optimizing scarce healthcare resources.

Methods: We conducted a retrospective cohort analysis of 284 breast cancer patients treated across 12 major cancer treatment centers in Kenya between 2018 and 2022. The primary outcome was pathological complete response to neoadjuvant chemotherapy (NAC). Three progressive Bayesian hierarchical models were developed, culminating in an enhanced model (M₃) that incorporated institutional variation, key clinical predictors, and biologically plausible interaction effects to capture complex stage-subtype relationships. Posterior inference was performed using Markov Chain Monte Carlo simulation, with comprehensive validation across multiple scenarios. A total of 48,000 simulations were completed.

Results: Higher clinical stage was associated with larger odds ratios (ORs) for the modeled outcome in the hormone receptors (HR)- reference group (Stage III vs I, OR 9.81), and HR positivity had a smaller OR (0.37), with interaction terms trending against HR positivity at higher stage but with intervals including the null. Predicted pathological complete response (pCR) remained highest in HR negative (HR-)/human epidermal growth factor receptors 2 (HER2+) at 71.8% and lowest in HR+/HER2− (18.7%). The model’s composite robustness was 97.0%, with center-level variation in adjusted pCR (31.0%-45.1%). Stage IV was reported descriptively and excluded from inference.

Conclusion: The enhanced Bayesian hierarchical model (M₃) integrates complex patient-tumor interactions with institutional variation, delivering well-calibrated, robust, and clinically interpretable predictions of pCR. In resource-limited settings, such models can guide targeted treatment allocation, reduce outcome disparities, and optimize the use of scarce oncology resources.

Figures