Application of dynamic prediction models for longitudinal biomarkers and clinical outcomes in low and middle-income settings

Honwana, Frissiano Ernest

Application of dynamic prediction models for longitudinal biomarkers and clinical outcomes in low and middle-income settings

Thesis / Dissertation

2025

Publisher

University of Cape Town

Department

Department of Public Health and Family Medicine

Faculty

Faculty of Health Sciences

Abstract

Routine monitoring of individuals with chronic diseases offers valuable data for understanding disease progression and treatment effectiveness, often using biomarkers. With the modernisation of clinical care, prediction models have received greater attention in analysing such data. Prognosis prediction modelling approaches have been widely adopted, especially with digitising health records into electronic health records (EHRs). Dynamic prediction modelling has emerged as a critical approach, allowing real-time updates of prognosis predictions based on available data. However, there is a notable scarcity of dynamic prediction models applied to routine data from EHRs, particularly in contexts such as HIV and type 2 diabetes (T2DM) in resource-limited settings. Existing dynamic prediction models are typically developed and validated in data with comprehensive follow-ups and covariate collection, leading to the assumption of their universally improved predictive performance over traditional approaches such as the Cox proportional hazards-based prediction model. In addition to applying an extension of existing models to correctly model semicontinuous biomarker data (two-part joint model), this thesis challenges this assumption by applying dynamic prediction models using large routine data from EHRs generated in resource-limited settings, specifically focusing on using longitudinal biomarkers to predict probabilities of clinical outcomes in individuals with HIV or T2DM in South Africa. The predictive performance of this model is compared with that of the Cox proportional hazards-based prediction model and the two-part joint model. The prediction models had comparable predictive performances. The Cox proportional hazards-based prediction model had area under the curve (AUC) values ranging from 0.63 to 0.89 and Brier scores between 0.042 and 0.088 across routine T2DM and HIV data. The joint model had AUCs ranging between 0.66 and 0.73 and Brier scores between 0.033 and 0.089. The two-part joint model had AUCs and Brier scores closer to 0.6 and 0.1, respectively. These findings highlight the importance of adopting a conceptual approach to inform predictive performance, emphasising the need to account for context, type of disease, characteristics of a biomarker, and data characteristics. Such an approach will enhance individualised predictions using dynamic prediction models, potentially enabling recommendations for differentiated care and improving routine monitoring for individuals with chronic diseases, especially in resource-limited settings.

Keywords

T2DM

HIV data

Reference:

Collections

PhD / Doctoral

Full item page