Clustering of longitudinal viral loads in the Western Cape

The abstract and a few summary figures of the MPH mini-dissertation submitted by Eke Arua for his degree.

Abstract

Introduction: Data from routine viral load monitoring is important for assessing the programmatic effectiveness of antiretroviral treatment (ART) in South Africa, but this is usually analysed in a cross-sectional manner and there are few analyses of longitudinal viral load trajectories over time. Cluster analysis was used to identify latent viral load trajectory patterns in a large dataset of routinely collected viral load measurements in the Western Cape, South Africa.

Methods: We analysed available VL measurements collected during routine care from the Western Cape public sector antiretroviral treatment programs, including all people living with HIV who were enrolled on ART and had experienced at least one elevated (>1000 copies/mL) viral load test result between 2008 and 2018. Empirical rules-based classification, nonparametric clustering using the KML-Shape algorithm and model-based clustering using latent class mixture modelling were used to cluster viral load trajectories.

Results: Both the nonparametric and model-based clustering techniques identified five latent viral load trajectory subgroups. The shapes and magnitudes of these subgroups differed according to method. Majority of individuals’ trajectories belonged to clusters that had a decreasing VL trend. Most of the trajectory subgroups identified had prolonged periods of low-level viremia. Both methods identified viral load trajectory clusters that increased over time.

Conclusion: Cluster analysis is a useful tool for identifying latent VL trajectory clusters in the Western Cape and the dynamic VL cluster patterns that exist among the large VL dataset can offer important insights. Further research is needed to understand factors associated with belonging to these clusters to improve population viral load suppression rates and aid HIV prevention.

Figure 1: Sample of 500 individual trajectories with LOESS smoother for different patterns over time: A) single elevated VL not at first visit; B) multiple elevated VL; C) single elevated VL at first visit and D) elevated VL throughout

Figure 2: The non-parametric approach “KLM-shape” identified five VL trajectories, mean trajectory plotted below with the percentage of individuals classified in that type indicated in top right corner.

Figure 3: The trajectories identified by the model based LCMM approach using fractional polynomials.

There are many more tables and details in the mini-dissertation itself, but it was a large and difficult analysis so well done to Eke!