Monitoring HIV disease progression among antiretroviral therapy patients in Zimbabwe from 2004 to 2017: spatial and multistate model approaches

Abstract

Globally, remarkable progress has been made on the treatment of HIV-infected people using antiretroviral therapy (ART), particularly in sub-Saharan Africa (SSA). SSA region is disproportionately affected by the pandemic as it accounts for more than 50% of people living with HIV (PLWHIV). With the global efforts to achieve zero HIV incidences by 2030, the 90- 90-90 UNAIDS targets were launched, whereby 90% of the PLWHIV should know their status, 90% of those diagnosed should be initiated on ART, and 90% of those on ART should achieve viral suppression(1). In Zimbabwe, strides have been made towards achieving these targets; however, limitations exist. Statistical modelling approaches like multistate models framework and geostatistical models can be used to determine how well ART patients are adhering to ART by estimating the ART attrition rate due to loss to follow-up (LTFU), immune deterioration patterns using CD4 cell counts and viral rebound trajectories. Geo-spatial modelling techniques can also be used to determine the spatial patterns of ART drop-out at the health facility and district levels and identify hot spots for ART drop-out to guide policy. Study objectives The main focus of this study was to answer the following questions crucial in HIV disease progression and monitoring: 1) what are the LTFU patterns after ART decentralization?; 2) what are the immune-deterioration patterns among ART patients?; 3) what are the underlying factors associated with poor viral suppression rates?; 4) what are the viral suppression among ART patients?; 5) which regions are heavily affected by poor ART outcomes (drop-outs, viral rebound or immune–deterioration) in Zimbabwe?; 6) and which statistical tools can be most appropriate iv to answer these epidemiological questions using routinely collected individual-level data and partially observed aggregated data? The specific objectives of this study were: 1. To demonstrate the use of Kolmogorov’s forward equations in estimating transition probabilities if data is partially observed using Bayesian estimation approach as an application in the multistate models' review. 2. To describe the spatio-temporal patterns of ART drop-out at the health facility and district levels and identify cold spots and hot spots for ART drop-out 3. To determine the time to occurrence of LTFU with mortality as a competing event over 13 years follow-up period across health facility levels and other associated factors 4. To describe the HIV disease progression, and immune recovery patterns based on CD4 cell counts among adult patients on ART in Zimbabwe accounting for decentralization of ART services to all levels of care 5. To determine spatial heterogeneity of viral suppression and viral rebound trajectories among patient on ART in Zimbabwe adjusting for individual frailty, non-linear and spatial effects using Bayesian estimation 6. To jointly model viral load (VL) and CD4 cell counts prognostic markers in a multistate model framework and identify the shared geographical patterns of low CD4 cell counts and viral rebound patterns Methods This study was a secondary data analysis of routinely collected individual-level data of HIV infected patients who initiated ART between 2004 and 2017. The data was compiled through the v electronic patient management system (ePMS) under the Zimbabwe national ART programme. The final study sample had 390,771 participants from 538 health facilities aged 15 years and above. The study endpoints were mortality if a patient was reported to have died, loss to LTFU defined as a failure of a patient to report for drug refill for at least 90 days from last appointment date or if the patient missed the next scheduled visit date and never showed up again; withdrawal and drop-outs. CD4 cell counts and VL prognostic markers were used to define the finite states of the HIV disease progression models. Semi-parametric time-homogeneous Multistate Markov model was used in the CD4 cell counts model: State 1(CD4≥500cells/uL), State 2(350cells/uL≤CD4<500cells/uL), State 3(200cells/uL≤CD4 <350cells/uL), State 4(CD4<200cells/uL) and the absorbing state death (State 5). The geoadditive Bayesian regression multistate model was fitted in the VL model. The geoadditive Bayesian regression multistate model accounted for the nonlinear effects of continuous variables, fixed effects and spatial variation. In this model, we assumed that individuals from state 1 (VL<50copies/uL) with an undetectable viral load may die (state 3) via state 2 (VL≥50copies/uL) which indicates a viral load rebound or detectable viral load, or directly from state 1 (VL<50copies/uL). Again individuals may move back to state 1(VL<50copies/uL) once they are in state 2 (VL≥50copies/uL) after ART which shows a reversible transition. Time-varying log-baseline effects of the transition intensities and nonlinear effects of continuous covariates were estimated as smoothed functions of time using penalized splines (P-splines). Non-parametric effects of fixed covariates and frailty effects to account for individual variability were also considered. Lastly, adjusted transition intensities, hazard ratios (HR) and regression coefficients were estimated from the joint modelling of CD4 cell counts and VL. Joint mapping of HIV disease progression conditions was done using the Bayesian intrinsic Multivariate Conditional vi Autoregressive (MCAR) prior model. The Kolmogorov-Chapman forward equations were also used to estimate transition rates on a partially observed aggregated data. Results: From a total of 390,771 participants enrolled in 538 health care facilities, the proportions of mortality, becoming LTFU and attrition were 4.7% (n=18,328), 22.7% (n=88,744) and 31.2% (n=121,875), respectively. LTFU rate was 5.75(95% confidence interval (CI): 5.71-5.78) per 100 person-years. Those who had CD4 cell counts below 200cells/uL had a mortality rate of 5.7 per 100 person-years (4,541 deaths). Adjustment for the competing event independently increased LTFU rate ratio in provincial and referral (adjusted sub-hazard ratios (AsHR) 1.24; 95%CI: 1.20- 1.28) and district and mission (AsHR 1.47; 95%CI: 1.45-1.45) hospitals; higher-level health facilities were associated with an increased risk of ART drop-out (RR=1.46; 95%CI: 1.43-1.48 (secondary level facilities) and RR=1.41; 95%CI: 1.43-1.45 (tertiary level facilities); higher levels of care had an increased risk of high rate ratios of LTFU and mortality while primary health care (PHC) facilities had an increased risk of poor immune recovery. Adolescence and young adults (15-24 years) group had an increase rate ratio of becoming LTFU. The effects of patients' age on mortality increased with age while the effects of age on LTFU increased with age for the 15-20 years age group and decreased after that; patients aged 20-60 years had a high risk of attrition, and patients aged 45+ years were more likely to immune deteriorate. Male patients had a higher risk of becoming LTFU; 32% increased risk of death; and a higher risk of immune deterioration. The longer the individuals were on ART, the less likely they were to drop-out (district level: RR=0.7552; 95%CI: 0.69-0.83); health facility level: RR=0.8194; 95%CI: 0.81-0.83). vii There existed a strong shared geographical pattern of 66% spatial correlation between the relative risks of VL rebound and low CD4 cell counts from the joint mapping model. High relative risks of VL  50copies/uL and CD4 cell counts  350cells/uL after ART initiation was observed in Matabeleland North and Mashonaland East provinces. The geoadditive Bayesian model showed that mortality rates decreased with increase in the time since ART initiation and viral rebound transition was significantly prevalent among patients living on the long-distance truck route region (Matabeleland North province) which borders with Botswana and Zambia. The southern and northern parts of Zimbabwe show a high risk of ART drop-out based on the structured effects. Conclusions: ART drop-outs or "leakages" within the ART treatment cascade persist after the decentralization of ART services. Interventions which immediately link HIV patients to care and strengthen patients’ retention are a priority to achieve viral suppression among ART patients. Though ART uptake has increased after ART decentralisation, poor immune recovery and viral suppression are high in PHC; hence, there is need to upgrade PHC infrastructure and strengthen their VL monitoring capacity for better outcomes. The observed poor immune recovery and viral rebound among subpopulation groups confirm the low viral suppression estimates. The poor viral suppression among males could be due to them being a blind spot in HIV response. Interventions which reach out to men in their "male spaces" and target their risk masculinity norms are encouraged to boost their HIV services uptake. The poor immune recovery among adolescents requires the launching of the social deprivation programmes to break food insecurity pathways and financial support, and the paediatric to adult transition programme to mitigate treatment lapses. Moreover, targeted interventions to highly burdened regions to improve patients' viii retention coupled with strategies which will enhance ART adherence are core in the global HIV control campaign. The main strength of this study was the statistical rigour which involves extending the Bayesian multistate model with the spatial covariate accounting for individual frailty terms and joint modelling of VL and CD4 cell counts. HIV decision-making programmes, model inferences based on BE perspective are the preferred as they provide coherent estimates despite model fitting challenges associated with Bayes models. However, future studies should consider assessing the time-varying effects of covariates; determine underlying latent covariates associated with ART outcomes, but the issues on the generalizability of research finding remain a priority to epidemiologists. Strategies which strengthen data quality of routinely collected programme data are a priority in this era of evidence-based decision making, for valid and precise findings.