Predictors of longitudinal viral load suppression among adults receiving first line antiretroviral therapy in Central Ethiopia Authors: Charuai Suwanbamrung, Awoke Abraham, Meklit Asheber, Mesfin Menza Jaldo, Tagese Yakob Published: 2025 ABSTRACT HIV/AIDS continues to be a major global public health issue and a significant cause of death. While the WHO advocates for viral load testing as the primary approach to monitor treatment and detect antiretroviral therapy (ART) failure, the factors affecting viral load trends are frequently neglected. This study exclusively analyzes viral load trajectories, identifying predictors of virological suppression independent of adverse outcomes. This study aimed to assess predictors of longitudinal viral load suppression HIV patients on first-line ART in Central Ethiopia: a mixed-effects analysis. A retrospective follow-up study was conducted in public hospital in Central Ethiopia. A total of 376 adult patients who started first-line ART and had at least two viral load measurements between March 1, 2017, and November 30, 2021, were selected using simple random sampling. Follow-up continued until November 30, 2022.Variables with a univariable association (p < 0.20) with viral load changes were included in the multivariable analysis. A linear mixed-effects model was applied, with statistical significance set at 5%, using adjusted coefficients and 95% confidence intervals. To maintain focus on viral load trends, we excluded time-to-event endpoints and adverse reaction analyses. A total of 376 adult patients on anti-retroviral therapy were assessed. WHO clinical stage (stage II) (B = 0.199: p < 0.0001), (stage III) (B = 0.2: p = 0.0318), (stage IV) (B 0.37: p = 0.0011), CD4 count ( > = 200) (B = − 0.2: p = 0.0070) and poor adherence increased log VL by 0.36, p < 0.001 were found to have a significant effect on the log of viral load. In this study, we have found an overtime decrement in the log of the Viral Load of patients with HIV on ART. Factors such as baseline CD4 count, WHO clinical stage, and adherence were found to be significant predictors of log Viral Load evolution. In order to maximize the results of the policy of test and treat, we suggest that health professionals focus on the interventions that will result in the initiation of ART and long-term viral suppression.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-29159-z. Introduction Acquired Immunodeficiency Syndrome (AIDS) is caused by the Human Immunodeficiency Virus (HIV), a retrovirus that attacks the immune system, specifically CD4 cells, weakening the body’s ability to fight infections1. The advent of antiretroviral therapy (ART) marked a turning point in HIV management, transforming it from a fatal disease to a chronic condition by reducing mortality and improving patients’ quality of life2,3. At the end of 2020, there were an estimated 37.7 million [30.2–45.1 million] persons living with HIV, with over two-thirds (25.4 million) living in the African Region. In the same year,680 000 [480 000–1.0 million] individuals died of HIV-related causes, while 1.5 million [1.0–2.0 million] more became infected with the virus and 68% of adults living with HIV received lifelong antiretroviral therapy (ART)1,4. East and Southern Africa remain the most seriously affected region in the world by the pandemic1,5. In Ethiopia, the first HIV case was identified in 1984, leading to widespread loss of life and leaving countless children orphaned6. The country faces a significant HIV burden, with an overall prevalence of 1.1% across all age groups. About 59% of those infected are on highly active antiretroviral therapy (HAART)7. Regional disparities exist, with Gambella (4.8%) and Addis Ababa (3.4%) having the highest prevalence, while Somali (< 0.1%) and SNNP (0.4%) report the lowest. Nationally, adult prevalence stands at 0.9%, with urban areas (2.9%) and women (1.2%) disproportionately affected compared to rural regions (0.4%) and men (0.6%)8. It remains a pandemic and serious public health concern. Access to HIV therapy is critical in the global drive to eradicate AIDS as a public health threat. The major goals of starting ART in HIV patients are to reduce viral replication and restore immunological function. Periodic viral load tests and/or CD4 cell counting are used in clinical practice to determine whether such goals have been met8. HIV remains a pressing global health challenge, and effective treatment is crucial to ending the epidemic. ART aims to suppress viral replication and restore immune function, with treatment success monitored through viral load (VL) and CD4 cell counts8. Viral load measures HIV particles per milliliter of blood and is a stronger predictor of disease progression than CD4 counts9. However, improper monitoring can lead to unnecessary treatment switches. Evidence demonstrate that 70% of suspected treatment failure cases did not actually have elevated VL, meaning many patients could have been incorrectly moved to second-line therapy without VL confirmation10. Failure to suppress viral replication increases the risk of drug resistance and HIV transmission11,12. High VL also makes individuals more infectious13. Regular VL testing is essential to assess treatment adherence, drug efficacy, and viral resistance14. The WHO recommends VL testing as the gold standard for detecting ART failure15. A suppressed VL (< 1000 copies/mL) indicates effective treatment, while a high VL (> 1000 copies/mL) suggests poor adherence or drug resistance16. Despite WHO recommendations for routine viral load monitoring, few studies in Ethiopia have modeled longitudinal VL trends independent of clinical outcomes like adverse drug reactions. While prior research has linked virological failure to ART toxicity17, this study uniquely isolates viral suppression dynamics to inform monitoring protocols in resource-limited settings. Therefore, this study aimed to assess predictors of longitudinal viral load suppression HIV patients on first-line ART in Central Ethiopia and this study uniquely focuses on modeling longitudinal viral load trends independent of clinical outcomes, applying linear mixed-effects models to account for within-patient correlations inadequately addressed by classical methods. Modeling longitudinal VL changes in HIV patients on ART will identify predictors of VL trends using advanced statistical methods and providing evidence for better clinical management and targeted interventions. The findings will serve as a baseline for future research and help policymakers develop data-driven strategies to reduce HIV-related morbidity and mortality. Methods and materials Study settings, study design, and period This retrospective follow-up study examined adult ART patients at Wachemo University Comprehensive Specialized Hospital in Hosanna Town, Central Ethiopia (235 km from Addis Ababa; 194 km from Hawassa) from March 2017 to November 2021. The hospital offers comprehensive HIV services including VCT, PITC, PMTCT and ART programs. Among 832 eligible ART patients (March 2017-November 2021), medical record data were extracted from December 2022-January 2023. Study population The source population comprised all ART patients aged ≥ 15 years at the hospital’s clinic. The study population included adults (≥ 15 years) who: initiated first-line ART between March 2017-November 2021, had ≥ 2 viral load measurements and had documented treatment initiation dates. Patients without recorded initiation dates were excluded. Sample size and sampling procedure As with cross-sectional studies, investigators conducting longitudinal studies need to know in advance the number of subjects approximately required to achieve a specified statistical power. The sample size was calculated by using the Diggle et al.18 formula for sample size calculation of repeated measurements.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:n=\frac{4{\left(\frac{Z\alpha\:}{2}+Z\beta\:\right)}^{2}{\sigma\:}^{2}\left(1+\left(m-1\right)\rho\:\right)}{m{d}^{2}}$$\end{document} Where. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d$$\end{document} is the effect size sample.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} is the number of time points repeated measurement was taken,\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} is the correlation between repeated measurements and.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:2$$\end{document} is the variance of the outcome variable. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d$$\end{document} is the effect size sample. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d$$\end{document} is the effect size sample. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} is the number of time points repeated measurement was taken, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} is the number of time points repeated measurement was taken, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} is the correlation between repeated measurements and. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} is the correlation between repeated measurements and. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:2$$\end{document} is the variance of the outcome variable. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:2$$\end{document} is the variance of the outcome variable. Assuming a significance level of 0.05, power of 0.8, ρ = 0.9, and effect size of 0.3 (poor adherence) from the previous study conducted in Zewditu Hospital19, we have \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\:=\:6$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}^{2}$$\end{document}= \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:0.8$$\end{document} (random intercept and random slope model) Using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{Z\alpha\:}{2}=\:1.96$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Z\beta\:=0.842$$\end{document} and inserting all quantities in the formula:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n = \frac{{4\left( {1.96 + 0.842} \right)^{2} 0.8\left( {1 + \left( {6 - 1} \right)0.9} \right)}}{{6*0.3*0.3}} = 376$$\end{document} Therefore, the final sample size for this study was 376. The sampling frame was prepared by collecting the identification number of patients on ART from the registration book. After identifying the patients who fulfilled the inclusion criteria, study subjects were selected by simple random sampling technique. Operational definitions BMI: Based on the WHO classification, normal weight is between18.5–24.9 kg/m2, BMI less than 18.5 and greater than 25 are underweight and overweight, respectively20. Adherence was classified as good, fair, or poor, according to the percentage of drug dosage calculated from a monthly total dose of ART drugs. Hence, good was reported if equal to or greater than 95% or ≤ 3 doses missing per month, fair if 85–94% or 4–8 doses missing per month, or poor if less than 85% or ≥ 9 doses missing per month21. Co-morbidities: Diabetes Mellitus, Chronic Kidney Diseases, Hypertension, Asthma, Heart disease and others22. Normal haemoglobin level: indicates free of anemia with an interval of ≥ 11.9 g/dl23. Data collection procedure and quality assurance A standardized checklist was developed for systematic extraction of patient data from medical records, utilizing routinely collected secondary data from hospital follow-ups, with ART clinic health professionals serving as data collectors who retrieved information from patient charts and follow-up cards using unique HMIS identification numbers; socio-demographic characteristics, baseline clinical parameters, and follow-up immunological data were collected electronically using ODK (v1.25.2). To ensure data quality, a 5% sample pre-test validated the checklist’s effectiveness, followed by two days of comprehensive training for data collectors and supervisors to standardize procedures, while daily monitoring by supervisors and the principal investigator verified data completeness and consistency through rigorous checks throughout the collection period. Data management The data were checked for inconsistencies, coding errors, completeness, clarity, and missing values before being entered. The data were exported to R statistical software Version 3.6.1 for further analysis. To avoid misspecification of missing data mechanisms, sensitivity analysis was applied and, we also applied the multiple imputation method and contrasted the outcomes. Statistical analysis Descriptive measures such as mean with standard deviation, median with Inter Quartile Range (IQRs), percentages, and frequencies were used to characterize the study population. Besides, the data were explored graphically by using an individual profile, mean profile; variance-covariance structure, and correlation structure. Consistent with our focus on isolated viral load trends, this analysis employed linear mixed-effects models (LMMs) to assess longitudinal VL patterns. The multicollinearity between predictors was evaluated by using Variance Inflation Factor (VIF) and VIF < 3.5 indicating that there no collinearity among independent predictors. Exploratory data analysis Exploratory data analysis can serve to discover as much of the information regarding raw data as possible, plotting individual curves to carefully determine the data should be performed first before any formal model fitting is carried out. Thus, this study explored the data by using descriptive statistics and a profile plot of the viral load throughout the study and assessed the nature of the data by exploring individual profiles, the average evolution, the variance function, and the correlation structure. Exploring the individual profile To explore the individual profile, a plot of the response with time is used to show whether there is a visible pattern common to most subjects. These individual profiles can also provide some information within and between subject variability. Exploring the mean profile The major purpose of exploring the mean structure is to choose the fixed effects for the model. To explore the overall mean, plotting the response variable against time including individual and overall mean profiles is done. In line with the overall mean, the possible difference between the groups is studied by plotting the mean of each group separately with the same figure. It is used to visualize the difference among the subgroups of mean profiles. Exploring the variance structure To explore the variance structure of the data three plots were used. These are the average evolution of the variance as a function of time, the individual profile plots of the data which show whether there is a considerable within and between subject variability, and the interaction plot which is used to plot the variance functions separated for different groups as a function of time. Exploring the correlation structure It helps to describe how measurements within an individual correlate. Pair-wise scatter plots and pair-wise correlation matrices can be employed for exploring the correlation structures. Linear mixed-effects model LMM is an extension of a linear regression model to model longitudinal data for the continuous response. It contains fixed effects and random effects and the model can be written as follow. Suppose we have N, several subjects, = 1, 2…. And there are repeated measures on each individual. Then the Linear Mixed Model is defined as: Yi = Xi β + Zibi + εi. Where: Yi is the response vector for the ith subject: Yi = (1, 2. . ,)’. Xi is a ni x p design matrix of fixed effects. β stands for a p-dimensional vector of fixed-effect parameters. Zi is a ni x q design matrix of random effects. bi is a q-dimensional vector of random-effect parameters. εi is ni x 1 vector of random errors. The parameter β is the same for all patients and bi is specific to patient i. The random effects, , and εi are assumed to be independent and are normally distributed with mean vector 0 and covariance matrices D and δ2 respectively. Thus the two basic distributional assumptions for the general linear mixed effects model are: β i ~ N(0, D) and. εi ~ N(0, δ2). In the context of longitudinal data, there are three special cases of LMMs. The first one is the random intercept model, which is the simplest type of linear mixed model, the second one is the random slope model and the last one is the random intercept and slope model. The random intercept model As the name indicates, the Random Intercept Model (RIM) is characterized by including an individual intercept as the sole random effect. In our case RIM can be given as: yi = β0 + β1 (x1) + β2 (x2) + …. . + βj (ℎ ) + b0i + εij. b0i ~ N(0, 2 ) ; εi ~ N(0, 2 ). In this case, the distribution of the random effect is a one-dimensional normal distribution with variance 2. Results From a total of 840 patients on ART follow-up at ART clinics of WUCSH in the enrolment period, 238 children and 88 patients with less than two viral load measurements were excluded by inclusion criteria. After all, the final analysis was done on 376 patients on ART (Fig. 1). Fig. 1Patient flow diagram presentation that shows how samples were selected for patients on ART follow-up at WUCSH, Central Ethiopia, 2023. Patient flow diagram presentation that shows how samples were selected for patients on ART follow-up at WUCSH, Central Ethiopia, 2023. Socio-demographic characteristics of the participants In this study, data from 376 patients who were on ART between 2017 and 2021 were collected. The mean baseline age (± standard deviation) of the study participant was 34.8 (± 10) years. Among the study participants, half (52.2%) were females and nearly half of the participants (53.2%) were married. More than a quarter (28.7%) of the study participants attended college and above followed by secondary (26.8%) (Table 1). Table 1Socio-demographic characteristics of patients on ART follow-up at WUCSH, central Ethiopia, 2023.Variables (n = 376)Frequency (%)Sex Male180 (47.8) Female196 (52.2)Marital status Single77 (20.5) Married200 (53.2) Divorced79 (21.0) Widowed20 (5.3)Educational status No formal education80 (21.2) Primary87 (23.1) Secondary101 (26.8) College and above108 (28.7)Occupation Governmental employed121 (32.2) Non-governmental employed66 (17.6) Farmer21 (5.5) Student55 (14.6) Housewife90 (23.9) Table 1Socio-demographic characteristics of patients on ART follow-up at WUCSH, central Ethiopia, 2023.Variables (n = 376)Frequency (%)Sex Male180 (47.8) Female196 (52.2)Marital status Single77 (20.5) Married200 (53.2) Divorced79 (21.0) Widowed20 (5.3)Educational status No formal education80 (21.2) Primary87 (23.1) Secondary101 (26.8) College and above108 (28.7)Occupation Governmental employed121 (32.2) Non-governmental employed66 (17.6) Farmer21 (5.5) Student55 (14.6) Housewife90 (23.9) Socio-demographic characteristics of patients on ART follow-up at WUCSH, central Ethiopia, 2023. Baseline clinical and behavioral characteristics Most, 332(88.3%) of the study participants had baseline CD4 count greater than 200 cells/mm3, and 190 (50.8%) were in the first WHO stage at the baseline. The median baseline CD4 count of study participants was 517 cells/ mm3 with an IQR of (255, 722). The predominant HAART regimen initially prescribed for the patients was a combination of Tenofovir, Lamivudine, and Efavirenz (TDF-3TC-EFV) 203 (55.0) followed by Zidovudine, Lamivudine, and Nevirapine (AZT-3TC-NVP) 80 (21.3%). In this study, while 32.5% of the study participants had ever smoked tobacco products, nearly 38.4% had ever drunk alcoholic drinks (Table 2). Table 2Baseline clinical and behavioral characteristics of patients on ART on follow-up at WUCS Hospital, central Ethiopia, 2023.Variables (n = 376)Frequency / %CD4 count <= 20044 (11.7) > 200332 (88.3)WHO stage (374) Stage I190 (50.8) Stage II117 (31.3) Stage III37 (9.9) Stage IV30 (8.0)BMI Underweight58 (15.4) Normal297 (78.9) Overweight21 (5.6)Past OI Yes38 (10.1) No338 (89.9)Functional status (375) Working280 (74.7) Ambulatory75 (20.0) Bedridden20 (5.3)TB co-infection (370) Yes50 (13.5) No320 (86.5)Initial ART regimen AZT, 3TC, NVP80 (21.3) AZT, 3TC, EFZ73 (19.4) TDF, 3TC, NVP20 (5.3) TDF, 3TC, EFZ203 (54.0)Adherence (371) Good278 (74.9) Fair70 (18.9) Poor23 (6.2)Comorbidity Yes41 (10.9) No337 (89.1)Alcohol consumption (365) Yes140 (38.4) No225 (61.6)Smoking cigarettes(366) Yes119 (32.5) No247 (67.5)AZT: Zidovudine, 3TC: Lamivudine, TDF: Tenofovir, EFV: Efavirenz, NVP: Nevirapine. Table 2Baseline clinical and behavioral characteristics of patients on ART on follow-up at WUCS Hospital, central Ethiopia, 2023.Variables (n = 376)Frequency / %CD4 count <= 20044 (11.7) > 200332 (88.3)WHO stage (374) Stage I190 (50.8) Stage II117 (31.3) Stage III37 (9.9) Stage IV30 (8.0)BMI Underweight58 (15.4) Normal297 (78.9) Overweight21 (5.6)Past OI Yes38 (10.1) No338 (89.9)Functional status (375) Working280 (74.7) Ambulatory75 (20.0) Bedridden20 (5.3)TB co-infection (370) Yes50 (13.5) No320 (86.5)Initial ART regimen AZT, 3TC, NVP80 (21.3) AZT, 3TC, EFZ73 (19.4) TDF, 3TC, NVP20 (5.3) TDF, 3TC, EFZ203 (54.0)Adherence (371) Good278 (74.9) Fair70 (18.9) Poor23 (6.2)Comorbidity Yes41 (10.9) No337 (89.1)Alcohol consumption (365) Yes140 (38.4) No225 (61.6)Smoking cigarettes(366) Yes119 (32.5) No247 (67.5)AZT: Zidovudine, 3TC: Lamivudine, TDF: Tenofovir, EFV: Efavirenz, NVP: Nevirapine. Baseline clinical and behavioral characteristics of patients on ART on follow-up at WUCS Hospital, central Ethiopia, 2023. AZT: Zidovudine, 3TC: Lamivudine, TDF: Tenofovir, EFV: Efavirenz, NVP: Nevirapine. Summaries of the response variable Normality assumption checking of viral load The normality of the data was checked by using a histogram and normal quintile Q-Q plot. Q-Q plot showed that residuals are far from the fitted line and it indicated viral load is not normally distributed. So, transformation was needed then after transforming by log it gets normal distribution. Generally, the repeatedly measured viral load of the patients was measured as continuous and its log transformation satisfies the assumption of normality, indicating that Gaussian longitudinal data analysis which was a linear mixed effect model(LMM) is reasonable. The descriptions (number of observations, mean and standard deviation) of the longitudinal response variable are presented. To provide clear clinical overview of treatment response, we assessed the proportion of patients achieving viral suppression (VL < 1000 copies/ML) over time. The viral suppression rate increased sharply from 0% at baseline to 75.8% at 6 months and 88.3% at 12 months. Suppression rates were sustained above 89% throughout the remaining follow-up period, reaching 100% in the same group of patients with data at 36 and 42 months. This show that a high level of overall treatment effectiveness in study population (Table 3). Table 3Number of observations, the mean and standard deviation of log viral load at each time point for patients on ART at WUCSH, central Ethiopia, 2023.VariablesObservations at a timeMeanS. deviationPatients with viral suppressionLog of viral load at 63766.240.45285(75.8%)Log of viral load at 123765.530.66332(88.3%)Log of viral load at 182265.480.62203(89.3%)Log of viral load at 24505.520.5146(92.0)Log of viral load at 30135.510.4512(92.3%)Log of viral load at 3655.500.465(100%)Log of viral load at 4255.270.315(100%) Table 3Number of observations, the mean and standard deviation of log viral load at each time point for patients on ART at WUCSH, central Ethiopia, 2023.VariablesObservations at a timeMeanS. deviationPatients with viral suppressionLog of viral load at 63766.240.45285(75.8%)Log of viral load at 123765.530.66332(88.3%)Log of viral load at 182265.480.62203(89.3%)Log of viral load at 24505.520.5146(92.0)Log of viral load at 30135.510.4512(92.3%)Log of viral load at 3655.500.465(100%)Log of viral load at 4255.270.315(100%) Number of observations, the mean and standard deviation of log viral load at each time point for patients on ART at WUCSH, central Ethiopia, 2023. Exploratory data analysis Exploring viral load change A minimum of two and a maximum of seven measurements were taken for patients on ART during the study period. From the transformed viral load, there is evidence of between-subject variability and within-subject variability. This means the data were unbalanced and there were a different number of measurements per subject. Subjects have a different log of viral load values at the starting point suggesting that the random intercept model and non-constancy of the trajectory of the log of viral load over time for patients suggests considering a mixed model with random slope could be a plausible starting point (Fig. 2). Fig. 2Individual profile plots showing the progression of the log of viral load over time for patients on ART at WUCSH, Central Ethiopia, 2023. Individual profile plots showing the progression of the log of viral load over time for patients on ART at WUCSH, Central Ethiopia, 2023. The evidence is also supported by a high between and within individual variability over time from both male and female profile plots (Fig. 3). Fig. 3Individual profile plots by Sex showing the progression of log viral load over time for patients on ART at WUCSH, Central Ethiopia, 2023. Individual profile plots by Sex showing the progression of log viral load over time for patients on ART at WUCSH, Central Ethiopia, 2023. Exploring the mean structure The average profile plot was biphasic in nature with a steep reduction in log viral load in the first 12 months, and then, a relative stabilization. Although it may seem that the overall downward trend was dominant and the overall pattern of the mean profile was non-linear, a linear time effect was chosen in the final model in order to represent the overall trend parsimoniously. The clue is also explained by the fact that we did not aim at best trying to predict the predictors of the viral load at the patient level, but instead we took into consideration the temporal trajectory and tried to explain its variation, rather than perfectly predicting it. Such a biphasic trend a sharp drop then a level plateau justifies the application of linear time effect to represent the overall trend. In more granular detail of the treatment response, our sensitivity analysis and the analysis of viral suppression rates (Table 3a) can be used (Fig. 4). Fig. 4The mean profile plot showing the progression of the log of viral load over time for patients on ART at WUCSH, Central Ethiopia, 2023. The mean profile plot showing the progression of the log of viral load over time for patients on ART at WUCSH, Central Ethiopia, 2023. Exploring correlation structure From Table 4, the correlation structure looks irregular overtime. That means the off-diagonal correlation has no known pattern over time, thus unstructured correlation might be appropriate. Table 4Pearson correlation coefficients matrix at each time point for a log of viral load over time for patients on ART at WUCSH, Central, Ethiopia, 2023.Time 1Time 2Time 3Time 4Time 5Time 6Time 7Time 11.00000.59260.0673-0.03100.05670.0022-0.0150Time 20.59261.00000.06590.00100.0339-0.0639-0.0900Time 30.06730.06591.0000-0.00100.0785-0.0224-0.0190Time 4-0.03060.0008-0.00141.00000.03360.02290.0140Time 50.05670.03390.07850.03401.0000-0.00180.0040Time 60.0022-0.0639-0.02240.0230-0.00181.00000.8210Time 7-0.0146-0.0895-0.0193-0.01400.00390.81241.0000 Table 4Pearson correlation coefficients matrix at each time point for a log of viral load over time for patients on ART at WUCSH, Central, Ethiopia, 2023.Time 1Time 2Time 3Time 4Time 5Time 6Time 7Time 11.00000.59260.0673-0.03100.05670.0022-0.0150Time 20.59261.00000.06590.00100.0339-0.0639-0.0900Time 30.06730.06591.0000-0.00100.0785-0.0224-0.0190Time 4-0.03060.0008-0.00141.00000.03360.02290.0140Time 50.05670.03390.07850.03401.0000-0.00180.0040Time 60.0022-0.0639-0.02240.0230-0.00181.00000.8210Time 7-0.0146-0.0895-0.0193-0.01400.00390.81241.0000 Pearson correlation coefficients matrix at each time point for a log of viral load over time for patients on ART at WUCSH, Central, Ethiopia, 2023. Longitudinal data modelling Random effect selection As clearly shown in Table 5, the random intercept and slope model is a much better fit because it has a much lower AIC (1274.605) as compared with the other models. It is in line with the individual profile plot. Table 5Model of fit indices for comparing of random effect structure on ART at WUCSH, Central, Ethiopia, 2023.ModelAICBICLog-likelihoodMixed effect model with random intercept only2114.4082129.263-1054.2038Mixed effect model with random slope only1810.2651825.120-902.1325Mixed effect model with random intercept and random slope1274.6051418.206-608.3023AIC: Akaike information criterion, BIC: bayesian information criterion Table 5Model of fit indices for comparing of random effect structure on ART at WUCSH, Central, Ethiopia, 2023.ModelAICBICLog-likelihoodMixed effect model with random intercept only2114.4082129.263-1054.2038Mixed effect model with random slope only1810.2651825.120-902.1325Mixed effect model with random intercept and random slope1274.6051418.206-608.3023AIC: Akaike information criterion, BIC: bayesian information criterion Model of fit indices for comparing of random effect structure on ART at WUCSH, Central, Ethiopia, 2023. AIC: Akaike information criterion, BIC: bayesian information criterion Linear mixed effect model with random intercept and slope To explore the relationship between the log of VL and each of the independent variables, uni-variable analysis for each independent variable was assessed and those found to be significant (p < 0.20) were selected for the multi-variable analysis. WHO clinical stage, baselined CD4 count, and adherence was significant predictor of log VL, according to current study finding from the linear mixed-effects models (Table 6). The linear effect of time was not statistically significant (β = -0.0018, p = 0.554), shows that after accounting for the sharp initial decline visible (Fig. 4), the overall longitudinal trend was stable. Initially, a positive relationship with ART duration was counter-intuitive and significant in this model but was subsequently found to be due to attrition bias (see Sensitivity Analysis). The estimated model for a log of VL is stated as follows: Log (viral load\bi)) = 5.54–0.2CD4 ( > = 200) + 0.199Stage II + 0.2Stage III + 0.37Stage IV + 0.36poor adherence + 0.0003 ART duration +\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:bio$$\end{document} + \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:bi1$$\end{document} + \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:i$$\end{document}. Baseline CD4 count had a significant effect on the log of VL of patients. Keeping all other covariates constant, Patients with a baseline CD4 count ≥ 200 had a 0.20-unit lower log₁₀ VL., corresponds to an approximate 37% (95% CI: 25% to 47%) reduction in viral load, compared to patients with CD4 count < 200 (Table 6). Table 6Parameter estimates for the full linear mixed-effects model.Fixed EffectEstimatesS. ErrorZ-Values95% CIp-valuesIntercept5.5430910.472638211.7279784.617–6.469< 0.0001Time-0.0017600.0029713-0.592393-0.008-0.00410.5540Baseline age0.0030000.00206411.453662-0.001-0.0070.1469Educational StatusNo formal EducationRefPrimary-0.0450410.0588027-0.765969-0.160-0.0700.4442Secondary-0.0236200.0548177-0.4308910.131 − 0.0840.6668College n above0.0417320.05462130.764027-0.065-0.1490.4454Baseline CD4< 200Ref>=200-0.2003880.0738623-2.712997− (0.345 − 0.056)0.0070 *WHO clinical stageStage IRefStage II0.1991820.04762044.1826980.106–0.293< 0.0001 *Stage III0.2002870.09290682.1557780.018–0.3820.0318 *Stage IV0.3759160.11447653.2837840.151-0.6000.0011 *AdherenceGoodRefFair0.0701900.06065921.157118-0.048-0.1890.2480Poor0.3640970.10123363.5966010.165–0.5630.0004 *Functional statusWorkingRefAmbulatory0.0373460.06063170.615941-0.018-0.1560.5383Bedridden-0.1787300.1337494-1.33630-0.441-0.0830.1823Haemoglobin0.0275550.03377330.815888-0.038-0.0930.4151ART duration0.0003000.00010282.6992800.0001–0.00050.0073 *Other OIYesRefNo-0.1212990.0749291-1.618857-0.268-0.0250.1064TB co-infectionYesRefNo-0.0243430.0762348-0.319317-0.174-0.1250.7497AlcoholYesRefNo-0.0268880.0688239-0.390676-0.162-0.1080.6963SmokingYesRefNo-0.0314880.0744536-0.422923-0.177-0.1140.6726ComorbidityYesNo-0.0137280.0666861-0.205860-0.144-0.1170.8370Ref: Reference, OI: Opportunistic infection ,TB: Tuberculosis, CI: confidence interval, ART: Antiretroviral Therapy, WHO: World Health Organization clinical stage. Table 6Parameter estimates for the full linear mixed-effects model.Fixed EffectEstimatesS. ErrorZ-Values95% CIp-valuesIntercept5.5430910.472638211.7279784.617–6.469< 0.0001Time-0.0017600.0029713-0.592393-0.008-0.00410.5540Baseline age0.0030000.00206411.453662-0.001-0.0070.1469Educational StatusNo formal EducationRefPrimary-0.0450410.0588027-0.765969-0.160-0.0700.4442Secondary-0.0236200.0548177-0.4308910.131 − 0.0840.6668College n above0.0417320.05462130.764027-0.065-0.1490.4454Baseline CD4< 200Ref>=200-0.2003880.0738623-2.712997− (0.345 − 0.056)0.0070 *WHO clinical stageStage IRefStage II0.1991820.04762044.1826980.106–0.293< 0.0001 *Stage III0.2002870.09290682.1557780.018–0.3820.0318 *Stage IV0.3759160.11447653.2837840.151-0.6000.0011 *AdherenceGoodRefFair0.0701900.06065921.157118-0.048-0.1890.2480Poor0.3640970.10123363.5966010.165–0.5630.0004 *Functional statusWorkingRefAmbulatory0.0373460.06063170.615941-0.018-0.1560.5383Bedridden-0.1787300.1337494-1.33630-0.441-0.0830.1823Haemoglobin0.0275550.03377330.815888-0.038-0.0930.4151ART duration0.0003000.00010282.6992800.0001–0.00050.0073 *Other OIYesRefNo-0.1212990.0749291-1.618857-0.268-0.0250.1064TB co-infectionYesRefNo-0.0243430.0762348-0.319317-0.174-0.1250.7497AlcoholYesRefNo-0.0268880.0688239-0.390676-0.162-0.1080.6963SmokingYesRefNo-0.0314880.0744536-0.422923-0.177-0.1140.6726ComorbidityYesNo-0.0137280.0666861-0.205860-0.144-0.1170.8370Ref: Reference, OI: Opportunistic infection ,TB: Tuberculosis, CI: confidence interval, ART: Antiretroviral Therapy, WHO: World Health Organization clinical stage. Parameter estimates for the full linear mixed-effects model. Ref: Reference, OI: Opportunistic infection ,TB: Tuberculosis, CI: confidence interval, ART: Antiretroviral Therapy, WHO: World Health Organization clinical stage. There are two estimated variance components; these are the random effects standard deviation and the residual standard deviation. The residual standard deviation is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\sigma\:}\boldsymbol{\epsilon\:}=0.59$$\end{document} and for the random effects, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\delta\:}\boldsymbol{b}\boldsymbol{o}\boldsymbol{i}$$\end{document}=0.19 and\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\delta\:}\boldsymbol{b}1\boldsymbol{i}=0.12$$\end{document}. The total variability between individuals is estimated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\delta\:}\boldsymbol{b}\boldsymbol{o}\boldsymbol{i}$$\end{document}= 0.19 +\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\delta\:}\boldsymbol{b}1\boldsymbol{i}=0.12\:\boldsymbol{i}\boldsymbol{s}\:\boldsymbol{e}\boldsymbol{q}\boldsymbol{u}\boldsymbol{a}\boldsymbol{l}\:\boldsymbol{t}\boldsymbol{o}\:0.31$$\end{document}, whereas the total variability within the individual is\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\:\boldsymbol{\sigma\:}\boldsymbol{\epsilon\:}=0.59$$\end{document}. However, the total variation in the log of VL is estimated to be 0.31 + 0.59 = 0.90. The proportion of total variability that is attributed to within-person variation is given by 0.59/0.90 is 65.5% while the proportion of total variability attributed to between individual variations in their general level of log VL is 0.31 /0.90 is 34.4% (Table 7). Table 7Estimates of random components.Random componentsEstimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:boi$$\end{document} 0.19 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:b1i$$\end{document} 0.12 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\epsilon\:ij$$\end{document} 0.59 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:\delta\:boi\:\delta\:b1i$$\end{document} 0ICC0.31 Table 7Estimates of random components.Random componentsEstimates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:boi$$\end{document} 0.19 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:b1i$$\end{document} 0.12 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\epsilon\:ij$$\end{document} 0.59 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:\delta\:boi\:\delta\:b1i$$\end{document} 0ICC0.31 Estimates of random components. Sensitivity analysis and investigation of attrition bias The initial analysis showed the statistically significant, counter-intuitive, positive correlation between ART duration and log viral load. To examine the possibility of the effect of attrition bias on this finding, we did a sensitivity analysis and compared the baseline attributes of patients who remained in the study with those who were lost to follow-up. The sensitivity analysis was conducted by re-running the linear mixed-effects model with data only available during a time that has better patient retention (reaching a high patient retention of 50 at this time point). The relationship of ART duration and log viral load was no longer significant in this model ( -0.000102, p = 0.451) indicating that the paradoxical outcome of the full model was caused by later follow-up periods with high dropout (Supplementary Table 1). In order to estimate informative censoring, we compared the baseline parameters between patients who were lost to follow-up (LTFU) at 24 months (n = 326) to the patients who were followed up after 24 months (n = 50). There was a crucial difference in both poor or fair adherence at baseline (32% vs. 18%, p = 0.04) and no significant change in the tendency of higher viral loads at the time of baseline and reduced counts of CD4 + in the LTFU group. This shows that the cohort followed-up in the long term was a non-random and healthier subset and this biased the effect of ART duration in the complete model (Supplementary Table 1). Model diagnostics The most crucial part of model development is checking the underlined assumption. Residuals have an undeniable role in the process of checking those assumptions. Firstly, we have a residual versus predicted value plot for the final linear mixed model. Discussion Unlike studies examining viral load as a predictor of clinical outcomes, our analysis demonstrates that longitudinal VL trends in Ethiopian ART patients are primarily driven by modifiable factors advanced WHO stages (β=+0.38 for stage IV, p < 0.001), optimal adherence (β=-0.36, p < 0.001), and higher baseline CD4 (β=-0.20, p = 0.007) independent of adverse events. These findings uniquely inform routine VL monitoring protocols in resource-limited settings, where isolating virological suppression from toxicity endpoints is critical for programmatic decision-making. This study has found a decrement in the log of VL of patients throughout the treatment. There were differences between subjects at baseline and during their progress over time. Factors found to have a significant effect on the evolution of the log of VL at a multivariable linear mixed-effects model, CD4 count, WHO clinical stage, and adherence. Baseline CD4 count was significant predictor of viral load. Patients with a CD4 count ≥ 200 associated with a 0.2-unit decrease in log of VL, when all other factors in the model were keep constant. This finding is supported by other studies conducted in Botswana24, in Vietnam25, and in Ethiopia26. When all other factors are held constant, individuals with poor adherence have an extra 0.36 copies/ml increment in their log of VL for each additional unit of treatment duration. This is because a low level of antiretrovirals in the body brought on by non-adherence is insufficient to stop viral reproduction, which causes a larger viral load to be found in the blood. This suggests that adherence is crucial for patients’ viral loads to continue to decline successfully. This finding is in agreement with previous studies conducted in the University of Gondar referral hospital27 and Zewditu Hospital at Addis Ababa26. Current study revealed important changes that can be altered to enhance viral load suppression which is which align with global evidence. ART response is a biphasic process: the initial quick reduction in viral load, indicated by the fact that the rates of suppression increase to 88.3% in 12 months (Table 3), and then the stable maintenance phase is achieved. This stability is also manifested in the non-significant linear impact of time in our main model (β = -0.0018, p = 0.554). Therefore, the initial year of treatment is very important in attaining suppression. One important methodological insight was also identified in our research. Sensitivity analysis revealed that the large, counterintuitive positive effect between ART length and suppression, as found in the complete linear model was an artefact of attrition bias (see Supplementary Table 1). When the follow-up was confined to 24 months, whereby the retention increased, the effect disappeared. This indicates a higher level of healthiness of patients that were booked in 90-day follow-ups in the long term (see Supplementary Table 2). Long-term therapy was apparently negative, but it was merely a statistical artefact. With this bias corrected, the combined data have validated the long-held principle that appropriately used ART can be used to provide uniform virological suppression. The Advanced WHO clinical stage had a significant effect on the log of VL. Advanced WHO clinical stage (II, III, IV) are poorer viral load suppression (high viral logVL) compared to WHO clinical stage I.This study was in line with the study conducted showed that beginning ART at a later stage of the WHO had a higher chance of bad outcomes than beginning ART at an earlier stage28. Policy implications Our findings support WHO’s call for decentralized VL monitoring29. Task-shifting to nurses could improve coverage, as pilot programs in Malawi achieved 80% VL suppression30. Strengths and limitations of the study As per our knowledge, only one study was done to identify predictors of longitudinal change in viral load among patients on ART. Therefore, this study is the baseline for future studies and can be an input for policymakers and program planners. The study applied an advanced statistical model. Despite the above strength, this study is not without a limitation. As the current study was based on secondary data obtained from patient medical records, some important variables like socio-economic status and clinical variables which are assumed to have an association with the outcome variable were not obtained in the record. Measurement of biomarkers such as viral load and CD4 count was not consistently recorded. Besides, since viral load measurement started recently there are not enough repeatedly measured viral loads to see the perfect association. While we analysed VL trends independently of clinical outcomes, future studies could integrate ADRs via joint modelling. while sensitivity analysis was used to probe attrition bias, the possibility of residual bias remains. Conclusion A linear mixed model with random intercept and slope was selected to fit the data based on different information criteria. The effects of several factors on the evolution of the log of VL were identified. Among these baseline CD4 counts, WHO clinical stage, and adherence were found to be significant predictors of log VL evolution. In this research, we observed a robust primary virological reaction to ART. The rates of viral suppression indicated high rates of viral loads dropped very quickly during the first year and then flattened. To reduce viral load, strengthening routine and consistent measurement and registration of basic biomarkers of ART treatment outcome such as CD4 count is needed to monitor patient’s progress. By considering the limitation of the current study, a prospective follow-up study is needed to identify all predictors and the potential biomarkers such as weight, and CD4 count. To fully exploit the association between the repeated measurement of viral load and the factors, we recommend a joint modelling approach with different parameterizations at different Centres having large sample sizes. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2