A prediction model for virologic failure in adolescents living with HIV in Uganda: Findings from the Suubi+Adherence study

Abstract

Objectives:

Adolescents living with HIV (ALHIV) have low viral suppression levels, with 1 in 3 ALHIV experiencing virologic failure, calling for more efforts to reverse these trends. We developed and validated a model that predicts the risk of virologic failure (VF) among ALHIV.

Study design:

Cross-sectional study.

Methods:

We used baseline data from 702 ALHIV enrolled in the Suubi + Adherence cluster-randomized clinical trial. Participants were aged 10–16 years, living with HIV and aware of their HIV status, and are living with a family. We developed a risk-prediction model for VF (viral load of ≥200 copies/mL) using sociodemographic, behavioral, psychological, economic, and treatment-related factors. LASSO logistic regression using 10-fold cross-validation with bootstrapping was used to select the predictors for the final model. Model performance was assessed by determining the discrimination using the area under the curve and calibration by drawing a calibration plot.

Results:

Using a lambda value of 0.007, the final model had 24 predictors (and interaction terms). The predictors included the participants’ age, sex, work status, stigma, depressive symptoms, adherence self-efficacy, HIV knowledge, duration with HIV, time spent on ART, communication with the caregiver, family cohesion, social support, orphanhood status, number of people in the household, HIV disclosure, years spent at the current residence, and household asset ownership. The model predicted VF with AUC of 73.8 (95 % CI: 68.3–78.0) and calibration slope of 0.985.

Conclusions:

We developed and validated a model to predict the risk of virologic failure among ALHIV in Uganda, demonstrating its potential utility in identifying individuals at elevated risk for VF. Future models could be refined by incorporating clinical characteristics such as CD4 count to further improve predictive accuracy.

1. Introduction

Over 1 million youths live with HIV in 2023, with the majority living in Sub-Saharan Africa (SSA).¹ Successful management of HIV/AIDS is contingent on suppressing the HIV, primarily through sustained antiretroviral therapy (ART) adherence. Viral suppression carries individual benefits by slowing disease progression and increased survival and benefits the community by limiting HIV transmission. However, over one third of adolescents living with HIV (ALHIV) on ART experience virological failure (VF)—characterized by uncontrolled viral replication despite ongoing treatment²— with the studies showing higher VF rates for Uganda (the location of the proposed study), reported at 45 %.³

The World Health Organization (WHO) recommends viral load testing six months after antiretroviral therapy (ART) initiation and every 12 months thereafter.⁴ However, SSA, the region affected most by HIV, faces a scarcity of healthcare resources,⁵ necessitating the prioritization of the care for ALHIV that require the most support. However, this approach demands an effective strategy to identify the most vulnerable ALHIV. While traditional regression models have provided valuable insights into the factors influencing VF such as HIV knowledge, mental health functioning, adherence self-efficacy, HIV stigma, and access to HIV care,^6,7 they fall short in identifying specific ALHIV at elevated risk for VF and who would benefit most from targeted interventions.

Predictive modeling has emerged as a powerful data-driven tool in patient-centered care, capable of stratifying patients based on their risk, and enables providers to individualize treatment.^8,9 Developing a reliable prediction model for VF would contribute to precision medicine in HIV management, emphasizing the need for personalized care based on individual risk profiles. Moreover, early prediction of increased VF risk can guide allocation of services, such as enhanced adherence counseling, multi-month prescription, and prompt referral for those in most need However, to our knowledge, no model has been developed to explicitly predict VF in ALHIV in SSA. Therefore, this study aimed to develop and validate a model to predict the risk of VF among ALHIV in Uganda. Guided by the socioecological model in the selection of candidate predictors, and informed by evidence from risk prediction models for other HIV outcomes (ART adherence) among ALHIV,¹⁰ we hypothesize our model would predict the risk of VF among ALHIV in Uganda, with an area under the curve (AUC) of at least 0.8.

2. Methods

2.1. Study design

This was a cross-sectional study that used baseline data from the Suubi + Adherence Study (R01HD074949: Ssewamala, PI), a two-group cluster-randomized controlled trial.

2.2. Study setting and participants

The study recruited participants from 39 health clinics located in Southern Uganda. The inclusion criteria for the current study were the same as those used in the parent Suubi + Adherence Study. Specifically, adolescents were included in the study if they were: 1) Aged 10–16 years; 2) Receiving medical confirmation of their HIV-positive status; 3) Aware of their HIV status; 4) Living within a family setting. Recruitment took place between September 2013 and July 2014, during which a total of 702 ALHIV were enrolled. Upon enrollment, all participants underwent an interviewer-administered questionnaire, which typically lasted between 60 and 90 min.

2.3. Theoretical framework

We adopted the social-ecological model¹¹ to inform the selection of the candidate predictors to include in the risk-prediction model. It divides the ecological system into levels that interact with and influence each other.¹² The individual (intrapersonal) level factors mainly include biological and behavioral attributes, such as depressive symptoms and adherence self-efficacy.^10,13 The interpersonal/relationship level comprise factors such as family and social support for the ALHIV to achieve and maintain health-promoting behavior including adhering to their medication.¹⁴ The community level is more concerned with the broader context in which ALHIV reside. For ALHIV, factors such as HIV stigma and disclosure of HIV status influence the viral suppression at this level.¹⁵ At the structural level in LMICs, poverty is an important factor that has undermined many efforts against HIV including ART adherence and downstream outcomes.^16,17

2.4. Study measures

2.4.1. Viral load

The outcome for this study was virologic failure, defined as viral load of >200 copies/ml, as recommended by the 2023 Uganda national HIV treatment guidelines.¹⁸ This virological definition was chosen as it reflects the most objective measure of treatment failure, distinct from clinical or immunological criteria. To perform the viral load test, blood samples were collected from each ALHIV using EDTA tubes, and the assays were conducted using the Abbott Real-Time HIV-1 RNA Assay PCR, version 5.

2.4.2. Candidate predictors

Supplementary Table 1 describes the candidate predictors used in developing the risk-prediction model. For all variables measured using a scale, a cumulative score was calculated by summing the individual item scores, ensuring that items phrased in the opposite direction were reverse coded where necessary.

2.5. Data analysis

All descriptive analyses were conducted using Stata, version 18.0. Continuous variables were summarized using means and standard deviations, while categorical variables were summarized with frequencies and percentages. Baseline characteristics were compared across the two study groups to ensure the effectiveness of the randomization. To compare the groups, while adjusting for clustering, survey data commands including svy: regress and svy: tab for continuous and categorical data were used, respectively.

2.6. Model development

Candidate predictors were examined to ensure they met the model assumptions, including linearity and no outliers. Variables that demonstrated a non-linear relationship with virologic failure were transformed using logs (depressive symptoms and years spent while HIV positive), inverse transformation (HIV stigma), square root (social support), and restricted cubic splines, which was used for HIV knowledge, and the number of years spent at the current address.

The least absolute shrinkage and selection operator (LASSO) penalized regression was used to select the subset of predictors that contributed the most in predicting VF.⁸ LASSO performs data shrinkage by applying a penalty factor (λ) to the regression coefficients. The best penalty factor (λ) is then selected using 10-fold cross-validation with bootstrapping. Applying the penalty results in shrinking the small coefficients to zero hence eliminating the least contributing variables from the model. The penalization also ensures that very large coefficients are shrunk to lesser values. This way, penalization avoids extreme predictions.

We then determined the contribution of each variable to VF and calculated the individualized risk for each participant using the formula, $risk=1/(1+e[-riskscore\left]\right)$.⁸ However, evidence shows that regression coefficients obtained directly from the LASSO models are usually biased and should not be used to accurately compute individualized risks.¹⁹ Instead, to obtain unbiased estimates, LASSO inference models should be used. Therefore, we fitted a cross-fit partialing out logistic regression model using the xpologit command to generate regression coefficients for calculating individualized risk for VF.^19,20 In fitting this model, the variables retained in the LASSO model were included as predictors, while the rest of the variables (initially eliminated by the LASSO model) were added as control variables.

2.7. . Model performance

Model performance was examined by determining the discrimination and calibration. Discrimination was assessed by computing the area under the curve (AUC), which indicates the ability of the model to differentiate between ALHIV with VF versus those who achieved viral suppression. The AUC was obtained by plotting a ROC curve with sensitivity y axis and 1-specificity on the x axis. AUC values between 50 % and <70 % suggest poor discrimination, while AUC values of 70 % to <80 %, <80 %–90 %, and >90 % indicate acceptable, excellent, and outstanding model performance, respectively. The 95 % confidence intervals for the AUC were also generated using bootstrapped resampling. Calibration involved determining the agreement between model-predicted and the observed risk of VF. A calibration plot was generated using the pmcalplot package, with the predicted risk of VF on the x-axis, and the observed risk on the y-axis. In a calibration plot, perfect prediction follows a straight line at 45° with a calibration slope of 1 and intercept through 0. While assessing the model performance, cluster (clinic) bootstrapping with 1000 replicates was performed to further correct for overfitting. Statistical significance was set at a p-value of <0.05. Throughout the analysis, the clinic ID was included in the model to adjust for clustering due to clinic membership.

2.8. Sensitivity analysis and missing data

The risk prediction model was developed using 530 participants who had complete data on all variables. We compared the baseline characteristics of the participants included in developing the models and those that were excluded (see supplementary Table 2). We also performed sensitivity analyses using complete datasets generated through multiple imputation using chained equations (MICE). During multiple imputation (MI), clinic ID was included in the imputation model to account for data clustering at the clinic level. A total of five (5) MI datasets were generated. A limitation of multiple imputation in Stata is that there is no provision for LASSO regression to pool regression estimates from the MI datasets. Therefore, after generating the MI datasets, the mi extract command was used to store separate complete MI datasets. LASSO models were then performed on each of the MI datasets and determined the variables that were consistently selected across the datasets.

3. Results

The mean age of the ALHIV was 12.4 years (SD = 1.98), and more than half (56.4 %) were females. The mean time the ALHIV have been taking ART was 3.9 years (SD = 3.36). Only 35.7 % (n = 246) of the participants were not orphans, while 38 % (n = 262) had lost one of their parents. On average, each ALHIV had lived in the current residence for 8.5 years (SD = 3.54). Overall, 231 participants (32.9 %) experienced virologic failure. Details are shown in Table 1.

Table 1.

Baseline characteristics of 702 adolescents living with HIV in Uganda.

Characteristics	Mean (SD)/number (Percentage)
	Control	Intervention	Total
Individual level
Participant age (min/max: 10 to 16)	12.4 (1.97)	12.5 (1.98)	12.4 (1.98)
Sex assigned at birth (Female)	193 (56.1)	203 (56.7)	396 (56.4)
The participant is in school	301 (87.5)	312 (87.2)	613 (87.3)
Adolescent works for pay	33 (9.59)	32 (8.94)	65 (9.3)
HIV stigma (min/max: 9 to 35)	18.4 (5.95)	18.3 (5.90)	18.6 (5.93)
Depressive symptoms (min/max: 0 to 20)	5.2 (3.69)	5.2 (3.84)	5.2 (3.76)
Adherence self-efficacy (min/max: 20 to 120)	95.2 (23.5)	93.4 (23.0)	94.3 (23.3)
Quality of life (min/max: 4 to 20)	14.4 (3.49)	14.1 (3.64)	14.2 (3.56)
HIV/AIDS clinical knowledge (min/max: 0 to 9)	6.0 (2.00)	6.0 (1.97)	6.0 (1.98)
Duration (years) with HIV (min/max: 0 to 16)	3.7 (2.97)	3.6 (3.12)	3.64 (3.05)
Duration (years) on ART (min/max: 0 to 16)	3.9 (3.22)	4.0 (3.51)	3.94 (3.36)
Interpersonal level
Family cohesion (min/max: 12 to 40)	31.4 (6.72)	32.1 (6.75)	31.8 (6.74)
Social support (min/max: 56 to 120)	87.6 (14.3)	89.2 (13.9)	88.4 (14.1)
Orphanhood status
Non-orphan	115 (33.9)	131 (37.3)	246 (35.7)
Single orphan	129 (38.1)	133 (37.9)	262 (38.0)
Double orphan	95 (28.0)	87 (24.8)	182 (26.4)
People in the household (min/max: 2 to 18)	5.78 (2.46)	5.70 (2.65)	5.74 (2.56)
Community level
Disclosure of HIV status (min/max: 4 to 16)	7.3 (3.08)	7.6 (3.24)	7.5 (3.16)
Years at current residence (min/max: 0 to 16)	8.5 (4.51)	8.6 (4.58)	8.5 (4.54)
Structural level
Household asset ownership (min/max: 1 to 20)	10.9 (3.38)	10.6 (3.69)	10.8 (3.54)
Outcome
Viral suppression	236 (68.6)	235 (65.6)	471 (67.1)

ART: Antiretroviral therapy. SD: Standard deviation.

3.1. Model development

Initially, 37 candidate predictors were included in developing the model. However, after performing LASSO logistic regression with 10-fold cross-validation, 24 predictors were retained in the final model, using a lambda value of 0.0071304. (see Table 2).

Table 2.

LASSO regression for the predictors of virologic failure among ALHIV.

Predictor variable	Penalized coefficients
Individual level
Participant age	0.132
Sex assigned at birth	−0.382
The participant is in school	x
Adolescent works for pay	−1.341
HIV stigmaa	1.913
Depressive symptomsa	0.502
Hopelessness	x
Self-concept	x
Adherence self-efficacy	0.0137
Quality of life	x
Personal health	x
HIV/AIDS clinical knowledge (min/max: 0 to 9)
Spline 1 (0–3)	x
Spline 2 (4–6)	−0.145
Spline 3 (7–9)	0.586
Duration (years) with HIVa	x
Duration (years) on ART
Below five years	x
Five to nine years	x
Ten to 16 years	x
Less than five years on ART# poor ART adherence	−0.382
Less than five years on ART# good ART adherence	x
Five to nine years on ART# poor ART adherence	−0.253
Five to nine years on ART# good ART adherence	0.521
Ten to 16 years on ART# poor ART adherence	0.412
Ten to 16 years on ART# good ART adherence	−0.820
Duration with HIV # poor ART adherence	0.331
Duration with HIV # good ART adherence	x
Interpersonal level
Communication with the caregiver about HIV/AIDS	0.787
Family cohesion (min/max: 12 to 40)
Spline 1 (12–24)	−0.168
Spline 2 (25–34)	0.051
Spline 3 (35–40)	x
Social supporta	−0.018
The participant is an orphan	0.211
Number of people living in the household	−0.031
Community level
Disclosure of HIV status	x
Years spent at current residence (min/max: 0 to 16)
Spline 1 (0–3)	−0.040
Spline 2 (4–9)	0.107
Spline 3 (10–13)	−0.293
Spline 4 (14–16)	x
Structural level
Household asset ownership	−0.001

^aDepressive symptoms and duration with HIV were log-transformed; HIV stigma was inverse-transformed; Social support was transformed using square roots. The model was developed using data from 530 ALHIV, who had complete data on all variables.

3.1.1. Contribution of the predictors to VF

Table 3 shows the results of the cross-fit partialing out logistic regression model, which were used to calculate the individualized risk for VF.

Table 3.

Cross-fit partialing out model for the predictors of virologic failure among 530 ALHIV.

Predictor variable	coefficients
Individual level
Participant age	0.130
Sex assigned at birth	−0.877
Adolescent works for pay	−1.733
HIV stigmaa	8.098
Depressive symptomsa	0.932
Adherence self-efficacy	0.025
HIV/AIDS clinical knowledge (min/max: 0 to 9)
Spline 2 (4–6)	−0.513
Spline 3 (7–9)	0.733
Duration (years on ART
Five to nine years	−0.619
Ten to 16 years	0.242
Less than five years on ART# good ART adherence	1.092
Five to nine years on ART# good ART adherence	3.510
Ten to 16 years on ART# good ART adherence	1.115
Duration with HIV # poor ART adherence	0.984
Duration with HIV # good ART adherence	0.377
Interpersonal level
Communication with the caregiver about HIV/AIDS	0.575
Family cohesion (min/max: 12 to 40)
Spline 1 (12–24)	−0.334
Spline 2 (25–34)	0.173
Social supporta	0.109
The participant is an orphan	0.196
Number of people living in the household	0.016
Community level
Years spent at current residence (min/max: 0 to 16)
Spline 1 (0–3)	−0.224
Spline 2 (4–9)	0.216
Spline 3 (10–13)	0.073
Structural level
Household asset ownership	−0.027

^aDepressive symptoms and duration with HIV were log-transformed; HIV stigma was inverse-transformed; Social.

To demonstrate the application of the results presented in Table 3, the model was used to calculate the risk for VF for a hypothetical adolescent with the following characteristics. Aged 17 years, female, employed, has HIV stigma score of 30, depressive score of 7, adherence self-efficacy score of 95, HIV clinical knowledge of 7, has had HIV for five years, was on ART for five years, has poor ART adherence, never talks with the caregiver about HIV, family cohesion score of 23, social support score of 98, non-orphaned, residing in a household with 13 members, has lived in the current residence for the last eight years, and has asset ownership of 4 items. The risk score for this participant is calculated in the equation below.

$$\mathrm{R}\mathrm{i}\mathrm{s}\mathrm{k}=1/(1+\mathrm{e}\mathrm{x}\mathrm{p}(-(0.013\mathrm{*}\mathrm{A}\mathrm{g}\mathrm{e}-0.877\mathrm{*}\mathrm{S}\mathrm{e}\mathrm{x}-1.733\mathrm{*}\mathrm{E}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{o}\mathrm{y}\mathrm{e}\mathrm{d}+\dots -0.027\mathrm{*}\mathrm{A}\mathrm{s}\mathrm{s}\mathrm{e}\mathrm{t}\mathrm{s}))$$

The risk for VF for this participant, derived from the model, is 36.98 %.

3.1.2. Model discrimination

The final risk prediction model demonstrated strong discriminative power to distinguish ALHIV with or without VF, evidenced by an AUC of 73.8 (95 % CI: 68.3–78.0). Refer to Fig. 1.

Fig. 1.

ROC for a model predicting the risk of virologic failure.

3.1.3. Model calibration

The model had excellent calibration, with a calibration slope of 0.985 (Fig. 2). Further analysis using the calibrationbelt package in Stata showed no miscalibration regions on the calibration plot, as evidenced by a non-significant test statistic value of 2.86, p-value = 0.091. This observation further supports the conclusion of excellent model calibration. Throughout the plot, the perfect-fit (reference) line consistently fell within the calibration band (lime green region), indicating that all the predicted risks of VF agreed with the observed virologic failure.

Fig. 2.

A calibration plot for the performance of the risk prediction model for virologic failure. The calibration plot was generated using the ‘pmcalplot’ package in Stata. The expected risk was partitioned into deciles, and the observed virologic failure rates were plotted against these deciles, with a smoother added (blue line). The green circles, and the dotted vertical lines on either side denote the mean predicted risks for virologic failure and their corresponding 95 % confidence intervals. The reference line is represented by a red straight dotted line, serving as a point of comparison to assess the alignment between predicted and observed virologic failure rates.

3.1.4. Sensitivity analysis

Upon applying the LASSO model to each dataset we extracted using the mi extract command in Stata, we found that the same set of variables—similar to those in the complete case analysis— were consistently retained in the models derived from each MI dataset (Supplementary Table 3).

4. Discussion

Currently, achieving viral suppression is the final goal in the HIV care continuum, and represents treatment success. However, even with the availability of ART, many ALHIV fail to achieve viral suppression. We developed a model that predicts VF among ALHIV, using sociodemographic, psychological, and economic characteristics.

The current model builds on Brathwaite’s model that predicted ART adherence among the same population with AUC of 0.699,¹⁰ and makes the following advancements. (1) The outcome for the current model is VF, which is the end goal for HIV care and carries critical public health implications regarding the increased risk of HIV spread. Moreover, the model applied the recently updated definition of VF of having >200 copies/ml;¹⁸ (2) Additional predictors were incorporated in the current model, such as the duration on ART, and the duration that the participant has lived at the current address; and (3) More rigorous data preparation has been employed including linearization of variables and adding interaction terms to the model.

The model demonstrated good discrimination, and almost perfect calibration. Given the scanty risk prediction models for VF in ALHIV in SSA, the current model could only be compared to those developed to predict VF in adults. Nonetheless, the performance of the current model was comparable to that in some of the models developed among adult HIV populations. For instance, in their model predicting VF among adults with HIV in the USA, Gebrezgi and colleagues found a AUC of 0.763, which was included within the confidence interval of the current model—suggesting comparable performance.²¹ It is worth noting that in their model, they used a different set of predictors—such as alcohol use and having AIDS. Contrastingly, other studies reported higher performance, with a AUC of ~0.8.^22,23 However, a key strength of these studies was the incorporation of clinical characteristics such as the CD4 count as candidate predictors, which are significant predictors of VF.

The predictors retained in our final model fall within various levels of the socioecological framework,¹¹ justifying how VF in ALHIV is shaped by multilevel factors. For instance, at the individual level, factors such as depressive symptoms and low adherence self-efficacy reflect the role of intrapersonal barriers to HIV treatment success. While factors like poor caregiver communication and HIV disclosure status reveal interpersonal and community level risks to viral suppression. Finally, structural predictors (e.g., household asset ownership) emphasize socioeconomic disparities in HIV care outcomes. These findings highlight the need for interventions that address not only the clinical factors, but also psychosocial and structural barriers to care if we are to achieve the 95–95-95 HIV targets.²⁴

Our results have direct implications for public health policy and practice in low-resource settings. The model identifies specific subgroups of ALHIV at highest risk for virologic failure, enabling policymakers to prioritize the scarce resources—such as intensified adherence counseling or community-based support programs—toward the more vulnerable adolescents. For practitioners, our model offers a practical tool for risk stratification in clinic or community settings. The predictors that were retained in the final model can be assessed during routine visits using brief, standardized tools. Clinics could adopt a two-tiered approach: (1) universal viral load monitoring per WHO guidelines, coupled with (2) targeted interventions for high-risk adolescents flagged by the model (e.g., peer support groups for those with low social support, or family-based counseling for poor caregiver communication). The implementation of such an approach would require minimal infrastructural support, as the model relies on non-laboratory variables feasible for low-resource settings.

A major strength of the current model is the use of a large sample size, recruited from 39 health clinics across Southern Uganda, covering diverse socioeconomic and geographic settings. This enhances the generalizability of the model to similar populations in low-resource settings. In addition, we employed robust methods such as the use of LASSO regression model in selecting the predictors (which is recommended due to its ability to handle multicollinearity, minimize overfitting, and shrink less important predictors to zero, producing a more parsimonious model⁸), bootstrapping, and adjusting for clustering, utilizing standardized scales to collect the information on the predictors, using an objective biomarker (viral load) as the outcome. However, our study also faces some limitations. First, we used self-reported data, for some variables such as ART adherence, which may be subject to respondent bias. However, this was minimized by relying on standardized measures, such as the three-item self-reported adherence measure. Secondly, due to the lack of a suitable dataset, we did not externally validate our model. Therefore, it is not known how well the current model would perform in different populations and settings.

4.1. Conclusion

In summary, our model offers a pragmatic tool for identifying ALHIV at elevated risk of virologic failure in resource-limited settings. By focusing on modifiable predictors like mental health, family support, and economic stability, the model aligns with global strategies for differentiated HIV service delivery. Scaling up such approaches could enhance the efficiency of HIV programs, ensuring that interventions reach those most in need while maximizing limited resources. Our study is among the first studies in SSA to develop a risk prediction model for VF in ALHIV. Researchers could use this model as a foundation for refining future models to achieve better precision. Future models could be improved by, incorporating clinically relevant predictors such as the HIV clinical stage and the CD4 count, which could potentially improve the overall discrimination ability of the model. Also, researchers should aim to produce simpler models containing fewer variables and collected using less extensive scales for easy implementation in busy routine care settings in LMICs. On the other hand, implementation research should evaluate the cost-effectiveness of integrating this model into routine HIV care pathways.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.puhe.2025.105753.

Data availability

The deidentified datasets included in the analysis for this study are available from the principal investigator (Prof. Fred Ssewamala) on reasonable request.