The research-utilized data from a publicly available demographic and health survey conducted in the country. Out of the 12,688 data points collected, 8,186 were excluded from the study as they pertained to individuals who were either younger than 15 or older than 24 years old. Following the exclusion of these data points, the analysis concentrated on a group of 4,502 adolescents who were included in the study.

In this study, we referred to the outcome variable as “tested,” which indicated whether the adolescent had undergone HIV testing within the past 5 years. If the adolescent had been tested during the survey period, we coded the variable as “Yes.” On the other hand, if the adolescent had not been tested for HIV within the past 5 years, we coded the variable as “No”.

In this specific study, we made the deliberate choice to include individuals aged 15–24 years old. We extracted from publically available data set of demographic and health survey of 2016.¹ Then we filtered the data of adolescent in this data set. Our data was sourced from a de-identified demographic and health survey database, encompassing information from a total of 4,502 adolescent individuals. To ensure the integrity of our data, we enlisted the expertise of two epidemiologists who collaborated on the project. Their valuable insights were instrumental in identifying and addressing any discrepancies such as noisy or abnormal values, errors, duplicates, and irrelevant data. Additionally, we meticulously reviewed the initial list of parameters to guarantee consistency in the data preprocessing stage. Ultimately, our analysis focused solely on the data pertaining to the 4,502 adolescents within the age range of 15–24 years old. To enhance clarity and coherence, we assigned appropriate labels to both nominal and continuous variables based on previous literature.

The random forest (RF) algorithm is known for significantly improving the classification accuracy of a model. This is achieved by generating multiple decision trees. Each decision tree produces a result for a given sample, and the final result is determined based on the majority of the decision trees’ results (34). To maximize the performance of the RF algorithm, we conducted hyperparameter tuning and 10-fold cross-validation for all our experiments. We focused on tuning the same set of hyperparameters. These hyperparameters include the maximum depth values of the decision trees and the minimum number of samples required in a leaf node. In order to assess the quality of each node’s split, we used two criteria: the Gini index and entropy. Since the RF algorithm generates multiple decision trees, we utilized the “number of estimators” hyperparameter to control the number of trees created in the forest. The values we experimented with for this hyperparameter were 100, 200, and 300. It’s worth noting that each decision tree in the RF algorithm learns from random subsets of samples drawn from the dataset. The use of bootstrap sampling, which involves drawing samples with replacement, ensures the diversity and robustness of the decision trees. Additionally, we tuned the “bootstrap” hyperparameter, which determines whether the entire dataset is used to build a decision tree. When set to “False,” the bootstrap parameter ensures that the entire dataset is used. On the other hand, setting it to “True” allows bootstrap sampling to take place. By carefully tuning these hyperparameters, we aimed to achieve the highest possible performance for the RF algorithm in our experiments.

In machine learning, dealing with imbalanced data can be a significant challenge. This occurs when the distribution of classes in a dataset is uneven, which can lead to biased results in favor of the majority class. In the current dataset being analyzed, there is a substantial imbalance between the “not tested” and “tested” classes, with 2,950 and 1,552 cases, respectively. This imbalance can lead to inaccurate results and make it likely for new observations to be categorized into the majority class. To address this issue, the study utilized a method called synthetic minority over-sampling technique (SMOTE) from the imbalanced-learn toolbox. SMOTE generates synthetic samples for the minority class by interpolating between existing minority class samples. By applying SMOTE, the dataset was balanced, allowing for more accurate and unbiased training of machine learning models. If you are interested in learning more about the imbalanced-learn toolbox and the SMOTE method, you can visit their website at https://imbalanced-learn.org/stable/.

Predictive accuracy is a commonly used metric to evaluate the performance of machine learning algorithms. However, when working with imbalanced datasets, accuracy can be misleading and hinder the identification of underlying causes such as HIV testing. In this particular study, researchers employed various techniques to address the class imbalance issue in their dataset. They utilized Synthetic Minority Oversampling Technique (SMOTE) (35), which generates new samples by interpolating between existing samples and their neighbors (36, 37). Additionally, they employed random under-sampling, which involves discarding samples from the majority class until the minority class reaches a predetermined percentage of the majority class (35). Another method used was Adaptive Synthetic (ADASYN), which generates synthetic data for harder-to-learn minority class samples, thereby reducing bias introduced by imbalanced data distribution. Through the application of these techniques, the researchers successfully achieved a balanced dataset (38). The outcome of their endeavors is discussed in detail in the result section of the study (39).

In the initial phase of our study, our primary aim was to identify key features that could accurately predict the HIV testing status of adolescents in Ethiopia. To achieve this, we conducted a thorough review of scientific literature by searching various databases. The findings from this review were then used to create a comprehensive questionnaire, which covered a wide range of predictors for HIV testing services. By incorporating these identified predictors, we aimed to develop a reliable tool for predicting HIV testing status. The SHAP analysis was done for the features. SHAP is a powerful technique that is independent of any specific machine learning model. It is used to calculate the Shapley values of the different features in a model, providing explanations for the model’s predictions. SHAP feature importance is determined by taking the average of the absolute values of the Shapley values for each feature. This approach allows for a comprehensive understanding of the impact each feature has on the model’s predictions.

During this stage, our main objective was to identify specific features that could accurately predict HIV testing status. We began by conducting a comprehensive review of scientific databases to determine the most relevant features. Based on this review, we developed a questionnaire that included predictors. We carefully compiled all the pertinent information to create the final version of our data collection tool. By doing so, we aimed to ensure that our study would yield accurate and reliable results that could help inform future efforts to improve HIV testing services in Ethiopia. WEKA version 3.9 was used to select important features, R version 4.0.2 and python version 3.2 was used for the analysis of data. Figure 1 displays a flowchart that outlines the process used to select the final variables for a machine-learning model. This process involved five distinct steps. The first step involved removing features from the dataset that had a missing value greater than 30%. In the second step, features that were deemed irrelevant to the final outcome variable, such as reference date, patient ID, and accompanying information, were eliminated. The third step addressed collinearity, which can lead to duplicated features and skew the model’s results. Features with a collinearity greater than 0.95 were removed from the dataset. By implementing these procedures, the most relevant and informative features for the machine learning model were identified, as shown in Figure 1. In this study, various feature selection methods were employed to identify the most relevant predictive features. These methods included recursive feature elimination (RFE), random forest feature importance, and the Boruta feature selection method. RFE is a technique used for feature selection, which begins with all the features in the training dataset and iteratively eliminates features until the desired number of features is reached. This method is particularly effective in reducing model complexity and improving the efficiency of machine learning algorithms. By utilizing these feature selection methods, the study aimed to identify the most informative features that contribute significantly to the predictive power of the model (40). This approach helps to streamline the data processing and enhance the accuracy of machine learning algorithms.

A comprehensive literature review was conducted to develop accurate predictive classifier models for HIV testing status (19, 33, 41–45). The selection of suitable machine learning (ML) algorithms was based on the type and quality of the dataset utilized. Eight ML algorithms were employed to construct the individual testing prediction model: J48 decision tree, random forest (RF), k-nearest neighborhood (k-NN), Support Vector Machine (SVM), multi-layer perceptron (MLP), Naïve Bayes (NB), logistic gradient boosting (logit Boost), and logistic regression (LR). The data were analyzed using Weka software v3.9.2 python, and R software to implement the algorithms, analyze and calculate curves and criteria, and draw the confusion matrix.

We used the EXPLORER module in WEKA to find the best hyperparameters for our models and evaluated their performance using tenfold cross-validation. We ran experiments with WEKA’s EXPERIMENTER module and repeated the cross-validation process to ensure reliable results. We used an 80:20 ratio for training and testing, and calculated average performance metrics across ten runs. We choose stratified tenfold cross-validation to accurately estimate accuracy. Our approach aims to minimize errors and biases by using more data for training and validation. We used WEKA’s EXPLORER and EXPERIMENTER modules along with tenfold cross-validation for a robust evaluation and comparison of classification models. We selected stratified tenfold cross-validation as it strikes a favorable balance between bias and variance, making it a preferred technique for accurately estimating accuracy. It is worth emphasizing that tenfold cross-validation is widely employed in the fields of machine learning and data mining due to its advantages over traditional instance splitting methods. This provides a reliable and robust method for evaluating and comparing the effectiveness of classification models. The summary of machine learning pipeline presented on Figure 2.

Evaluating the performance of a machine-learning model is essential for its success. In our study, we thoroughly assessed the performance of our predictive models using various performance metrics, as outlined in Table 1. These metrics encompassed accuracy, specificity, precision, sensitivity, and the receiver operating characteristic (ROC) chart criteria. By leveraging these metrics, we effectively measured the effectiveness of our models in predicting HIV testing services of adolescents. To determine the best model for predicting HIV testing service of adolescent, we compared the performance of each model using the aforementioned evaluation criteria. The results of this comparison are summarized in Table 2. Through a meticulous analysis and comparison of these evaluation criteria, we successfully identified the model that demonstrated the highest performance in predicting HIV testing. Our comprehensive evaluation process enabled us to select the most effective model and gain valuable insights and confidence in its predictive capabilities.

The technique of association rule mining examines the correlations between multiple variables in a group, and it was first developed by Agarwal and Srikanth (46). In this study, an additional method was utilized to support the classification of machine learning algorithms for predicting adolescent HIV testing using R software. The apriori algorithm was employed to uncover associations between the selected features and the target feature (47). A minimal support degree of 0.00095 and a minimum confidence threshold of 90% were set to identify all potential association rules. This is because a rule is considered reliable if its confidence level is more than 80% (48). In this study, the main focus was on identifying features that are associated with adolescent HIV testing through the use of association rules. Specifically, the study utilized a technique known as classification association rules (49), which involves analyzing the features that are implied by the target features (Antecedent = > Consequent). The goal of this approach was to classify all the variables that contribute to HIV testing among adolescents and to identify the predictors that each category contributes to the testing. To evaluate the strength of each rule, the study used the metrics of Support, Confidence, and Lift. It is important to note that in this context, the sets of features represented by X and Y are mutually exclusive.

Rule X = > Y:

Support =  Fequence ( X, Y ) N , Confidence =  Frequency ( X,Y ) Fequency ( X ) , Lift =  Frequency ( X,Y ) Fequency ( X ) Frequency ( Y )

After applying our exclusion criteria and conducting a quantitative analysis of case records, we have identified a total of 4,502 adolescents in the database who met the eligibility criteria for our study. Among these participants, 3,339 (74.2%) adolescents were male, while 1,163 (25.8%) adolescents were female. The mean age of the study participants was 19.09 (±2.841) years old. A majority of the study participants, 3,137 (69.7%), reported residing in rural areas. Additionally, 55.2% of the total 4,502 study participants had attended primary education. An overwhelming majority, 4,292 (95.3%) of the adolescents, had previously heard about AIDS. Out of the total study participants, 79.4% of them knew the place for HIV testing. Regarding awareness of sexually transmitted infections (STIs), the majority (95.7%) of the adolescents were aware of STIs (Table 3).

In this particular study, various pre-processing steps were undertaken to handle missing or null values, encode categorical labels, and balance the dataset. Initially, the incomplete raw data and missing values were addressed using the imputation technique. Supplementary Figure S1 provides a breakdown of the percentage of missing values for each feature, which shows that the feature know the place of HIV testing 210 (4.66%), alcohol drinking 16 (0.3554%), and stigma 210 (4.66%) having the highest and the only percentage missing data in adolescent HIV testing data set. These missing values were imputed using the simple imputation technique. To prepare the data for the model, it was necessary to convert input and output features into numerical values. This was accomplished by utilizing the one-hot encoder to encode categorical variables found in the dataset. The dataset consisted of thirty one features with a single target feature. One-hot encoding is a valuable encoding technique for classification tasks. This technique transformed each categorical value into a new column with one-hot encoding, and the label values were created as new columns with values of either 1 or 0.

The current study aimed to address the issue of data imbalance and enhance the effectiveness of machine learning algorithms. Various techniques were employed to balance the dataset. The outcome feature revealed that the majority of observations (65.5%) were classified as not tested for HIV, while a smaller portion (34.5%) were tested. To balance the dataset, under-resampling, SMOTE, and ADASYN techniques were utilized, each employing a different approach to either maximize the minority class or decrease the majority class. The performance of each balancing technique was compared using selected supervised classification machine learning algorithms, with a focus on accuracy and AUC. In the unbalanced dataset, the random forest algorithm achieved a higher AUC of 84.9% compared to other classifiers, while the J48 decision tree had a higher accuracy of 82.4%. In the ADASYN approach, the J48 decision tree classifier achieved a higher accuracy of 83.1%, while logistic regression achieved a higher AUC of 79.9%. When comparing different balanced sampling methods using a J48 decision tree classifier, the SMOTE sampling technique performed the best, with an accuracy of 89.3% and an AUC of 86.3%. ADASYN was the second most effective method, with an accuracy and AUC of 83.1 and 76.9%, respectively. However, the under-sampling technique was found to be the least effective, with an AUC of 73.4% and an accuracy of 78.2% (Table 4).

Overall, the study demonstrated that employing balanced sampling methods can significantly improve the performance of machine learning algorithms on imbalanced datasets. The results suggest that SMOTE and ADASYN are effective techniques for balancing imbalanced data, while under-sampling may not be the most effective approach. The study in question faced a significant issue of imbalanced data, which could potentially hinder the performance of the classifying algorithm. To tackle this problem, a balanced sampling method was deemed crucial. Unbalanced data poses a challenge to machine learning, as values from the minority class or rarely occurring classes may be mistakenly classified as instances of the majority class. To address this issue, the study applied SMOTE to the unbalanced dataset, resulting in an increase in the total number of records (Figure 3). The classifier and the balanced sampling method were compared with other classifiers using training accuracy and AUC. Therefore, the study primarily relied on AUC to compare the classifier and the balanced sampling method. Overall, the use of a balanced sampling method proved to be effective in overcoming the issue of imbalanced data, and the study’s findings highlight the importance of considering AUC when evaluating the performance of classification algorithms on imbalanced datasets.

In this experiment, the objective was to evaluate the effectiveness of different classifiers in predicting the HIV testing status of adolescents. The primary goal of the analysis was to assess the accuracy of the predictions made by the chosen classifier. Among the classifiers that were selected, the random forest classifier exhibited particularly robust performance on the balanced dataset compared to the others. Once the best model was identified, hyper-parameter tuning and feature selection were performed. Additionally, the significance of the predictor for viral failure was determined to further evaluate the model’s performance.

This study aimed to assess the effectiveness of the J48 decision tree classifier in predicting HIV testing among adolescents and compare it with two other feature selection methods: Boruta feature selection and recursive feature elimination. The findings, as depicted in Figure 4, revealed that the J48 decision tree feature importance method outperformed both Boruta feature selection and recursive feature elimination. It achieved impressive results, including a sensitivity of 81.3%, specificity of 80.9%, precision of 81.0%, f1-score of 81.14%, and an AUC of 0.863. These outcomes strongly suggest that the J48 decision tree feature importance method is the most suitable approach for accurately predicting adolescent HIV testing. Conversely, Boruta feature selection and recursive feature elimination were found to be less effective in this context. Consequently, the J48 decision tree classifier was selected as the feature selection method for this study. Among all the features examined, the most significant predictors of HIV testing were age, knowledge of the place for HIV testing, age at first sexual encounter, recent sexual activity, exposure to family planning, and the number of sexual partners (see Table 5).

After carefully selecting the most suitable classifier, this study proceeded to apply hyperparameter tuning or optimization to compare it with the default hyperparameter settings. Figure 5 visually represents the performance of the default hyperparameter configuration against the hyperparameter tuning approach when using the random forest classifier. The results indicated that the default hyperparameter configuration performed better than the tuned hyperparameters. Based on these findings, it was concluded that the random forest classifier with default hyperparameters outperformed the one with tuned hyperparameters. Consequently, the study decided to utilize the random forest classifier with default hyperparameter settings. The J48 decision tree classifier, after undergoing hyperparameter tuning, achieved impressive metrics, including a precision of 88.6%, an f1-score of 91.1%, a sensitivity of 89.3%, and a specificity of 80.9%. It is worth noting that the J48 decision tree classifier with hyperparameter tuning outperformed all other classifiers in terms of the area under the curve (AUC). In summary, the study ultimately presented the default hyperparameter tuning and hyperparameter tuned shown in Figure 5.

In our analysis of the dataset, we utilized the Pearson Correlation Coefficient (PCC) as another feature selection method. PCC is a statistical measure that helps determine the correlation between two random attributes. The correlation value, represented by “r,” ranges from −1 to +1. By applying PCC to measure the correlation between the attributes and the target variable, which in this case is “adolescent HIV testing,” we were able to identify which features were positively or negatively correlated with the target class. The results of this analysis are presented in Table 6, which lists the correlation values of each feature with respect to the target variable. The feature with the highest correlation value is “know the place of HIV testing,” with a value of 0.337. Additionally, “age” exhibits a correlation of 0.317, while “age at first sex” has a correlation of 0.306 (Table 6). A positive correlation indicates that the variables are positively associated, meaning that as the value of x increases, the value of y also increases, and vice versa.

After studying the correlation of the predictors to the target, we also took into consideration the collinearity. Collinearity happens when two predictors are linearly associated or having a high correlation to each other, and both were used as predictors of the target variable (21). Multicollinearity may also happen, which is a situation wherein the variable has collinearity with more than one predictors in the dataset. We used the Variance Inflation Factor (VIF) to detect the collinearity of the predictors in the dataset. The VIF starts from 1 to infinity, and the value of 1 means that the features were not correlated. VIF values less than 5 are moderately correlated, while VIF values of 10 and above are highly correlated and a cause of concern (21). The VIF values of each predictor in the dataset can be seen in Table 7.

Table 7 presents the Variance Inflation Factor (VIF) for each predictor in the dataset. The highest VIF value is observed for the variable “total children ever born,” which stands at 25.109. Other predictors such as “working in the last 12 months,” “current marital status,” “recent sexual activity,” and “residence” also exhibit relatively high VIF scores, although they are lower than 5. A VIF between 1 and 5 suggests that these predictors are not strongly correlated and can be considered for inclusion in the adolescent HIV testing model. Therefore, it is recommended to include “working in the last 12 months,” “current marital status,” “recent sexual activity,” and “residence” as predictors when building the adolescent HIV testing model. In addition, we used SHAP plot of feature selection. Based on SHAP plot age, wealth, sexual activity, educational status, exposure to family planning, IPV, know the place of HIV testing and awareness on STI and AIDS were identified (Figure 6).

A comprehensive literature review has thoroughly examined 31 features that contribute to HIV testing services for adolescents. Through a feature evaluator, the significance of these factors was assessed, resulting in the identification of 20 highly important variables during the feature selection process. However, only 20 features were included in the analysis, while others were excluded based on specific criteria outlined in Figure 1.

To predict the HIV testing status of adolescents, the significance of each factor was calculated, leading to the selection of 20 predictors for machine learning (ML) algorithms. Among these predictors, age emerged as the most important factor for HIV testing services, with a value of 0.31733. On the other hand, the total number of children born was found to be the least important predictor, with a value of 0.08367. The importance of each feature in the dataset was calculated and presented in Table 5, displaying the variables in descending order of ranking.

In this study, our objective was to predict HIV testing among adolescents by selecting the most optimal features and utilizing eight different machine learning (ML) algorithms. These algorithms included J48, RF, LR, MLP, logit Boost, k-NN, SVM, and NB. To evaluate the performance of each algorithm, we conducted 10-fold cross-validation with a seed value of two and assessed various metrics, such as sensitivity, specificity, accuracy, precision, and the receiver operating characteristic (ROC) curve. The results of the cross-validation are presented in Table 8. Our experimental findings revealed that the J48 decision tree algorithm outperformed the other ML algorithms in accurately predicting adolescent HIV testing. It achieved impressive performance metrics, including a sensitivity of 81.30%, specificity of 80.90%, accuracy of 81.3%, precision of 81.0%, and an ROC value of 86.3%. Figure 7 visually depicts the performance metrics of the ML algorithms used in this study, while Figure 8 presents a comparison of the area under the ROC curve for these algorithms. Notably, the SVM algorithm exhibited the lowest performance with an ROC value of 66.8% according to the ROC analysis. Figure 9 presents the false positive rate and true positive rate of each algorithm. For a comprehensive summary of the performance evaluation of each algorithm, please refer to Table 8.

This study utilized the J48 decision tree feature importance method to select relevant features. Subsequently, the association mining rules were applied using the apriori algorithm for interpretation and a comparison of the best-selected features. From the association mining rules, a total of nine rules were identified with a confidence value of over 90% and the highest lift or interestingness. Among these rules, the twenty most significant ones were chosen for predicting adolescent HIV testing. The absolute minimum support count of the apriori algorithm was 2,251 in support of 0.5 and confidence of 0.9. The summary of quality measures of association rule mining analysis indicated that minimum support value = 0.5018, confidence = 0.9461, lift = 0.9924. The total rules of this apriori algorithm of this study was 712,602 done by 48 s.

Rule 1 indicated that individuals know the place of HIV testing (yes), intimate partner violence (=no), had media exposure (=yes), number of sexual partner (=more than one), awareness on AIDS (=yes), awareness on STI (=yes), sexual activity (=active in last 4 weeks), age at first sex (=between16 to 24 years), marital status (=married), knowledge on contraceptive (=knowledgeable), number of children (=one and above), knowledge on ovulatory cycle (=yes) and exposure to family planning (=yes) then the possibility of HIV testing was 95.33% of confidence with support value = 95% and Lift value = 31.444.

Rule 2 means that being rural residence, awareness on AIDS (=yes), awareness on STI (=yes), sexual activity (=active in last 4 weeks), age at first sex (=between16 to 24 years), marital status (=married), knowledge on contraceptive (=knowledgeable), number of children (=one and above), knowledge on ovulatory cycle (=yes) and literacy (able to read and write) the adolescent will be 92.24% chance of HIV testing with support = 93% and Lift = 33.334.

Rule 3 means that if age greater than 18 years old, education (=secondary), age at first sex (16–24 years), recent sexual activity (active in last 4 weeks), exposure to family planning (=yes), and had media exposure the probability of adolescent HIV testing is 100% of confidence with support of 0.00255, Lift = 31.233%.

Rule 4 showed that if the adolescent attended secondary education, age above 18 years, intimate partner violence (=no), had media exposure, number of sexual partner (more than one), had awareness on AIDS, and had awareness on STI a then the possibility of the adolescent for testing HIV is 100% of confidence, support = 0.00222, Lift = 33.377.

Rule 5 means that if the adolescent attended secondary education, know the place of HIV testing, age at first sex (16–24), urban residence, and had knowledge on contraceptives then the possibility of adolescent HIV testing is 100% confidence and support = 0.001033, Lift = 32.64.

According to the association rule analysis, it was found that several factors strongly predict adolescent HIV testing in Ethiopia. These factors include knowing the place of HIV testing, being above 18 years old, having a secondary education or higher, engaging in first sexual activity above the age of 16, living in urban areas, having recent sexual activity, being exposed to family planning information, having more than one sexual partner, being aware of STIs and AIDS, and having media exposure. The association rules indicate a strong relationship between these independent features and the dependent feature (adolescent HIV testing). This is supported by the lift values, which are greater than one. This suggests that the presence of these factors increases the likelihood of adolescents getting tested for HIV in Ethiopia. In summary, the association rule analysis highlights the significant predictors of adolescent HIV testing in Ethiopia, emphasizing the interconnections between various factors and the importance of these factors in promoting HIV testing among adolescents.

This retrospective study analyzed demographic and health survey data that exhibited irregularities and imbalances. To address this issue, the researchers took measures to balance the dataset by removing noisy and inadequate records. Specifically, they focused on tackling the problem of imbalanced classes, where the number of records related to the HIV-tested class was significantly lower than the tested or never tested (1,552 vs. 2,950).

To evaluate the performance of each machine learning algorithm, various criteria were utilized. Furthermore, external validation of the proposed model was conducted using multi-center country-level data, aiming to enhance the generalizability of the predictions. However, it is important to note that the researchers relied on self-reported data from the demographic and health survey, which may introduce inconclusiveness and potentially impact the training data.

For future research, the researchers recommend that behavioral features be incorporated to further enhance prediction accuracy. Additionally, it is crucial to monitor the dynamic variations of significant features over time to better identify adolescents HIV testing in the countries. While the study provides valuable insights using the available data, there are areas for improvement and avenues for further research to enhance the understanding and prediction of HIV testing services in this specific sub-population group in Ethiopia.