Page 1827
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
A Hybrid Machine Learning Framework for Early Detection of Hiv-
Associated Opportunistic Infections
Udoinyang G Inyang
1
, Abraham C Inyang
1
, Asukwo E Onukak
3
, Emmanuel A Ubong
2
1
Department of Data Science, Faculty of Computing, University of Uyo, Nigeria
2
School of Computing and Information Technology, Federal University of Technology, Ikot Abasi.
Nigeria.
3
Department of Internal Medicine, University of Uyo, Nigeria
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150600129
Received: 26 June 2026; Accepted: 01 July 2026; Published: 17 July 2026
ABSTRACT
The increasing burden of OIs among people living with HIV necessitates accurate and timely predictive systems
to support clinical decision-making. The framework leverages clinical, demographic, and treatment-related
variables extracted from a real-world dataset obtained from the University of Uyo Teaching Hospital. Data
preprocessing techniques, including normalization, feature engineering, and dimensionality reduction were
applied to enhance model performance. The proposed model was evaluated using multiple performance metrics,
including accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix analysis. Experimental results
show that the best-performing configuration achieved an accuracy of 96.79% with strong sensitivity and
specificity, demonstrating its effectiveness in distinguishing between infected and non-infected cases.
Furthermore, the model demonstrated the use of the ReLU-based activation function. SHAP analysis further
enhanced model interpretability by identifying key predictive features such as CD4 count, treatment type, and
age. Overall, the results confirm that the proposed hybrid framework is a reliable and clinically meaningful tool
for early detection and risk stratification of HIV-associated opportunistic infections.
Keywords: HIV, Infection, Machine Learning, Opportunistic Infections, DNN
INTRODUCTION
Opportunistic infections (OIs) remain a leading cause of morbidity and mortality among people living with HIV,
making early detection of infection risk and disease onset a critical clinical priority (Xu et al., 2022; Zhou et al.,
2019). The progression of HIV-related complications and OI development is often rapid and difficult to predict
using conventional clinical assessment alone, creating a strong need for effective risk-stratification frameworks
that can identify high-risk patients early and support timely clinical intervention (Fu et al., 2025). Recent studies
have shown that Artificial Intelligence (AI) and Machine Learning (ML) techniques can leverage key HIV-
related indicators such as viral load, CD4 count, and co-infection status to predict disease progression and
forecast opportunistic infections including tuberculosis and pneumonia, thereby enabling more proactive,
preventive, and personalized care (Sah et al., 2025).
The adoption of ML-based approaches is particularly valuable because they can process large, complex clinical
datasets, integrate heterogeneous data sources, and uncover hidden patterns without strict statistical assumptions
(Fu et al., 2025; Waring et al., 2020). In particular, Deep Neural Networks (DNNs) require careful optimization
of parameters such as activation functions, hidden layers, learning rates, and batch sizes to effectively learn
meaningful patterns from data (Inyang et al., 2025). Despite these advantages, diagnosing OIs in HIV/AIDS
patients remains challenging due to limitations in conventional approaches such as culture-based methods,
targeted molecular assays, and invasive diagnostic procedures. Although emerging techniques like next-
generation sequencing (NGS) and metagenomic sequencing provide faster and more comprehensive pathogen
Page 1828
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
detection, the most effective diagnostic strategy combines these advanced tools with traditional microbiological
methods rather than relying on either alone (Zhou et al., 2019; Xu et al., 2022).
Therefore, an effective early-detection framework for HIV-associated opportunistic infections should not only
identify patients at risk of rapid clinical deterioration but also support the timely identification of likely infectious
causes, thereby improving triage, treatment selection, and healthcare resource allocation (Fu et al., 2025; Sun et
al., 2025). In line with this objective, this study proposes a hybrid modeling approach that integrates DNNs and
XGBoost for predictive analytics. While DNNs are effective in capturing complex nonlinear relationships within
high-dimensional clinical data, XGBoost enhances performance through efficient feature selection and robust
classification of structured datasets, leading to improved overall prediction accuracy for opportunistic infections
in HIV patients. The study utilizes the Opportunistic Infections dataset obtained from the University of Uyo
Teaching Hospital for the early detection and monitoring of HIV-induced opportunistic infections (OIs).
Related Works
Hybrid Framework for Early Detection of HIV-Associated Opportunistic Infections
A hybrid framework for the early detection of HIV-associated opportunistic infections (OIs) can be structured
into four interconnected components: data preparation, multimodal prediction, rule-guided clinical reasoning,
and explainable decision support. The data preparation stage integrates structured clinical variables, including
CD4 count, viral load, symptoms, treatment history, and co-infection indicators, with longitudinal electronic
health record (EHR) data. This approach is supported by hybrid deep learning models that combine Multilayer
Perceptrons (MLPs) and Long Short-Term Memory (LSTM) networks, enabling the effective extraction of both
static and temporal patient information for improved disease prediction and monitoring of evolving OI risk (Chen
et al., 2024). To enhance clinical applicability, a rule-aware reasoning layer can be incorporated using fuzzy
rule-based systems. Such systems combine Type-2 Fuzzy Logic with deep learning techniques to provide
transparent decision rules for early referral and treatment intervention, particularly when signs of treatment
failure emerge (Ekpenyong et al., 2019).
At the predictive level, the framework benefits from an ensemble-based architecture rather than relying on a
single learning model. Multiple machine learning algorithms can be combined to specialize in different data
patterns, with their outputs aggregated into a unified risk score to improve predictive accuracy and robustness
(Qasem, 2024). This approach can be further strengthened through feature selection and data-balancing
techniques, which address the high dimensionality and class imbalance commonly found in HIV and OI datasets
(Sa'adah et al., 2026). Additionally, the framework may incorporate molecular data sources, such as
transcriptomic and immune-cell-specific biomarkers, to support biomarker discovery and improve prediction in
complex clinical cases (Mahmud et al., 2025). Consequently, an effective early-detection framework should
integrate clinical and temporal data analysis, ensemble learning, fuzzy reasoning, and interpretable artificial
intelligence to support accurate risk prediction, timely intervention, and improved management of HIV-
associated opportunistic infections (Ekpenyong et al., 2019; Chen et al., 2024; Qasem, 2024; Mahmud et al.,
2025; Sa'adah et al., 2026).
Machine Learning Tasks for HIV-Associated Opportunistic Infection Detection and Prediction
Machine learning (ML) has been applied to a wide range of tasks aimed at improving the detection, diagnosis,
and prediction of opportunistic infections (OIs) among people living with HIV. One important application is
broad-spectrum OI screening, where models are designed to determine whether an HIV-positive patient is likely
to have any opportunistic infection rather than focusing on a specific pathogen. Such approaches support early
triage and reduce the likelihood of missed diagnoses by identifying high-risk patients using a small number of
clinically relevant features (Chen et al., 2025).Beyond general screening, ML techniques have also been used
for pathogen-specific diagnosis. For example, predictive models have been developed for Pneumocystis jirovecii
pneumonia (PJP), demonstrating that the integration of routine clinical data with microbiological indicators such
as P. jirovecii PCR results can improve diagnostic sensitivity while maintaining acceptable specificity (Chew et
al., 2024). Similarly, ML models incorporating inflammatory biomarkers have been proposed for the early
detection of nontuberculous mycobacterial (NTM) disease in HIV-positive individuals, enabling the
Page 1829
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
identification of less common but clinically significant infections through interpretable prediction frameworks
(Li et al., 2025).In addition to diagnostic applications, ML has shown considerable potential for prognostic
prediction. Several studies have focused on forecasting the future occurrence of active tuberculosis (TB) among
HIV patients using routinely collected clinical data, thereby providing opportunities for preventive intervention
before disease onset (Bartl et al., 2025; Ngiam et al., 2019; Scott, 2021). ML has also been employed to model
complex interactions between coexisting diseases such as TB and HIV, facilitating a deeper understanding of
how multiple clinical conditions jointly influence patient outcomes (Cherezov et al., 2025). Furthermore,
predictive models have been developed to estimate the risk of antimicrobial-resistant bacterial colonization or
infection among people living with HIV, supporting treatment planning and infection-control strategies
(Henderson et al., 2022).
Emerging research has extended these applications to microbiome-based classification, where gut microbial
signatures are used as predictive markers of AIDS progression and immune decline. Such approaches
demonstrate the potential of integrating molecular and biological data into HIV-related prediction systems (Zhou
et al., 2023). Collectively, these studies indicate that an effective hybrid framework for HIV-associated OI
management should support multiple interconnected tasks, including broad OI screening, pathogen-specific
diagnosis, future disease-risk prediction, and co-morbidity or antimicrobial-resistance assessment. This diversity
of predictive objectives highlights the need for a hybrid architecture capable of integrating multiple data sources
and learning models rather than relying on a single-purpose classifier (Chen et al., 2025; Chew et al., 2024; Li
et al., 2025; Bartl et al., 2025; Henderson et al., 2022; Cherezov et al., 2025).
Interpretability and Clinical Decision Support in HIV-Associated Opportunistic Infection Detection
Interpretability is essential in the early detection of HIV-associated opportunistic infections (OIs) because model
predictions often guide clinical decisions such as diagnostic testing, referral, treatment modification, and patient
monitoring. To enhance transparency and clinician trust, hybrid frameworks increasingly combine machine
learning with rule-based reasoning. For example, integrating deep learning with Type-2 fuzzy rule systems
enables proactive detection of treatment failure while providing clear decision thresholds for intervention
(Ekpenyong et al., 2019). Explainable AI techniques, such as SHAP, further improve model transparency by
identifying the key factors influencing prediction outcomes, including immune suppression, inflammatory
markers, symptoms, and co-infection indicators. Such approaches have been successfully applied in HIV-
associated NTM prediction, allowing clinicians to better understand and validate model recommendations (Li et
al., 2025). Similarly, lightweight screening models can support rapid identification of patients at risk of any OI
and trigger further investigation when necessary (Chen et al., 2025). When combined with interpretable
molecular biomarkers derived from transcriptomic data, these approaches create a multi-layered decision-
support framework that improves early detection, supports targeted diagnosis, and facilitates timely, evidence-
based clinical intervention (Mahmud et al., 2025; Chen et al., 2025; Li et al., 2025).
METHODOLOGICAL WORKFLOW
The architecture comprises a structured collection of essential clinical, demographic, and treatment-related
variables used to monitor the health status, treatment outcomes, and comorbid conditions of individuals living
with HIV/AIDS, as presented in Table 1. It is designed to serve as a comprehensive reference framework for
healthcare professionals managing patients undergoing antiretroviral therapy (ART). Each variable is assigned
a unique code, data type, and descriptive definition to clearly indicate its clinical significance. The parameters
include baseline demographic and identification information, such as Hospital Number, Age, and Sex, as well
as key immunological indicators, including CD4 Count and Viral Load, which are critical for assessing immune
status and evaluating treatment response. Treatment-related variables, such as the initial ART regimen, current
ART regimen, and treatment failure records, are incorporated to monitor regimen modifications and identify
potential drug resistance. In addition, clinical monitoring indicators, including WHO clinical stage, adherence
level, and duration on ART, provide valuable insights into disease progression, treatment effectiveness, and
patient compliance.
Page 1830
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 1: Architecture for Analytics Hybrid Modelling for Early Detection and Monitoring of HIV
Induced Opportunistic Infections
Mathematical Formulation of Deep Neural Networks for HIV_OIs Prediction
DNNs are multilayered models capable of learning hierarchical representations from complex clinical and
biological datasets. In the context of HIV_OIs prediction, the DNN processes patient features such as CD4 count,
viral load, treatment regimen, comorbidities, and demographic variables to detect patterns indicative of
opportunistic infections. A typical DNN consists of an input layer, several hidden layers, and an output layer.
Each layer transforms its input using a weighted sum followed by a non-linear activation function. The
mathematical formulation is as shown in Equations 1 to 7:
1. Input Layer:
(1)
where
represents a normalized feature of an HIV patient (e.g., CD4 count, duration of ART, weight loss,
fever, etc.), and is the number of features.
2. Hidden Layers:
For the -th layer, the transformation is given by Equation (2):
(2)
where:
is the weight matrix for layer ,
Page 1831
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
is the bias vector,
is the activation from the previous layer (or input if ),
is the activation function, typically ReLU:
3. Output Layer:
The final output (probability of HIV_OI occurrence) is shown in Equation (3) and (4):
(3)
(4)
where:
is the index of the output layer,
is the sigmoid function for binary classification:This yield
, representing the predicted
probability of an HIV-related opportunistic infection being present.
4. Loss Function (Binary Cross-Entropy):
The loss function quantifies the error between predicted and actual outcomes which is shown in Equation
(5):
(5)
where:
is the true label (1 for HIV_OI, 0 for no infection); is the predicted probability from the network.
5. Optimization (Gradient Descent with Backpropagation):
The parameters
and
are updated to minimize Equation (6):
(6)
where: represents all trainable parameters; is the learning rate;
is the gradient of the loss with respect
to the parameters.
6. Prediction Decision Rule:
After training, predictions are made using a thresholdEquation (7):
Predict HIV_OI
if
if
(7)
Mathematical Formulation of XGBoost for HIV-OIS Prediction
XGBoost (Extreme Gradient Boosting) is a scalable and efficient implementation of gradient boosting decision
trees. Its goal is to minimize the loss function through additive learning of decision trees, making it suitable for
medical prediction tasks such as HIV-related opportunistic infection (HIV-OIS) monitoring, as shown in
Equations (8) – (13).
1. Objective Function
The objective of XGBoost is to minimize a regularized loss:
Page 1832
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
(8)
Where:
is the actual HIV_OIS status for patient
is the predicted value at iteration
is a differentiable convex loss function (e.g., logistic loss for classification)
is the -th decision tree
(9)
Equation (9)is the regularization term that penalizes complexity:
Where:
: number of leaves; :vector of scores on leaves; , :regularization parameters
2. Additive Training Process
At each boosting round , a new function
is added to minimize the loss shown in Equation (10):
(10)
Using second-order Taylor expansion for efficiency, the approximate objective becomes in Equation (11):
(11)
Where:
: first-order gradient
: second-order gradient
Optimal Leaf Weights
For a given tree structure with leaves, the optimal weight
for leaf is:
(12)
The optimal objective function reduction as shown in Equation (13):
opt
(13)
Page 1833
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Experimentation and Analysis
The experimentation and analysis were conducted in three phases: Data Collection and Statistical Distribution
Analysis, Modeling Phase.
Data Collection
The study utilized the Opportunistic Infectionsdataset obtained from the University of Uyo Teaching Hospital
for the early detection and monitoring of HIV-induced opportunistic infections (OIs). The dataset contains 3,982
patient records and 88 variables, comprising 35 categorical features, 45 numerical features, and 8 text-based
attributes. The variables capture a wide range of clinical, demographic, and treatment-related information,
including CD4 count, viral load, body mass index, treatment regimens, and patient adherence indicators. This
diverse combination of quantitative and qualitative data provides a comprehensive representation of patient
health status, supporting both statistical analysis and machine learning applications. Consequently, the dataset
serves as a robust foundation for developing predictive models aimed at improving the early detection,
monitoring, and management of HIV-induced opportunistic infections.
Statistical Distribution
The kurtosis values further reveal the nature of data distribution in terms of peakedness and the presence of
extreme observations as in Table 1. Variables such as Current Weight (44.67) and Repeat Viral Load Result
(20.68) display very high kurtosis, indicating heavy-tailed distributions with potential outliers, which may reflect
patients with unusual body weights or viral load fluctuations. Conversely, Months of ARV Refill (-0.85) and
Number of Fingers Captured (-0.19) exhibit negative kurtosis, indicating flatter distributions with fewer extreme
cases. The Age variable (0.53) demonstrates near-normal kurtosis, confirming a relatively even spread of age
data. However, the Number of Fingers Recaptured shows a strong negative skewness (-4.54) with a high kurtosis
(23.82), implying that most entries cluster around higher values with a few extreme low cases. These results
suggest that while most variables show moderate variability, others, particularly weight and viral load, require
normalization and outlier treatment to ensure accurate performance of the Hybrid predictive models.
Table 1: Skewness and Kurtosis of Patient Data Variables
Variable
Skewness
Kurtosis
Current Weight (kg)
2.359874
44.665707
Age
-0.236635
0.526328
Months of ARV Refill
0.804408
-0.852294
Number of EAC Sessions Completed
1.169557
1.852281
Repeat Viral Load Result (c/ml) – POST EAC
4.461786
20.678513
Number of Fingers Captured
-1.342856
-0.186446
Number of Fingers Recaptured
-4.543142
23.820179
Pairwise relationships between multiple features such as Age_Clean, CD4_Numerical, BMI_Clean,
Viral_Load_Clean, and Creatinine_Clean against the binary OI status as shown in Figure 2. Scatter plots reveal
the degree of separation between OI and non-OI groups, with CD4_Numerical showing the clearest distinction
indicating that lower CD4 counts correlate strongly with OI presence. The diagonal histograms depict individual
feature distributions, helping identify skewed variables and feature overlaps. This visualization is instrumental
in feature interaction analysis, supporting model feature selection for improving OI prediction accuracy.
Page 1834
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 2: Pair Plot Feature for OI
Data Preprocessing
The Data Preprocessing stage, as shown in Table 2, involved applying Principal Component Analysis (PCA) to
identify the most influential features contributing to variability within the dataset used for early detection and
monitoring of HIV-induced opportunistic infections (OI). The analysis revealed four key principal components
(PCs) with corresponding eigenvalues and explained variance ratios. The first principal component (PC1) was
dominated by Current Weight, with an eigenvalue of 1.3208 and an explained variance ratio of 0.1887, indicating
that this feature alone accounts for approximately 18.9% of the total data variance. This suggests that variations
in patient weight play a major role in influencing health outcomes related to OI risk. The second component
(PC2) was associated with Age, which had an eigenvalue of 1.1793 and contributed 16.9% of the variance,
highlighting age as a significant factor affecting viral suppression and immune response. The third component
(PC3), represented by Months of ARV Refill, had an eigenvalue of 1.0717 and explained 15.3% of the variance,
emphasizing the importance of treatment adherence and medication refill patterns in determining OI occurrence.
Lastly, the fourth component (PC4) corresponded to the Number of EAC Sessions Completed, with an
eigenvalue of 1.0433 and a 14.9% variance contribution, reflecting the influence of counseling engagement on
adherence and health outcomes. Overall, these four components capture a substantial portion of the dataset’s
total variance, showing that both clinical factors (weight, age) and behavioral indicators (ARV refill frequency,
EAC sessions) are essential for accurate prediction and monitoring of HIV-induced opportunistic infections.
Table 2: Data Preprocessing for HIV-induced opportunistic infections
S/N
Feature
Principal
Component
Eigenvalue
Explained Variance Ratio
1
Current Weight
PC1
1.3208
0.1887
2
Age
PC2
1.1793
0.1685
Page 1835
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
3
Months of ARV
Refill
PC3
1.0717
0.1531
4
Number of EAC
Sessions Completed
PC4
1.0433
0.1490
Figure 3 shows the variance explained by each principal component in a dimensionality reduction analysis for
OI prediction. It displays a steep decline from an eigenvalue of 1.30 (first component) to 1.05 (fourth
component), forming an “elbow” after the second component (1.20). This indicates that the first two principal
components capture most of the dataset’s variance, making them ideal for simplifying data while preserving
essential information. The plot is critical in optimizing computational efficiency during model training and
improving interpretability of features influencing OI occurrence.
Figure 3: Scree plot for OI
Modeling Phase
Table 3 presents the performance of the proposed hybrid machine learning framework for early detection of
HIV-associated opportunistic infections. The model achieved a high overall accuracy of 96.79%, indicating
strong classification performance. Class 0 (non-opportunistic infection cases) shows near-perfect performance
with precision (0.9799), recall (0.9793), and F1-score (0.9796), reflecting excellent identification of non-
infection cases. Class 1 (opportunistic infection cases) also performs strongly with precision (0.9234), recall
(0.9255), and F1-score (0.9245), demonstrating good sensitivity in detecting infection cases despite class
imbalance.
The macro and weighted averages (0.95 and 0.9679 respectively) confirm balanced and consistent performance
across both classes, showing that the model is reliable for early detection of HIV-associated opportunistic
infections.
Table 3: Classification Performance Metrics
Class / Metric
Precision
Recall
F1-score
Support
Class 0
0.9799
0.9793
0.9796
1645
Page 1836
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Class 1
0.9234
0.9255
0.9245
443
Accuracy
—
—
0.9679
2088
Macro Average
0.9517
0.9524
0.9520
2088
Weighted Average
0.9679
0.9679
0.9679
2088
The confusion matrix in Figure 4 illustrates the performance of the proposed hybrid machine learning model in
classifying HIV-associated opportunistic infections. The model correctly identified 1,611 true negative cases,
indicating a strong ability to accurately classify patients without infection. It also correctly detected 410 true
positive cases, demonstrating good sensitivity in identifying infected individuals. However, the model produced
34 false positives, where non-infected patients were incorrectly classified as infected, and 33 false negatives,
where infected patients were missed. Overall, the relatively balanced number of false positives and false
negatives suggests that the model maintains a good trade-off between sensitivity and specificity. The results
indicate that the framework is highly effective for early screening, with strong predictive reliability and minimal
misclassification, making it suitable for clinical decision support in HIV-associated opportunistic infection
detection
Figure 4. Confusion Matrix of the Hybrid Machine Learning
The loss curves presented in Figure 5 provide critical insight into the model's learning dynamics and convergence
behavior over 500 training epochs. The training loss (blue line) demonstrates a sharp initial decline, indicating
that the hybrid framework rapidly learns the underlying patterns distinguishing opportunistic infections from
negative cases during the early stages of training. Simultaneously, the validation loss (orange line) follows a
similar downward trajectory, suggesting that the model generalizes well to unseen data, with no immediate signs
of severe overfitting. However, as training progresses beyond the initial epochs, both curves appear to stabilize
with the validation loss settling at a slightly higher value than the training loss. This small but persistent gap is
expected and reflects the inherent challenge of generalizing from training data to new patient samples in a
clinically heterogeneous population. Importantly, the absence of a widening divergence between the two curves
throughout the remaining epochs indicates that the model maintains stable learning without significant
overfitting, making it reliable for deployment in early detection settings where robust generalization is
paramount for patient safety.
Page 1837
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 5: Training and validation loss curves showing the model's convergence behavior and
generalization gap over 500 training epochs.
The accuracy trends illustrated in Figure 6 showcase the model's classification performance and its ability to
correctly identify both positive and negative cases of HIV-associated opportunistic infections across the training
duration. Both the training and validation accuracy curves exhibit a steep ascent during the first 100 epochs,
rising from baseline levels to approximately 0.98, indicating that the hybrid framework efficiently captures
distinguishing features early in the learning process. Beyond this point, both curves remain consistently high and
tightly aligned, fluctuating minimally around the 0.98 mark for the balance of the 500 epochs. This sustained
high performance on both the training and validation sets suggests that the model has achieved near-optimal
classification capability without sacrificing generalizability. The close alignment between the two curves is
particularly encouraging, as it demonstrates that the framework does not merely memorize training data but
instead learns robust patterns that translate effectively to new, unseen patient data. For early detection of
opportunistic infections, such high accuracy is clinically promising, as it implies a low rate of both missed
diagnoses and false alarms, thereby supporting timely and reliable clinical decision-making.
Figure 6. Training and validation accuracy
The Area Under the ROC Curve (AUC) values presented in Figure 7 offer a comprehensive measure of the
model's discriminatory power in distinguishing between patients with and without HIV-associated opportunistic
Page 1838
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
infections throughout the training process. At epoch 0, the training and validation AUC begin at 0.950 and 0.960,
respectively, indicating that even before extensive training, the model possesses considerable inherent ability to
differentiate between the two classes. By epoch 10, the training AUC rapidly ascends to 0.980, while the
validation AUC reaches 0.978, demonstrating the framework's swift capacity to refine its decision boundary with
minimal exposure to the training data. Subsequently, from epoch 20 through epoch 500, the training AUC
stabilizes at 0.988 and the validation AUC plateaus at 0.974, maintaining this near-perfect discriminatory
performance consistently across the remaining epochs. The slight but stable margin between the training and
validation AUC values suggests a modest degree of overfitting, yet the sustained validation AUC of 0.974
remains exceptionally high and clinically valuable. This enduring performance indicates that the hybrid
framework reliably separates positive from negative cases with excellent sensitivity and specificity, making it a
powerful tool for the early detection of opportunistic infections in HIV-positive individuals, where early
intervention can significantly improve patient outcomes.
Figure 7. Training and validation Area Under the ROC Curve (AUC) .
The data presented in Figure 8, which displays the fraction of actual positive cases within each predicted risk
tier, serves as a direct validation of the hybrid framework's risk calibration and clinical utility. The table reveals
a clear and monotonic gradient: only 1% (0.01) of patients assigned to the Low-risk tier were actually positive
for opportunistic infections, whereas 35% (0.35) of those in the Medium-risk tier and a striking 90% (0.90) of
those in the High-risk tier were confirmed positive. This progressive increase in the observed positivity rate
across tiers demonstrates exceptional risk stratification capability, indicating that the model's risk scores are
well-calibrated and clinically meaningful. The extremely low positivity rate (1%) in the Low-risk tier is
particularly valuable, as it validates the model's ability to safely identify patients who can be spared from invasive
diagnostic procedures and intensive monitoring, thereby reducing healthcare costs and patient anxiety.
Conversely, the 90% positivity rate in the High-risk tier represents a powerful clinical tool, as it enables clinicians
to prioritize this subgroup for immediate confirmatory testing, early initiation of prophylactic therapies, and
close follow-up, potentially averting life-threatening opportunistic infections. This near-perfect separation
between tiers underscores the framework's high predictive validity and reinforces its potential as a reliable
decision-support system for personalized risk-based screening in HIV/AIDS care, ultimately facilitating earlier
detection and improved patient outcomes.
Page 1839
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 8: Fraction of Actual Positive Cases Within Each Predicted Risk Tier
The SHAP analysis provides valuable insights into the relative importance of features used in predicting HIV-
induced opportunistic infections. The results reveal that TPT_TYPE_ENCODED, CD4_NUMERICAL, and
AGE_CLEAN are the most influential predictors, with higher SHAP values indicating a stronger contribution
to the model’s output. Among these variables, TPT_TYPE_ENCODED demonstrates the greatest predictive
influence, followed by CD4 count and age, underscoring the importance of treatment type, immune status, and
demographic characteristics in assessing infection risk. As illustrated in Figure 9, the distribution of SHAP
values enhances model interpretability by showing how individual features affect prediction outcomes, thereby
providing actionable, data-driven insights that can support clinical decision-making and improve the early
detection and management of opportunistic infections.
Figure 9: SHAP value for OI
Page 1840
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 10 presents a comprehensive analysis of the linear relationships between clinical and demographic
variables associated with viral load suppression among HIV patients. The results indicate that
“Clinical_Risk_Score_28” has a strong correlation coefficient of -0.28, establishing it as a valuable indicator in
understanding suppression outcomes. “Age” shows a smaller correlation of -0.10, while “CD4_Count_Numeric”
(0.05), “Regimen_Line_Complexity” (0.08), and “WHO_Stage” (0.12) exhibit positive correlations, signifying
that these factors positively support effective viral suppression. These relationships enhance the interpretability
of the models, offering valuable insights for identifying patient patterns and improving monitoring precision in
HIV-induced OI prediction
Figure 10: Accuracy by Activation and layers
5.1 Results and Discussion
The results of this study demonstrate the effectiveness of the proposed hybrid machine learning framework for
the early detection of HIV-associated opportunistic infections. The model achieved a high classification
performance, with an overall accuracy of 96.79%, indicating strong capability in distinguishing between infected
and non-infected cases. The classification report further shows that the model performs exceptionally well on
Class 0 (non-infection cases), with precision (0.9799), recall (0.9793), and F1-score (0.9796), reflecting its
reliability in correctly identifying healthy patients.For Class 1 (opportunistic infection cases), the model also
achieved strong performance with precision (0.9234), recall (0.9255), and F1-score (0.9245), indicating good
sensitivity in detecting infected individuals even under class imbalance conditions. The macro-average (0.9520)
and weighted-average (0.9679) scores further confirm that the model maintains balanced performance across
both classes without significant bias toward the majority class.
The confusion matrix analysis supports these findings, showing 1,611 true negatives and 410 true positives, with
relatively low misclassification rates of 34 false positives and 33 false negatives. This indicates that the model
maintains a strong balance between sensitivity and specificity, making it suitable for clinical screening where
both missed diagnoses and false alarms must be minimized.Further analysis of the training dynamics revealed
stable convergence behavior, with loss curves showing a steady decrease and minimal overfitting. Similarly, the
accuracy curves indicate consistent learning performance across training and validation sets, stabilizing at high
accuracy levels. The ROC-AUC results further confirm strong discriminatory power, with values remaining
above 0.97, demonstrating an excellent ability to separate positive and negative cases.
Comparative evaluation of activation functions and network depth shows that simpler architectures, particularly
the ReLU model, outperform deeper configurations. The SHAP analysis provides additional interpretability,
revealing that CD4 count, age, and treatment type are the most influential predictors of opportunistic infections.
This enhances the clinical relevance of the model by aligning its predictions with known medical risk factors.
Page 1841
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
The results confirm that the proposed hybrid framework is robust, accurate, and clinically interpretable, making
it suitable for early detection and decision support in HIV-associated opportunistic infections.
6.1 Conclusion
This study developed a hybrid machine learning framework combining Deep Neural Networks and XGBoost for
early detection of HIV-associated opportunistic infections. The model effectively integrates clinical,
demographic, and treatment-related features to support accurate risk prediction.
The findings show that the proposed system achieves high predictive performance, with strong accuracy,
sensitivity, and ROC-AUC values, making it suitable for clinical screening applications. Comparative
experiments confirmed that simpler architectures with ReLU activation outperform deeper and more complex
models. Furthermore, interpretability analysis using SHAP highlighted key clinical factors such as CD4 count,
age, and treatment type as major determinants of infection risk. The framework provides a reliable, interpretable,
and high-performing decision-support tool that can assist healthcare professionals in early identification of HIV-
associated opportunistic infections, ultimately supporting timely intervention and improved patient outcomes.
REFERENCES
1. Bartl, L., Zeeb, M., Kälin, M., Loosli, T., Notter, J., Furrer, H., and Nemeth, J. (2025). Machine Learning-
based Prediction of Active Tuberculosis in People With HIV Using Clinical Data. Clinical Infectious
Diseases.
2. Chen, H., Chen, F., Wang, Y., Cai, E., Pan, W., Li, Y., andSu, F. (2025). A Machine Learning Model for
Diagnosing Opportunistic Infections in HIV Patients: Broad Applicability Across Infection Types. Journal
of Cellular and Molecular Medicine.
3. Chen, J., Liu, L., Huang, J., Jiang, Y.-J., Yin, C., Zhang, L., andLu, H. (2024). LSTM-Based Prediction
Model for Tuberculosis Among HIV-Infected Patients Using Structured Electronic Medical Records: A
Retrospective Machine Learning Study. Journal of Multidisciplinary Healthcare.
4. Cherezov, D., Dam, T., Najjingo, I., Mbabazi, M., Kisembo, H., Kirenga, B., andMadabhushi, A. (2025).
Opportunistic use of artificial intelligence with X-ray imaging for diagnosis of HIV status in tuberculosis
patients in Uganda and Tanzania. PLOS Digital Health.
5. Chew, R., Woods, M., and Paterson, D. L. (2024). Development and validation of supervised machine
learning multivariable prediction models for the diagnosis of Pneumocystis jirovecii pneumonia using
nasopharyngeal swab PCR in adults in a low-HIV prevalence setting. International Health.
6. Ekpenyong, M., Etebong, P. I., and Jackson, T. (2019). Fuzzy-multidimensional deep learning for efficient
prediction of patient response to antiretroviral therapy. Heliyon.
7. Fu, X., Wu, L., Xun, J., Pütz, B., Zheng, Z., Li, Y., and Müller-Myhsok, B. (2025). Predictive survival
modelings for HIV-related cryptococcosis: comparing machine learning approaches. Frontiers in Cellular
and Infection Microbiology.
8. Henderson, H., Napravnik, S., Kosorok, M., Gower, E. W., Kinlaw, A. C., Aiello, A., and van Duin, D.
(2022). Predicting Risk of Multidrug-Resistant Enterobacterales Infections Among People With
HIV. Open Forum Infectious Diseases.
9. Inyang, U. G., Eyoh, I. J., Umoh, U. A., Ubong, E. A., Ene, E. E., Obiyo, D. C., andAkponome, B. E.
(2025). Optimal deep neural network parameters for power loss minimization analytics. Nigerian Journal
of Technology (NIJOTECH), 44(4), pp.1–14.
https://doi.org/10.4314/njt.2025.4911
10. Li, L., Shi, C., Liu, X., Li, W., Luo, Y., Zhang, H., and Shen, L. (2025). A diagnostic model for NTM
disease in HIV-positive patients: a machine learning-based analysis with novel inflammatory
markers. Frontiers in Immunology.
11. Mahmud, S., and Hossain, F. (2025). Identification of HIV-Associated Gene Expression Biomarkers Using
Machine Learning and Interpretable Artificial Intelligence. bioRxiv.
12. Ngiam, K., & Khor, I. W. (2019). Big data and machine learning algorithms for health-care delivery. The
Lancet Oncology.
13. Qasem, S. N. (2024). Introducing HeliEns: A Novel Hybrid Ensemble Learning Algorithm for Early
Diagnosis of Helicobacter pylori Infection. De Computis.
Page 1842
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
14. Sa’adah, L. N., Fauzi, F., Prizka, Arum, R., Haris, M., Yan, andBisoumi, N. (2026). HYBRID
RESAMPLING METHOD AND HYPERPARAMETER OPTIMIZATION FOR HIV/AIDS
PREDICTION: EVIDENCE FROM EIGHT MACHINE-LEARNING MODELS. JITK
(JurnalIlmuPengetahuan Dan TeknologiKomputer).
15. Sah, A., Elshaikh, R., Shalabi, M., Abbas, A. M., Prabhakar, P. K., Babker, A. M., and Agarwal, S. (2025).
Role of Artificial Intelligence and Personalized Medicine in Enhancing HIV Management and Treatment
Outcomes. Life.
16. Scott, I. (2021). Demystifying machine learning: a primer for physicians. Internal Medicine Journal
(Print).
17. Sun, Q., Zhang, K., Xu, Y., Luo, M., Yang, Z., Liu, Q., and Liu, A. (2025). Explainable machine learning
for predicting clinical outcomes in HIV/TB co-infection: a comparative retrospective study. BMC
Infectious Diseases.
18. Waring, J., Lindvall, C., andUmeton, R. (2020). Automated machine learning: Review of the state-of-the-
art and opportunities for healthcare. Artif. Intell. Medicine.
19. Xu, J., Huang, Q., Yu, J., Liu, S., Yang, Z., Wang, F., and Xiao, Y. (2022). Metagenomic Next-Generation
Sequencing for the Diagnosis of Suspected Opportunistic Infections in People Living with HIV. Infection
and Drug Resistance.
20. Zhou, J., Yang, Y., Xie, Z., Lu, D., Huang, J., Lan, L., and Huang, J. (2023). Dysbiosis of gut microbiota
and metabolites during AIDS: implications for CD4+ T cell reduction and immune activation. AIDS
(London).
21. Zhou, Y., Hemmige, V., Dalai, S., Hong, D. K., Muldrew, K., andMohajer, M. (2019). Utility of Whole-
Genome Next-Generation Sequencing of Plasma in Identifying Opportunistic Infections in
HIV/AIDS. Open AIDS Journal.
