Ensemble Machine Learning for Malaria Diagnosis in Resource-Limited Settings Using Clinical and Demographic Features

1. Introduction

More than half of the world population is affected by malaria, a parasitic disease that remains a major public health problem [1]. The 2024 World Health Organisation (WHO) malaria report estimates that 263 million cases with 597000 deaths occurred in 2023. The malaria report also highlights that the WHO African regions, especially Sub-Saharan Africa, continues to shoulder the heaviest burden, contributing $94\%$ of the cases and $95\%$ of the deaths globally [2]. Malaria disproportionately affects the rural areas, especially poor communities with limited to no access to healthcare [3,4]. Annually, over five million people are at risk of contracting malaria in Zimbabwe. Zimbabwe malaria statistical reports indicated 16794 cases and 32 deaths in the first half of 2024 [5]. Of those cases, 199 were children under five years of age, an indication of the persistent burden of malaria within the country. Many at-risk groups continue to miss out on needed services to prevent, diagnose and treat malaria [2].

Malaria is an infectious disease caused by the Plasmodium parasite [6,7]. The most severe malaria cases are caused by the Plasmodium falciparum species [7,8]. Traditional parasitology diagnostic methods, such as malaria Rapid Diagnostic Tests (mRDTs) and microscopy, are widely used for malaria diagnosis [9]. Microscopy is the gold standard for malaria diagnosis since the early 20th century, but it is time-intensive, requires skilled personnel, and is prone to variability in accuracy depending on operator expertise [10,11]. Similarly, mRDTs are one of the most efficient tools used to accurately determine a patient’s malaria status. Still, their sensitivity is reduced at low parasite densities, producing false-negative or false-positive results [11,12]. Resource-limited settings are usually affected by the absence of a definitive diagnosis, which is a serious obstacle to treatment compliance, efficacy, and clinical care of severe malaria cases [8,13].

Traditional statistical methods, such as logistic regression (LR) and time series analysis, have been used in malaria research to identify risk factors and model malaria prevalence [14]. However, they have been found to depend on data assumptions, focus on explanation over prediction, and struggle with high-dimensional clinical data, but are very useful in understanding malaria trends [14,15]. Machine Learning (ML), a subfield of Artificial Intelligence (AI) that focuses on developing the algorithms that can learn patterns and relationships from data without being explicitly programmed, offers an alternative to supporting malaria diagnosis in resource-limited settings [16,17,18].

A study focusing on ML for malaria prediction using clinical and demographic features was carried out in Uganda [3]. Individual classifiers such as Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting (GB), Decision Trees (DT), Naive Bayes (NB) and K-Nearest Neighbors (KNN) were used to build models using the clinical and demographic features [3]. The individual ML model performance was evaluated using metrics such as accuracy, precision and recall, and they all achieved a good performance [3]. Ensemble models such as bagging, AdaBoost, soft and hard voting and stacking were created from the individual models. The ensemble models outperformed the individual classifiers, achieving over 0.98 across the same metrics [3].

In Yumman Province, China, a study compared the performance of traditional time series models and deep learning algorithms for malaria prediction [8]. The autoregressive integrated moving average (ARIMA), seasonal and trend decomposition using Loess (STL+ARIMA), back propagation artificial neural network (BP-ANN), and long short-term memory (LSTM) network models were applied separately in simulations using malaria data and meteorological data [8]. GB regression trees were used to combine the four models (ARIMA, STL+ARIMA, BP-ANN and LSTM) through stacking [8]. The findings reported that predictive performance of the stacking ensemble model was superior to that of the individual models, indicating that stacking may have significant implications for malaria disease prediction [8].

Despite the increasing volume of research utilising ML methods for malaria prediction and diagnosis, several significant gaps persist. First, numerous current studies depend on laboratory-based data [19,20] or microscopy images [21,22], that are frequently inaccessible in resource-limited environments. Second, although individual ML models have demonstrated encouraging outcomes in malaria predictions [23], there has been insufficient focus on the systematic assessment of ensemble learning approaches for malaria diagnosis utilising routine clinical and demographic information. Third, the majority of existing research is geographically concentrated outside of Southern Africa [24,25,26,27,28,29], which restricts the contextual relevance and generalizability of their results to nations like Zimbabwe, where malaria continues to be endemic, and healthcare resources are limited [5,30,31].

In addressing these challenges, this paper aims to bridge the existing gap by developing and evaluating an ensemble ML framework for malaria diagnosis, utilising readily available clinical symptoms and demographic characteristics. By leveraging retrospective patient data from Gutu Mission Hospital and Gweru Provincial Hospital in Zimbabwe, this research systematically compares the performance of several individual machine learning classifiers with ensemble learning techniques, including bagging, voting, and stacking. A particular focus is placed on assessing whether ensemble models can significantly enhance diagnostic accuracy, robustness, and generalizability compared to individual classifiers.

2. Materials and Methods

2.1. Study Design and Data Collection

This paper followed the methodology described in Figure 1. A retrospective quantitative study was conducted using anonymised patient records from Gweru Provincial Hospital (urban setting) and Gutu Mission Hospital (rural setting) in Zimbabwe, selected to represent high and low malaria transmission areas.

2.2. Inclusion and Exclusion Criteria

Patient records from January 2022 to December 2024 with a confirmed malaria diagnosis (positive/negative) either through microscopy or mRDTs, with comprehensive clinical and demographic data, were included in this study. However, records with missing key variables, such as malaria test results, cases with co-infections such as typhoid, or ambiguous diagnostic outcomes, were excluded from the study.

2.3. Sample Size and Sampling Procedure

This paper used the entire dataset of eligible records, 637 participants tested for malaria with complete details from Gutu Mission Hospital and Gweru Provincial Hospital. The dataset consisted of participants who tested either positive or negative for malaria using microscopy or mRDTs.

2.4. Dependent and Independent Variables

The outcome variable for this study, referred to as malaria status, is a binary variable that captures whether the patient was positive or negative for malaria. The predictors included clinical symptoms (fever, chills, headache, abdominal pain, diarrhea (binary variables)) and demographic features (age, gender, residence and travel history).

2.5. Data Preprocessing

A Microsoft Excel file was created for the malaria dataset to facilitate data pre-processing in R version 4.5.1. Data pre-processing in this paper comprised data cleaning, feature selection, class-imbalance handling, data encoding, and data splitting.

2.6. Data Cleaning

Data entry typos, wrong numerical values and incorrect data formats were identified by comparing the paper-based medical records and the computed spreadsheet, and they were corrected to align with the original dataset. Duplicate records were checked for by verifying key identifiers such as patient ID and test date, and were removed to avoid redundant data from skewing the ML models. The dataset was then checked for outliers to make sure no extreme values might affect predictive accuracy. Data points with extreme values were capped, which means replacing extreme values with upper and lower threshold values.

2.7. Feature Selection

Variance inflation factor (VIF) and Pairwise Pearson correlation coefficients ($\mathrm{r}$) were used to study the relationships between the variables. Recursive Feature Elimination (RFE) guided by RF was used for feature selection, choosing the most effective features for model building.

2.8. Handling Class Imbalance

An imbalanced dataset is when one class outnumbers the other [32]. This imbalance can lead to biases towards the majority class [33]. The majority of the malaria diagnosis results from the collected data were negative (562) compared to the 75 positive cases, very underrepresented in a study comprising 637 participants. Synthetic Minority Oversampling Technique (SMOTE) was used to address data imbalance using the SMOTE function in R. This was done through creating synthetic samples for the minority class by interpolating between existing minority class [34,35].

2.9. Data Encoding

One-Hot encoding was used to change categorical variables into numerical representations to make them compatible with ML algorithms. Age was grouped into categorical bins (0-5 years, 6-15 years, 16-30 years, 31-45 years, 46-60 and 61+ years) (Table 1). The decision to categorise age was based on clinical and epidemiological literature, which has successfully proven that working with age bins for malaria-related studies takes into consideration that malaria incidence and severity vary non-linearly across age groups and effectively improves model performance [36,37,38].

2.10. Data Splitting

The dataset was split using stratified sampling into training, evaluation and testing sets at a ratio of 70:20:10, respectively. The 70% for training allows for sufficient model training, and the 20% for evaluation ensures there is enough data for hyperparameter tuning and evaluation of model performance. The $10\%$ for testing was used for the final evaluation of model performance and was strictly unseen during training and hyperparameter tuning to provide an unbiased estimate for the model’s performance on unseen data.

2.11. Individual Model Selection and Training

Seven individual ML classification algorithms were selected and trained to evaluate their contributions to specific ensemble methods. The models included LR, GB, KNN, XGBoost, RF, DT, and NB. Model selection was guided by a comprehensive review of the literature on malaria status prediction [3,16,39]. All seven models were trained using 70% of both the balanced and unbalanced datasets to compare their predictive performance.

2.12. Hyperparameter Tuning

Optimisation of model performance was done by tuning hyperparameters relevant to specific ML classifiers. Lasso regression was used to prevent overfitting and improve the generalisation of LR. LR was also trained on 5 fold cross-validation to improve performance.

RF and XGBoost were tuned using Grid Search. The hyperparameter mtry in RF was tested for values $\mathrm{c}(1,2,3,4)$ and 10-fold cross-validation was applied to verify for favourable parameter settings. For XGBoost, the hyperparameters tuned included the number of trees, learning rate and maximum depth. 5 cross-fold validation was used to evaluate the most favourable hyperparameters for XGBoost.

Cross-validation-based pruning was used for DT. It was applied to optimise the tree’s complexity and prevent overfitting using the rpart function in R. Random search was used for tuning hyperparameters for KNN, where tune length was set to 10 to allow KNN to explore a range of values for k. For GB, Bayesian optimisation was used to tune the hyperparameters such as the number of trees, learning rate and tree dept. Cross-validation was used for NB, and tuneLength was set to 10.

2.13. Ensemble Model Selection and Building

Using ensemble techniques through the integration of various ML classification algorithms has proven to achieve greater precision performance compared to utilising a solitary technique [40,41]. The study employed bagging, stacking, soft voting and adaboost.

Bagging works by reducing variance and preventing overfitting by training multiple models on different subsets of the training data and averaging their predictions [42,43]. RF was preferred for the bagging method as it uses multiple decision trees during training [44]. Training of the bagging model was done using the randomForest function in R.

Stacking uses basic-level meta classifiers and amalgamates them with meta-learner classifiers [43,45]. The base learners used for stacking are $\mathrm{RF},\mathrm{DT},\mathrm{KNN},\mathrm{GB},\mathrm{NB}$ and XGBoost and LR as the base classifier. Prediction models were first obtained from each of the base models. A new dataset (predictions_stacked) was created, and it contained all base learners and the actual target variable (Diagnosis).

Soft Voting used LR, RF, DT, KNN, GB, NB and XGBoost as base models. These base models were averaged, and the predicted probabilities were combined [46]. AdaBoost used the ada package in R, with 100 iterations per iteration. AdaBoost adjusts the weights of misclassified instances to improve the performance of weak learners [45,47].

2.14. Individual and Ensemble Model Performance Evaluation

The individual and ensemble models were evaluated on their performance in malaria prediction using metrics such as accuracy, recall, precision, confusion matrix, F1 score and AUC-ROC. Predictions and predicted probabilities were generated for each model, and the performance metrics were calculated and compared. For ensemble models, the same metrics were used to assess the effectiveness. Model performance was also compared between individual and ensemble models.

2.15. Ethical Review

Ethical approval was obtained from Gweru Provincial Hospital’s Research Ethics Committee and Gutu Rural District Hospital’s Research Ethics Committee. Measures were implemented to ensure any ethical issues relating to the study were addressed. All patient identifiers, such as names, national identification numbers, and contact details, were removed during data collection. For issues related to data security, the dataset was kept secure and accessible to authorised people only.

3. Results

3.1. Analysis of Demographic and Clinical Variables Associated with Malaria Diagnosis

A total of 637 participants with an age range of 1 to 90 years were included in this paper. The mean age was 30.7 $\pm 21.4$ years, indicating a wide variation in participant ages. Out of the 637 participants $49.9\%$ were from Gweru Provincial Hospital while $50.1\%$ were from Gutu Mission Hospital. The distribution of gender was similar between groups, with no significant difference observed in their diagnosis results ($p=0.483$, Table 2).

However, participants who tested positive were significantly more likely to report fever ($84.0\%$ vs. $66.0\%,p=$ 0.002), chills ($81.3\%$ vs. $62.3\%,p=0.001$), diarrhea ($41.3\%$ vs. $26.5\%,p=0.007$), and abdominal pain ($50.7\%$ vs. $29.5\%,p<0.001$) compared to those who tested negative (Table 2). A higher proportion of positive cases had a recent travel history ($49.3\%$ vs. $35.2\%,p=0.017$) and resided in rural areas ($65.3\%$ vs. $51.1\%,p=0.020$). No significant differences were observed in the distribution of headache or age groups between the two diagnosis categories (Table 2).

3.2. Feature Selection

A correlation analysis of selected numeric and binary-encoded variables (age, gender, headache, fever, abdominal pain, diarrhoea, chills and travel history) confirmed that the predictors are relatively independent of each other. The pairwise Pearson correlation coefficients ($\mathrm{r}$) (Figure 2) show that most of the correlations are close to zero, indicating weak linear relationships between variables. The highest observed correlation was between chills and fever, with a coefficient of -0.40. Travel history showed weak to negligible correlations with all other features, except a moderate negative correlation with age ($\mathrm{r}=-0.08$).

VIF values for the predictors were all below 5, indicating no significant collinearity among predictors (Table 3). The highest VIF was 1.35 for chills, and the lowest VIF was 1.01 for gender. Findings align with the weak correlations observed in the correlation matrix (Figure 2).

Using RFE, the top five variables identified (Figure 3) were chills, fever, diarrhea, travel history, and abdominal pain. All variables demonstrated statistical significance based on chi-square tests (Table 2), with p-values of 0.0019 for chills, 0.003 for fever, 0.01 for diarrhea, 0.02 for travel history, and 0.0004 for abdominal pain.

SMOTE successfully addressed the data imbalance issue to improve model performance [Figure 4].

3.3. Machine Learning

The performance of the seven ML individual classifiers (LR, RF, DT, GB, KNN, NB and XGBoost) were evaluated using six metrics. XGBoost got the highest accuracy (0.95). GB and RF both got an accuracy of 0.94. DT achieved an accuracy of 0.89, while LR had 0.83 and NB had 0.82, showing lower but comparable performance. KNN had the lowest accuracy of 0.69. XGBoost achieved the highest precision (0.93). RF and GB had precision values of 0.91 and 0.92, respectively. DT (0.87) and LR (0.85) achieved moderate precision values. KNN had the lowest precision of 0.68. Recall was strongest for RF at 0.99. XGBoost and GB both got a recall of 0.98. DT obtained 0.93, showing a good recall performance. LR (0.82) and NB (0.77) demonstrated more limited sensitivity. KNN achieved 0.79 recall, performing better in this metric than in others.(Figure 5).

The AUC-ROC values, indicating overall classification ability across all thresholds, were strongest for XGBoost and GB (both 0.99). RF achieved 0.98, DT (0.95), and NB (0.90), while LR achieved 0.89. KNN (0.75) had the lowest AUC-ROC performance (Figure 6).

Confusion matrices were used to evaluate the models, where each matrix shows the number of actual malaria cases (positives) and non-cases (negatives) correctly or incorrectly predicted by the model (Figure 7). LR correctly classified $40.7\%$ true positives $(\mathrm{TP})$ and $42.6\%$ true negatives $(\mathrm{TN})$, then misclassified $7.4\%$ of actual positives as false negatives (FN) and 9.3% of actual negatives as false positives (FP). XGBoost achieved strong performance, with $44.4\%\mathrm{TP}$ and $51.9\%\mathrm{TN}$. Impressively, it recorded $3.7\%\mathrm{FN}$ and $0\%\mathrm{FP}$, making it the model with the lowest error. GB correctly identified $43.5\%$ of TP and $50.9\%$ of TN. It had minimal misclassifications of $4.6\%\mathrm{FN}$ and only $0.9\%\mathrm{FP}$.

Comparative performance analysis of models between the test and evaluation sets showed consistent patterns in predictive accuracy and reliability. XGBoost showed superior performance across all metrics, achieving the highest accuracy (0.95 test, 0.93 eval), F1-score (0.96 test, 0.93 eval), AUC-ROC (0.99), precision (0.93) and recall (0.98). RF and GB models had a nearly similar performance to XGBoost. DT showed moderate performance with accuracy and F1-scores around 0.89 (Figure 8).

The Stacking model demonstrated the highest overall performance across all evaluation metrics. Stacking has an accuracy of $96\%$ while Soft Voting, Bagging and AdaBoost classifiers achieved an accuracy score of $94\%$. Soft Voting had a slightly higher precision (93%). Stacking had the highest precision of 95% (Figure 9).

The ROC curves showed exceptional discriminative performance across all ensemble models. Soft Voting and Bagging achieved a high classification capability (AUC $=0.99$ ). AdaBoost and Stacking had AUC values of 0.98. All ensemble models maintained high sensitivity with their curves hugging the top-left corner of the plot, a nature only high-quality classifiers take. The clustering of AUC scores between 0.98 and 0.99 confirms that all ensemble methods provide clinically reliable diagnostic performance (Figure 10).

Confusion matrices for the four ensemble models show distinct performance quality in malaria classification (Figure 11). Soft Voting correctly identified 48 TP cases (44.9%) and 53 TN (49.1%), with error rates 4 FN (3.7%) and 3 FP (2.8%). Bagging had stronger specificity, with only 1 FP and 55 TN, with 6 FN at 5.6%. AdaBoost demonstrated the best sensitivity among all ensemble models, with 3 FN (2.8%) and 49 TP (45.4%) and only 1 FP. Stacking matched AdaBoost’s sensitivity with 3 FN at 2.8%, 49 TP at 45.4%, but with more FP. All models maintained FP rates below 3%.

4. Discussion

The findings of this research demonstrate the potential impact of ML ensemble models in enhancing malaria diagnosis, particularly in developing countries where there is poor accessibility and insufficient diagnostic tools for malaria. Our results show that the Stacking ensemble model outperformed all individual models, obtaining an accuracy of $0.96$, precision $=0.95$, recall $=0.98$, F1 score $=0.96$ and AUC-ROC $=0.98$.

Of the 637 participants from Gweru Provincial Hospital (49.9%) and Gutu Mission Hospital (50.1%), a proportion of 50.9 were males and $49.1\%$ were females. The age of the participants had a mean of $31\pm 21$ years. A total of

75 cases were positive, with $54.7\%$ of the 75 cases being males and $45.3\%$ females showing a slightly higher prevalence in males, although the difference did not attribute any statistical significance ( $p=0.56$ ). Clinical symptoms such as chills, fever, abdominal pain and diarrhea were significant predictors of malaria ( $p<0.05$ ), highlighting the dependence of clinical symptoms as diagnostic markers for malaria. While most demographic features were not statistically significant predictors of malaria, travel history proved to have a significant effect towards malaria incidence, as individuals who travel may have different risk exposure and prevention behaviours. These results align with existing literature, which emphasises how physical clinical symptoms of malaria can be relied on for diagnosis, especially in malaria endemic areas [48]. Some studies argue that demographic factors are as important as clinical symptoms, reporting gender and age as having a huge impact on malaria incidence [49,50]. Other literature reports that while demographics are found to be related to malaria incidence, different geographical and socio-economic conditions impact their significance to malaria epidemiology [51,52]. Early diagnosis of malaria through timely identification of clinical symptoms can potentially improve malaria management and control in endemic localities [49,53].

The performance evaluation of the ML models, measured by considering all the metrics, showed that ensemble models outperform individual ML models. XGBoost achieved an accuracy of 0.95 while RF and GB got 0.94, indicating strong generalizability. A comparison of the individual models’ performance on the evaluation and test datasets (Figure 8) showed consistent performance, proving that there was minimal overfitting due to effective hyperparameter tuning. Models such as KNN and NB, which demonstrated moderate model performance, could suggest their applicability for specific scenarios with fewer complexities.

While all the ML ensemble models achieved high diagnostic performance (Figure 9), the Stacking ensemble model outperformed the other ensemble models, obtaining a high malaria diagnostic performance across all metrics. Bagging, Stacking and AdaBoost achieved 0.98 recall, and Soft Voting got 0.95. High recall in malaria diagnosis means reliable identification of positive cases, reducing the risk of severe complications. Mahajan et al. [54] and Rajab et al. [55] obtained very high precision scores for Bagging, Stacking and AdaBoost, emphasising that ensemble models are reliable in boosting sensitivity in malaria diagnosis. F1 scores for the ensemble models were high, with Stacking and AdaBoost at 0.96. Soft Voting and Bagging got 0.94, indicating a strong overall performance. The paper’s results for the F1 score of the ensemble model performance highlight the potential of the ensemble models at balancing precision and recall. Ensemble models demonstrated exceptional reliability in making predictions with minimal errors (Figure 11). The top-performing individual models in our study were XGBoost, GB and RF, while the top-performing ensemble model was stacking.

5. Conclusion

The investigation demonstrated that the use of clinical and demographic data for malaria diagnosis with ensemble ML models is viable. These ensemble models provide a scalable and cost-effective alternative, or complementary tool, to traditional malaria diagnostic methods in resource-limited settings. Clinical symptoms such as chills, fever, and abdominal pain exhibited strong predictive power, underscoring their importance and suitability for symptom-based screening tools. These findings may support malaria control programs in high-burden areas.

Notably, limitations such as sample size, data diversity, model complexity and possible deployment issues were acknowledged. Integrating these stacking ensemble models into a user-friendly digital platform can help health workers gain access to reliable diagnostic support, reducing the burden on health systems. Validating the proposed models in real-time clinical workflows should be adopted to assess their clinical applicability. Model integration should be piloted into hospital systems or mobile health platforms to assess clinical acceptance and usability.

This paper contributes to the growing evidence supporting the use of AI in global health. By demonstrating the feasibility of ML-driven malaria prediction, this study lays a foundation for scalable and tailored data-informed diagnostic tools that can be used to support infectious disease management in resource-limited settings. It also adds to the countless efforts of developing nations towards the achievement of sustainable development in the healthcare sector under the SDG-3 pillar of the United Nations.

Future research includes focusing on large-scale validation and real-world implementation to ensure the model’s effectiveness in clinical settings. To further advance the performance of ensemble models, focus more on investigating the impact of more than base models within ensemble models and combining ensemble models to explore their impact.