Application of machine learning for early detection of chronic diseases in Africa

Chronic diseases in Africa

There has been an increase in chronic diseases in Africa with the trend expected to continue in the coming years. Diabetes is one of the chronic diseases that is becoming a major public health concern in the region. ¹⁸ Some of the reasons attributed to the growing cases of diabetes are urbanization, changes in lifestyles, and history of the disease in the family. The International Diabetes Federation (2020) reports that the prevalence of diabetes in sub-Saharan Africa is expected to increase by 143% by 2045. Another widespread chronic disease is hypertension with many cases going undiagnosed and untreated resulting in major complications such as stroke, heart failure, and kidney disease. ¹⁹ Moreover, According to a World Health Organization in 2021, it was reported that cardiovascular diseases were also becoming common in Africa, including coronary artery diseases and heart failures, driven by risk factors such as high blood pressure, excessive smoking, unhealthy diets, and inadequate physical activity. ²⁰

One way of overcoming the related health problems presented by chronic diseases is to ensure early detection and management. However, the existing detection and management practices for chronic diseases in Africa face several challenges. To begin with, there are limited resources with limited access to the appropriate diagnostic tools and unavailability of trained healthcare professionals. ⁸ In addition, the screening programs put in place are inadequate which has led to late diagnosis and poor management of the diseases. ⁴ The problem is worse in rural areas where access to even basic health care services is a challenge with patients having to travel long distances to receive medical attention. The disease management efforts are further complicated by the lack of standardized protocols and guidelines tailored to the African context. There have been attempts to try to overcome the challenges including initiatives such as mobile health programs that enable remote monitoring of patients and health awareness which are still in their initial stages and also have scalability and sustainability challenges. ²¹ There is, therefore, a need to explore additional solutions to help overcome the highlighted challenges.

Machine learning in chronic disease management

The use of Machine Learning is a possible solution and has shown potential in enhancing healthcare in the early detection and management of chronic diseases. ²² Recent studies highlight the use of ML in predicting chronic diseases such as diabetes, cardiovascular disease, chronic kidney disease, and cancer.^23,24 By analyzing clinical and demographic data, ML models achieve high prediction accuracies ranging from 80% to 100% depending on the disease type.^25,26 These studies show that ML can present disease risk assessment by identifying critical predictors like age, blood pressure, cholesterol levels, and other clinical markers. For example, diabetes prediction relies on factors such as waist size and sodium intake, while cardiovascular disease prediction considers age and blood pressure. ²⁷ Chronic kidney disease prediction benefits from indicators such as albumin levels and anemia, showing accuracies as high as 100% in some models. ²⁸ Additionally, ML’s integration with electronic health records, web-based tools, and standardized data models suggests promising real-world applications for early diagnosis and prevention. ²⁹ Some studies emphasize the importance of interpretability, ensuring clinicians can understand model predictions.

Some of the ML techniques that have been applied to develop predictive models for chronic diseases include logistic regression, decision trees, support vector machines, and neural networks.^20,30–32 The models are able to analyze various datasets and detected patterns and correlation that would not be possible to do with the traditional statistical methods. ³³ Logistic regression models have the capability of predicting the probability of disease occurrence by analyzing the related risk factors. Decision tree models can be used to identify the main possible predictors of a given disease.^34–36 Neural networks including deep learning models have the capability of analyzing complex data for example radiographic medical images to detect early signs of various diseases. ³⁷

At the global level, there have been several studies that demonstrate the effectiveness of ML models in the healthcare sector. To begin with, deep learning algorithms have been used to develop models for predicting cardiovascular diseases with high accuracy. ²⁰ Deep learning has also been used to train algorithms on electronic health records to predict diabetes by identifying at-risk patients up to 1 year in advance. ³⁸ In addition, convolutional neural networks have been used for the prediction of diabetes and were able to achieve a performance comparable to that of medical experts. ³⁹ There are additional studies that present the use of ML in the prediction of other chronic diseases, including cancer and asthma.^40–43 In addition to these, a study ⁴⁴ applied machine learning to predict the possible risk of cardiovascular diseases using data from wearable devices and smartphones. The models were able to enable timely interventions by providing real-time risk assessments. Routine health check-up data was also used in training a machine learning model for the prediction of hypertension with an accuracy of over 80%. ⁴⁵ However, most ML applications in healthcare have been developed in high-income countries. ⁴⁶ Such countries have well-established electronic health record (EHR) systems that provide extensive datasets for training and validation. This raises concerns about the direct applicability of these models in low-resource settings, such as Africa, where data limitations and infrastructure constraints may reduce predictive accuracy. ⁴⁷

Regional application of ML in predicting NCDs

Recent studies show the need for context-specific ML models that are adapted to regional healthcare needs. In India for example, localized ML models have improved diabetes detection rates by 30% compared to generic global models. ⁴⁸ In a similar study done in China, it was found that ML models integrating wearable device data significantly enhanced cardiovascular risk prediction. ⁴⁹ However, the African continent lags in ML-driven healthcare applications. Studies indicate that most predictive models rely on externally trained datasets, which fail to capture the genetic, environmental, and socio-economic variations unique to African populations. ⁵⁰ This gap show the need for regionally trained ML models that incorporate localized data and healthcare system constraints. Furthermore, the cloud-based implementation of the ML algorithms does not apply to most African regions with connectivity and energy limitations. In contrast, ML was applied to Indian health survey data for diabetes classification, with the study highlighting how regional adaptation improves accuracy—emphasizing the need for similar Africa-specific ML models. ⁴⁸ The challenges that limit the adoption of the solutions in Africa can be summarized as follows;

Given the challenges discussed, there is an urgent need for Africa-specific ML models that integrate regional data, infrastructure realities, and healthcare constraints. Unlike existing global models, locally trained ML systems can enhance predictive accuracy and improve early disease detection. This study contributes to addressing this gap by:

By implementing context-aware AI solutions, this research aims to bridge the gap between ML advancements and practical healthcare applications in Africa, paving the way for improved early detection and management of chronic diseases.

Dataset

The dataset used in this study was based on electronic health records of 768 patients with different attributes relevant to diabetes prediction. The sample size of 768 patient records was determined based on the availability of data from the clinic over a 3-month period. While not a probabilistic sample, this dataset provides a representative distribution of diabetes cases within the study setting. The dataset primarily comprises patients from an urban clinic, and may therefore underrepresent rural populations, who face different risk profiles and healthcare access barriers. The socioeconomic diversity within the dataset is also limited. These limitations suggest caution in generalizing findings across all African populations. To enhance reproducibility, a synthetic sample mirroring the data distribution is being prepared and will be made available through a GitHub repository. However, the use of data from a single clinic introduces limitations in generalizability. Future studies will prioritize multi-site and stratified sampling strategies to improve population diversity. The current dataset may be biased toward urban demographics or a specific socio-economic group. The key attributes include the number of pregnancies, glucose levels, blood pressure, skin thickness, insulin levels, Body Mass Index (BMI), and diabetes pedigree function based on family history and age.

The diagnosis of diabetes in this study was based on standard medical guidelines that utilize glycated hemoglobin (HbA1c) levels and fasting blood sugar (FBS) levels as key diagnostic markers. The following thresholds were applied:

To ensure the relevance and reliability of the dataset, the study included patients who met the following criteria:

Certain individuals were excluded to maintain data integrity and ensure consistency in the predictive model:

Model development process

The model was developed following the steps outlined in Figure 3.

OPEN IN VIEWER

Figure 3.

Model development steps.

The dataset was preprocessed before training the models to ensure accuracy and reliability. Missing values, mainly for BMI and glucose levels, were imputed using the median to maintain data consistency. To ensure robustness, we also conducted a sensitivity analysis by comparing model outputs after applying mean imputation and complete case analysis. The performance metrics showed less than 3% variation, justifying the choice of median imputation. Outliers were identified using the Interquartile Range (IQR) method and adjusted at the 1st and 99th percentiles to prevent them from affecting model performance. Feature selection was conducted through correlation analysis, which showed glucose levels, BMI, and age as the strongest predictors of diabetes.

Five machine learning models—Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), Naïve Bayes (NB), and an Optimized Neural Network (NN)—were developed and evaluated to balance model complexity, predictive performance, and interpretability. The Decision Tree and Naïve Bayes classifiers were selected for their high interpretability and transparent decision boundaries, which are essential for clinical decision support and can aid in gaining practitioner trust. SVM and Neural Networks were included to capture complex, non-linear relationships in the data that simpler models may not effectively handle. K-Nearest Neighbor (KNN) was chosen due to its simplicity, effectiveness on smaller datasets, and non-parametric nature, which allows it to adapt to complex decision boundaries without prior assumptions about data distribution. KNN also served as a useful benchmark to compare against more advanced models.

Logistic regression, while a common choice in clinical prediction tasks due to its interpretability, was not included as preliminary data exploration indicated non-linear interactions among key predictors—such as glucose level, BMI, and age—which it may not capture without significant feature engineering or transformation. Logistic regression offers limited optimization flexibility compared to models like decision trees and neural networks, which support techniques such as pruning and quantization for resource-constrained environments.

This diversity of models enabled a comparison of trade-offs between model interpretability and predictive performance. Hyperparameter tuning was conducted to improve accuracy, with the Decision Tree optimized using GridSearchCV and the Neural Network refined through node-level structured pruning and 8-bit quantization. These optimizations significantly reduced model size and memory footprint, making the NN model suitable for edge-based deployment with minimal latency. Robustness and generalizability were ensured using 5-fold cross-validation to mitigate overfitting.

Model evaluation

Statistical analysis was performed using Python and the Scikit-learn library. Descriptive statistics were computed for demographic and clinical variables. The performance of the developed models was evaluated using key metrics applied in ML model evaluation. The metrics include the model accuracy, recall, precision, F1 score, and the ROC-AUC (Receiver Operating Characteristic - Area Under Curve).

Ethical considerations

The study complies with ethical guidelines outlined by the World Medical Association (WMA) Declaration of Helsinki (2013) for research involving human health data. The datasets used in the study ensured the anonymity and confidentiality of patients. Ethical approval was obtained from the University of Rwanda Institutional Review Board, and a waiver of informed consent was granted due to the retrospective and anonymized nature of the data. For future edge deployment, on-device encryption and role-based access controls are being considered to mitigate data privacy risks in low-connectivity environments. Bias and fairness were addressed by mitigating any dataset imbalances. The machine learning model developed is interpretable and can only be used as a support tool for clinicians, not a replacement.

To address equity concerns, it is essential to ensure the model does not present healthcare disparities. Deployment must be accompanied by equitable access initiatives such as training local health workers and adapting tools to low-literacy users. Regular audits should be conducted to detect algorithmic bias, and datasets must be diversified across gender, geography, and socioeconomic status. Moreover, local data governance structures must be involved to ensure compliance with evolving data privacy laws and cultural sensitivities.

Baseline characteristics

Table 1 presents the baseline table summarizing patient characteristics, including mean, standard deviation, median, and range for key clinical variables. The p-values indicate statistical significance between diabetic and non-diabetic groups. The p-values were obtained using the independent samples t-test for continuous variables to compare the means between diabetic and non-diabetic groups. This test was chosen because it evaluates whether there is a statistically significant difference in the means of two independent groups.

OPEN IN VIEWER

Table 1.

Baseline table.

Variable	Mean ± SD	Median	Range	p-Value
Age (years)	33.2 ± 11.8	29.0	21–81	<0.001
BMI (kg/m²)	32.0 ± 7.9	32.0	0.0–67.1	<0.001
Glucose (mg/dL)	120.9 ± 32.0	117.0	0–199	<0.001
Blood pressure (mmHg)	69.1 ± 19.4	72.0	0–122	0.087

Age, BMI, and glucose levels showed statistically significant differences between diabetic and non-diabetic groups (p < 0.001), meaning these factors are likely important predictors of diabetes. Blood pressure, however, did not show a significant difference between groups (p = 0.087).

The feature importance analysis presented in Figure 4 was primarily derived from the Decision Tree (DT) model, which provides inherent feature ranking based on information gain. To validate these findings, feature importance was also extracted from the Support Vector Machine (SVM) model using a permutation-based method, and from the Optimized Neural Network (NN). Across all three models, glucose level, BMI, and age consistently emerged as the most influential predictors of diabetes. The Naïve Bayes model, while not inherently producing feature importance scores, also indirectly highlighted similar features through high conditional probability contributions. K-Nearest Neighbor (KNN), being a non-parametric model, does not provide traditional feature importance scores, but a univariate sensitivity analysis confirmed alignment with the top features identified by the other models.

OPEN IN VIEWER

Figure 4.

Feature importance for the DT model.

In addition, the correlation analysis reveals insights into the relationships between various medical attributes and the presence of diabetes, with glucose levels, BMI, and age showing the highest positive correlations with the output variable. The prioritization of features facilitated efficient feature selection for model development and thus enhanced the predictive accuracy. These findings show the importance of glucose and BMI as critical indicators in diabetes prediction models, which is consistent with established medical knowledge. This provides a foundation for developing targeted early detection systems in African healthcare settings.

Model performance

The results of the K-Nearest Neighbors (KNN) classifier are summarized by the accuracy scores for both the training and testing datasets across different numbers of neighbors as presented in Figure 5. The accuracies were calculated using a range of neighbors from 1 to 34. The training accuracies show a decline from 1.0 when the number of neighbors is 1, indicating overfitting, to 0.7378 as the number of neighbors increases to 34. This pattern suggests that the model complexity decreases with more neighbors. The test accuracies show an initial increase from 0.6688 to 0.7532 indicating that the model’s performance on unseen data improves as it becomes less overfitted and more generalized. This analysis shows the importance of selecting an appropriate number of neighbors in KNN to balance the trade-off between overfitting and generalization to enhance the model performance on unseen data.

OPEN IN VIEWER

Figure 5.

KNN training and test results.

The Decision Tree model from the initial training has an accuracy of 1.0, which indicates overfitting, with a test accuracy of 0.6948. To improve the model’s performance hyperparameter tuning was conducted using GridSearchCV identifying the best parameters as max_depth = 3, min_samples_leaf = 4, and min_samples_split = 2. With these optimized hyperparameters, the model’s test accuracy improved to 0.7532, and the training accuracy to 0.7769, showing a better balance between fitting the training data and generalizing to unseen data. Figure 6 shows the confusion matrix for the DT model

OPEN IN VIEWER

Figure 6.

Confusion matrix for the DT model.

The Gaussian Naive Bayes (GNB) classifier achieved a training accuracy of 75.24% and a test accuracy of 75.97%. This indicates that the model generalizes well, performing consistently on both the training and unseen test data. Figure 7 shows the confusion matrix for the GNB model.

OPEN IN VIEWER

Figure 7.

Confusion matrix for the GNB model.

The SVM model’s performance is evaluated through its accuracy on both training and test datasets, achieving 78.66% on the training data and 77.27% on the test data. This close alignment between training and test accuracies indicates that the model generalizes well to unseen data.

The initial neural network model achieved an accuracy of 66.4% and a loss of 0.77 on the validation set. The model was optimized by pruning and quantization with the best model achieving the highest accuracy at 89%. Further, the model maintained a minimal latency (1 ms out of 100 ms) and was resource-efficient with 1 kB of RAM and 15 kB of ROM requirements. Figure 8 shows the confusion matrix for the model.

OPEN IN VIEWER

Figure 8.

Confusion matrix for the optimized NN model.

Although the NN model was optimized for low-resource deployment, real-world testing on edge hardware such as Raspberry Pi or smartphones was not yet conducted. Simulations suggest feasibility based on latency and memory usage benchmarks, but future empirical evaluations are necessary.

Comparison of the models

The graph presented in Table 2 compares the performance of different machine learning models used for diabetes prediction based on five metrics: accuracy, precision, recall, F1 score, and ROC-AUC. The SVM model shows strong and balanced performance across all metrics. After hyperparameter tuning, the Decision Tree model achieved balanced metrics, though slightly lower than the SVM. The GNB model displayed consistent performance. The initial neural network had lower performance across all metrics, indicating the need for optimization. The optimized with NN outperformed all other models with the highest accuracy (89%) and better performance across all other metrics, indicating it is the most reliable model for diabetes prediction among those tested.

OPEN IN VIEWER

Table 2.

Model performance comparison.

Model	Accuracy (%)	Precision	Recall	F1-Score	ROC-AUC
DT	75.3	0.74	0.77	0.75	0.76
SVM	77.2	0.76	0.78	0.77	0.78
NB	75.9	0.75	0.76	0.75	0.76
Optimized NN	89.0	0.87	0.90	0.88	0.91