Published in PLOS One on 2026-07-28, this study is presented as an attributed evidence briefing. The excerpts below are the original researchers’ words, reproduced under their Creative Commons Attribution licence. Health Signal has selected the abstract and, where needed, opening background paragraphs; the complete methods, results, limitations and references remain in the linked paper.
Original authors: Birhanu Betela Warssamo
Source: PLOS One • DOI 10.1371/journal.pone.0354265
Read the complete primary study ↗ · Creative Commons Attribution licence
From the researchers
Abstract
Fever remains a leading cause of morbidity and mortality among children under the age of five, particularly in low-resource settings like Ethiopia. Most existing studies in Ethiopia have relied predominantly on traditional statistical techniques, which may not fully capture complex patterns and nonlinear relationships within the data. To address this gap, the present study applies multiple machine learning algorithms to determine the most accurate model for predicting fever among under-five children in Ethiopia. This study utilized a total of 8,592 weighted child samples from the 2016 EDHS. After preprocessing, the dataset was randomly split into 70% training and 30% testing sets. Five machine learning algorithms; LR, RF, SVM, GNB, and DT were developed to predict fever. SMOTE was applied to address class imbalance in the training data. Model performance was assessed using accuracy, AUC, sensitivity, specificity, precision, F1-score, and balanced accuracy. Association rule mining via the Apriori algorithm was used to identify frequent patterns associated with fever. The overall prevalence of fever among under-five children was 14% (n = 1,197). Among the algorithms tested, the RF classifier achieved the best performance (accuracy: 92.2%, AUC: 0.958, sensitivity: 98.7%, specificity: 85.6%, F1-score: 0.8761). RF Gini importance and SHAP values ranked region, diarrhea & maternal factors as top predictors. Association rule mining revealed ten strong rules linking predictor variables with fever occurrence. The RF classifier demonstrated superior performance in predicting fever and identifying key features among under-five children. The results confirm that machine learning algorithms can serve as effective tools for accurately predicting childhood fever. These findings provide valuable insights for policymakers and public health stakeholders, supporting data-driven decision-making for targeted fever prevention strategies in Ethiopia.
1. Introduction
Fever is an elevation of body temperature above 38°C and is one of the most common clinical manifestations of infection and illness among children. It may result from viral, bacterial, or parasitic infections, such as malaria, pneumonia, and influenza, as well as inflammatory conditions or adverse reactions to medications, including antibiotics and vaccines. Common symptoms accompanying fever include chills, sweating, dehydration, and malaise [1–3].
Fever remains a major public health concern in Sub-Saharan Africa (SSA), particularly among children under five years of age [4]. It is a leading cause of childhood morbidity and mortality and is associated with malnutrition, impaired growth and development, and increased risk of death [5–8]. In low- and middle-income countries (LMICs), the burden of fever among children is particularly high, with an estimated two-week prevalence of 18.8% [9]. In Ethiopia, fever is also a common health problem among under-five children. According to the Ethiopian Demographic and Health Survey (EDHS), a substantial proportion of children experience fever in the two weeks preceding the survey, indicating a persistent public health challenge and highlighting the need for improved preventive and predictive strategies [10,11].
Reading this evidence
Use the full paper to check the study design, participant selection, comparator and outcome definitions before drawing a practical conclusion. An association does not by itself show cause and effect. A reported performance measure needs its denominator and uncertainty. A finding from a particular facility, sample or time period should retain that scope when it is discussed elsewhere.
This briefing makes the source accessible within the wider technology & industry desk. It does not assess an individual patient or certify a product. When considering relevance to another health system, examine the original authors’ account of context and limitations, then seek local implementation evidence. The related reporting below provides other studies to read alongside this one; it is a topic connection, rather than a claim that their results are directly comparable.
© 2026 Birhanu Betela Warssamo. Excerpts reproduced under CC BY. Selection, formatting and contextual notes by Health Signal. The researchers are the authors of the source study, not members of Health Signal’s staff.
Explore connected evidence
- Evidence brief: Predicting severe stunting and its determinants among under-five in Eastern African Countries: A machine learning algorithms
- Evidence brief: Interpretable ensemble machine learning framework for cardiovascular disease prediction using EMR data and large language models in Ethiopia
- Evidence brief: Forecasting under-five stunting in Ethiopia using classical and machine learning time series models
