The central aim of the study is to use hybrid deep learning models to effectively predict the levels of diarrhea prevalence i.e., Minimal, Mild, Moderate and Severe- in various geographical areas using systematic, numerical health information. What was previously a comprehensive dataset was narrowed and pre-processed to extract the indicator of “DIARCARE” which is a diarrhea-related healthcare visit. Numerical variables were normalized, noise was introduced to enhance generalization and a categorical target variable was produced to capture the prevalence levels. In order to utilize the spatial and sequential patterns of the data, several deep learning models were executed such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs), Gated Recurrent Units (GRUs), Multi-layer Perceptrons (MLPs), and a combination of them. It is worth mentioning that hybrid architectures (CNN-RNN, RNN-GRU, LSTM-GRU, FCNN-MLP) were proposed to improve the capacity of learning since they can represent both long-term and short-term dependencies of the dataset. The effectiveness of the suggested method is proven by experimental results. CNN model recorded almost perfect results with a 100 percent validation rate and a loss of 6.24e-04 whereas LSTM, MLP and TabNet models demonstrated high accuracy and recall with over 99 percent in the majority of classes. A number of hybrid structures including Hybrid RNN-MLP and Hybrid GRU-LSTM also achieved 100 percent classification accuracy on the test set strengthening the power of deep learning ensembles in health prediction tasks. These findings highlight the possibility of incorporating artificial intelligence in the systems of social health surveillance. This piece has provided a precise framework to forecast childhood diarrhea and also in promoting larger objectives in community health. It assists in identifying dangerous locations, allocating medical supplies where required, participating in prevention efforts, and showing how AI may monitor the response to illnesses and make decisions on child healthcare practices. The findings presented in this research can be used by the government and non-governmental organizations to develop health interventions using actual data.

1.1 Key Contributions of the Study

Childhood diarrhea remains a major burden of poor health, particularly in low-resource environments despite the significant global interventions. Conventional approaches to surveillance are usually based on the use of manual data collection and report, which is time consuming, reactive and not proactive, and also limited in scope. The world is in urgent need of smart systems capable of using the available health information to anticipate the prevalence of illnesses, detect the risk patterns, and assist in timely interventions. To fill the gap, this paper proposes an AI-based framework that utilizes the child health indicators published by UNICEF to accurately forecast the prevalence of diarrhea. The contribution of the study is:

1.      To carry out preprocessing and feature engineering, such as standardization, noise injection and selection of relevant features to enhance model generalization and decrease overfitting. 2. To conduct feature importancy analysis to determine the major causes of diarrhea, including access to clean water, healthcare facilities, and food.

2.      To design and develop a hybrid deep learning model to classify childhood diarrhea prevalence using structured health indicator data from the UNICEF dataset, and to compare it with various deep learning architectures.

3.      To validate the deep learning approach in diarrhea prevalence prediction with high accuracy, precision, recall, and F1-score across all model evaluations, demonstrating the potential of AI in strengthening public health surveillance systems.

2. Literature Review

The major issue of childhood diarrhea in these parts of the world leads to a high number of sicknesses and deaths among children. By using clear health indicators and advanced technique, the prediction of health concerns can be made more accurate and early response can be implemented. In the same way, current research in different fields has illustrated how using hybrid AI is beneficial for detecting diseases and screening inhibitors. According to Alshammari et al. (2023), the optimized deep model showed great ability in recognizing olive leaf diseases. Even though agriculture is the topic, the way they use deep learning and optimization explains how this can improve the performance of image-based classification. It was demonstrated in their study that using optimization in CNN models better predicts diseases, making it possible to detect early cases of childhood diarrhea in health data. Mathematically, the optimization of CNN parameters (weights `w`) can be expressed as:

                  (1)

Where L is the loss function (e.g., cross-entropy), y is the ground truth, and f(x; w) is the CNN prediction. In the same manner, Asmare et al. (2022) [2] applied a QSAR model that combined deep learning for discovering IMPDH inhibitors in Cryptosporidium parvum through virtual screening. Even though this study is mainly about biomedicine, it proves how using structured biological signals and hybrid neural networks helps obtain accurate results. Using many different features in descriptors matches the practice of using various health indicators in public health databases, as UNICEF does. Therefore, Asmare and Yun (2024) [3] went on to apply an e-pharmacophore model based on deep learning to speedily scan a large range of molecules, locating substances that block CDPK1 from Cryptosporidium parvum. They showed how the use of pharmacophore modeling and deep learning can help identify difficult patterns of interaction. This proves the usefulness of hybrid models when it comes to making predictions in health care using structure data.Bautista-Castillo et al. (2025) [4] described a deep learning tool called AI-MET that aims to tell MIS-C apart from endemic typhus, as these illnesses present with similar early symptoms. Structured clinical indicators were part of the approach they used to train neural networks that were very good at spotting these two illnesses.

 Formally, the decision support system can be described as a classifier: 

Where W and b are trainable parameters, and x is the input feature vector of clinical indicators. Chen et al. (2023) [5] proposed a clinical narrative-aware deep neural network aimed at predicting critical

outcomes in emergency department (ED) settings. Their model uniquely combined structured electronic health records (EHR) with unstructured clinical narratives, thereby improving outcome prediction in real-time emergency scenarios. In 2025, Chidziwisano et al. [6] developed a poultry disease prediction model using pictures taken from poultry populations. What they did clearly showed that CNNs can identify disease patterns effectively from things people can see. A time-series analysis using data was performed by Domun et al. (2019) [7] to predict healthcare concerns in pigs such as tail-biting, fouling, and diarrhea. It is remarkable that they focus on discovering diarrhea through observation and the study of the environment. A general time-series prediction model can be represented as:

Where y_t is the predicted outcome at time t, and x_{t-k} are previous time-step features. Elhaminia and colleagues (2025) [8] introduced a way to use imaging, radiation data, and medical information to expect bowel toxicity in patients facing pelvic radiotherapy. Fatima and Unsal (2024) [9] carried out a geospatial study of possible causes of diarrhea in children under five in Pakistan by using Multiscale Geographically Weighted Regression (MGWR). The MGWR model can be expressed as:

                              (4)

Where (u_i, v_i) are coordinates for location i, and β_k are local coefficients for predictor x_k. In 2022, Ghaffari et al. [10] used deep convolutional neural networks (CNNs) to discover diarrhea and respiratory pathology in preweaning dairy calves by analyzing data taken from the animals’ milk feeders. Karaarslan et al. (2024) [11] created a model that uses deep learning and classical machine learning to classify and study diseases from sounds produced during breathing. In hybrid classification, the architecture can be formalized as:

                                                                      (5)
Where g could be a traditional classifier like SVM or Random Forest.

 

 

Figure 1: System design of hybrid deep learning for diarrhea prevalence prediction

3.1   Dataset Preprocessing and Visualization

The UNICEF Child Health Dataset available [19] is a valuable open-source dataset designed to support research and analysis related to child health outcomes, particularly focusing on the prevalence of diarrhea and malaria among children under the age of five. This dataset has been curated using UNICEF’s publicly available health survey data across multiple developing countries and regions. The primary purpose of this dataset is to monitor and assess the health status of children, particularly those vulnerable to infectious diseases. Analyze factors contributing to the prevalence of diarrhea and malaria, two major causes of child mortality in low- and middle-income countries. Whereas, an in-depth exploratory data analysis is conducted to understand the distribution and relationships of variables such aa statistical Summaries to measures such as mean, median, standard deviation, and interquartile ranges are calculated for continuous features. Frequency distributions are computed for categorical attributes. And histogram Plots are used to observe the distribution of numerical variables like child age, household wealth index, or birth weight.

 

Figure 2: Diarrhea Prevalence by Residence and Mother's Education Level

The figure 2 displays the distribution of diarrhea cases among children living in urban and rural areas. The x-axis represents the type of residence, while the y-axis shows the number of cases. Both urban and rural regions show equal numbers of diarrhea cases (~20 each), but the number of children not affected by diarrhea is consistently higher in both settings (~30 each). Although diarrhea prevalence appears similar across urban and rural areas in this sample, the proportion of healthy children is relatively high in both. However, deeper analysis with real data may reveal geographic disparities due to sanitation or healthcare access. Whereas right hand side, plot explores how a mother's level of education correlates with diarrhea occurrence in her child. Categories include None, Primary, Secondary, and Higher education levels.Children of mothers with no formal education and higher education both exhibit a similar number of diarrhea cases (~10). Meanwhile, mothers with primary or secondary education show comparable trends.The consistent diarrhea prevalence across all education levels in this simulated data suggests no strong effect. However, in real-world datasets, higher maternal education is typically associated with better hygiene practices and healthcare knowledge, which tends to reduce the risk of diarrhea.

 

 

Figure 3: Correlation Analysis and Age-wise Diarrhea Trend

The left-hand side of the figure 2 shows the heatmap visualization for the pairwise Pearson correlation coefficients between selected numerical and binary features such as Age, Wealth_Index, Sanitation_Binary, Diarrhea_Binary, Urban_Binary, and Female_Binary. Positive correlations are shown in red, and negative correlations in blue. Age and Wealth_Index show a strong positive correlation (0.93), indicating that older children in the dataset are more likely from wealthier households. The right-hand side of the figure 3 shows the line plot which illustrates how the average rate of diarrhea occurrence varies with child age. The x-axis represents age in months, and the y-axis shows the proportion of children with diarrhea at each age group.

3.3 Feature Engineering and Selection

The figure 4 represents the contribution of each feature to the prediction task as learned by a machine learning model such as a Random Forest Classifier or Gradient Boosting Machine.

The importance score for each feature fi is typically calculated as:

                       (6)

Where T is the total number of decision trees and Δi_t(fi) is the decrease in impurity from splits involving feature fi in tree t.

Figure 4: Feature Ranking Using Importance and Selection Metrics

Whereas, feature selections show the shows the statistical relevance of features based on univariate analysis techniques such as ANOVA F-test or Mutual Information (MI). For ANOVA F-test, the score for each feature fi is calculated as
For Mutual Information:

Where fi is a feature and y are the target variable (diarrhea status). And together, feature importance and selections confirm the consistency and reliability of features like Sanitation, Age, and Access to Healthcare, reinforcing their inclusion in final model development.

3.4   Deep Learning and Hybrid Model Implementation

To accurately predict the severity of childhood diarrhea prevalence using structured health indicators from the UNICEF dataset, a wide spectrum of deep learning models was implemented, ranging from traditional architectures such as CNN, LSTM, GRU, MLP, FCNN, and RNN to more advanced hybrid configurations [12-15]. Among these, five hybrid models consistently demonstrated superior performance across all evaluation metrics, making them the cornerstone of this study [16-18]. These hybrid models were designed to capture complex temporal and spatial relationships within the structured health indicator dataset. The Hybrid CNN-RNN model combines spatial feature extraction via CNN with temporal sequence modeling through RNN. The Hybrid RNN-GRU model integrates traditional RNN memory with gated GRU mechanisms to refine temporal flow. The Hybrid RNN-LSTM architecture connects short-term memory captured by RNN with long-term retention provided by LSTM units. The Hybrid GRU-LSTM model processes input sequentially through GRU followed by LSTM layers to achieve maximum memory context. Fully connected layers are stacked together with MLP blocks in the Hybrid FCNN-MLP to boost the learning of dense features. Most importantly, the Hybrid RNN-MLP model managed to outperform the others, thanks to RNN for pattern learning and MLP for its decision boundaries.

 

3.4.1 Proposed Algorithm Hybrid RNN-MLP for Childhood Diarrhea Classification

To manage the difficulties in predicting childhood diarrhea based on structured health indicators, the study creates a new Hybrid RNN-MLP architecture. A combination of the sequence modeling of Recurrent Neural Networks and the decision-making skills of Multilayer Perceptrons is achieved in the architecture. Because of this design, the model can pick up relationships between pieces of data and understand patterns in health-related tables. Because the model performed well in predicting four levels of diarrhea, it was selected as the base of the prediction system. The purpose of this type of hybrid model is to help predict whether childhood diarrhea occurs in low, medium, high, or very high rates based on United Nations Children’s Fund (UNICEF) health indicator data.

Algorithm 1: Hybrid RNN-MLP Framework for Childhood Diarrhea Classification

 

 

 

 

 

 

 

 

 

 

 

The results from this experiment prove that combining deep learning models leads to accurate predictions of the rate of diarrhea among children based on structured health indicators. Various models used in this study have demonstrated that adding deep learning techniques greatly helps predict outcomes in public health data. Among all the used models, the CNN performed exceptionally well, having a 100% validation accuracy, almost no validation loss, and 100% precision, recall, and F1-scores in every diarrhea category. The performance demonstrates that CNN learns from spatial patterns created from numbers that have been transformed into the right input forms. In addition, LSTM and MLP models proved effective by achieving accuracy greater than 99%, which shows they can identify related data patterns both sequentially and nonlinearly. TabNet also demonstrated this because it is very precise and justifies the significance of various features, making it a reliable source of information that the decision-makers in the healthcare sector can use. Generally, the hybrid models of RNN-MLP, GRU-LSTM, and CNN-RNN proved to be superior to the single models. They leverage the capabilities of every part: CNNs, short and detailed patterns and RNNs, series of successive events in time. This form of blending led to better performance and accuracy as can be observed in their almost perfect evaluations by all measures of evaluations such as AUC-ROC. In all of the models, the report indicated that their findings were balanced since none of the groups such as Minimal, Mild, Moderate, and Severe Diarrhea Prevalence had a significant decline in recall or precision. Based on this, it may be known that the models were not biased on the larger class and they were well suited to accommodate different classes sizes once they have been subjected to appropriate processing and normalization. The feature scaling, the process of eliminating redundant elements of the data, and the addition of Gaussian noise significantly enhanced the work of this model. The computation of the importance of each feature ensured that only the important features were retained and that the advantages of the high-speed training and reduced overfitting rates were enjoyed. The models are capable of forecasting even significant events on diarrhea and this offers scientists an effective tool of early supportive intervention in the face of scarce resources. High-risk areas can be identified early enough and as a result, medical care can be provided, clean water made available, nutritional programs launched, and educating activity arranged in the areas. In addition, the application of open UNICEF data also shows that the latter AI solutions may be applied to more health indicators or detect a range of diseases.

 

The paper presents a quantitative and descriptive method of predicting childhood diarrhea using the detailed health indicators suggested by UNICEF. In the framework, the prevalence of diarrhea is categorized into four classes: Minimal, Mild, Moderate and Severe due to preprocessing, feature extraction and improvement of model accuracy. Most of the different architectures, including CNN, LSTM, GRU, RNN, FCNN, and MLP, had been evaluated and their combinations with other architectures examined, like CNN-RNN, LSTM-GRU, and RNN-MLP. Overall, it happened that hybrid models could maintain their performance at a high level, and many of them could achieve almost perfect accuracy, precision, recall, and F1-scores.In particular, CNN and hybrid RNN-MLP models showed great results, showing that using spatial and sequential feature extraction helps in health prediction. Hybrid deep learning proves to be able to predict problems with diarrheal diseases, and the information found in public health data can be used to spot and monitor these diseases at an early stage. The framework developed in this work helps authorities in government, NGOs, and public health organizations by allowing them to easily spot high-risk places, use their resources well, and select the best approaches to intervention. Although the research indicates hybrid deep learning techniques can predict diarrhea by using specific routinely collected data, there are still ways to make these models work better in the real world. Going forward, it can be used for functions such as mobile health applications and those that monitor health in the community. These systems could provide instant risk predictions based on localized data inputs, helping frontline workers take quicker preventive actions.

 

Declarations and Statement

Conflict of interest: The authors declare that they have no conflict of interest.

Funding: Not applicable

Ethics Approval and Consent to Participate: This article does not contain any studies with human participants or animals performed by any of the authors.

Informed Consent: Not applicable

1.     Alshammari, Hamoud H, Ahmed I Taloba, and Osama R Shahin. 2023. “Identification of Olive Leaf Disease through Optimized Deep Learning Approach.” Alexandria Engineering Journal 72: 213–24. https://doi.org/https://doi.org/10.1016/j.aej.2023.03.081.

2.     Asmare, Misgana Mengistu, Nitin Nitin, Soon-IL Yun, and Rajani Kanta Mahapatra. 2022. “QSAR and Deep Learning Model for Virtual Screening of Potential Inhibitors against Inosine 5’ Monophosphate Dehydrogenase (IMPDH) of Cryptosporidium Parvum.” Journal of Molecular Graphics and Modelling 111: 108108. https://doi.org/https://doi.org/10.1016/j.jmgm.2021.108108.

3.     Asmare, Misgana Mengistu, and Soon-Il Yun. 2024. “E-Pharmacophore and Deep Learning Based High Throughput Virtual Screening for Identification of CDPK1 Inhibitors of Cryptosporidium Parvum.” Computational Biology and Chemistry 112: 108172. https://doi.org/https://doi.org/10.1016/j.compbiolchem.2024.108172.

4.     Bautista-Castillo, Abraham, Angela Chun, Tiphanie P Vogel, and Ioannis A Kakadiaris. 2025. “AI-MET: A Deep Learning-Based Clinical Decision Support System for Distinguishing Multisystem Inflammatory Syndrome in Children from Endemic Typhus.” Computers in Biology and Medicine 188: 109815. https://doi.org/https://doi.org/10.1016/j.compbiomed.2025.109815.

5.     Chen, Min-Chen, Ting-Yun Huang, Tzu-Ying Chen, Panchanit Boonyarat, and Yung-Chun Chang. 2023. “Clinical Narrative-Aware Deep Neural Network for Emergency Department Critical Outcome Prediction.” Journal of Biomedical Informatics 138: 104284. https://doi.org/https://doi.org/10.1016/j.jbi.2023.104284.

6.     Chidziwisano, George, Eric Samikwa, and Chisomo Daka. 2025. “Deep Learning Methods for Poultry Disease Prediction Using Images.” Computers and Electronics in Agriculture 230: 109765. https://doi.org/https://doi.org/10.1016/j.compag.2024.109765.

7.     Domun, Yuvraj, Lene Juul Pedersen, David White, Olutobi Adeyemi, and Tomas Norton. 2019. “Learning Patterns from Time-Series Data to Discriminate Predictions of Tail-Biting, Fouling and Diarrhoea in Pigs.” Computers and Electronics in Agriculture 163: 104878. https://doi.org/https://doi.org/10.1016/j.compag.2019.104878.

8.     Elhaminia, Behnaz, Alexandra Gilbert, Andrew Scarsbrook, John Lilley, Ane Appelt, and Ali Gooya. 2025. “Deep Learning Combining Imaging, Dose and Clinical Data for Predicting Bowel Toxicity after Pelvic Radiotherapy.” Physics and Imaging in Radiation Oncology 33: 100710. https://doi.org/https://doi.org/10.1016/j.phro.2025.100710.

9.     Fatima, Munazza, and Ömer Ünsal. 2024. “Geospatial Analysis of Diarrhoea Determinants among Children under Five in Pakistan Using Multiscale Geographically Weighted Regression (MGWR).” Applied Geography 172: 103434. https://doi.org/https://doi.org/10.1016/j.apgeog.2024.103434.

10.   Ghaffari, M. H., A. Monneret, H. M. Hammon, C. Post, U. Müller, D. Frieten, C. Gerbert, G. Dusel, and C. Koch. 2022. “Deep Convolutional Neural Networks for the Detection of Diarrhea and Respiratory Disease in Preweaning Dairy Calves Using Data from Automated Milk Feeders.” Journal of Dairy Science 105 (12): 9882–95. https://doi.org/10.3168/jds.2021-21547.

11.   Karaarslan, Oğuzhan, Kristen Dominica Belcastro, and Onur Ergen. 2024. “Respiratory Sound-Base Disease Classification and Characterization with Deep/Machine Learning Techniques.” Biomedical Signal Processing and Control 87: 105570. https://doi.org/https://doi.org/10.1016/j.bspc.2023.105570.

12.   Keshavamurthy, Ravikiran, Samuel Dixon, Karl T Pazdernik, and Lauren E Charles. 2022. “Predicting Infectious Disease for Biopreparedness and Response: A Systematic Review of Machine Learning and Deep Learning Approaches.” One Health 15: 100439. https://doi.org/https://doi.org/10.1016/j.onehlt.2022.100439.

13.   Kucukkara, Zeki, Ilker Ali Ozkan, Sakir Tasdemir, and Onur Ceylan. 2025. “Classification of Chicken Eimeria Species through Deep Transfer Learning Models: A Comparative Study on Model Efficacy.” Veterinary Parasitology 334: 110400. https://doi.org/https://doi.org/10.1016/j.vetpar.2025.110400.

14.   Li, Yiming, Deepthi Viswaroopan, William He, Jianfu Li, Xu Zuo, Hua Xu, and Cui Tao. 2025. “Improving Entity Recognition Using Ensembles of Deep Learning and Fine-Tuned Large Language Models: A Case Study on Adverse Event Extraction from VAERS and Social Media.” Journal of Biomedical Informatics 163: 104789. https://doi.org/https://doi.org/10.1016/j.jbi.2025.104789.

15.   Linu Babu, P., and S. Jana. 2025. “Gastrointestinal Tract Disease Detection via Deep Learning Based Duo-Feature Optimized Hexa-Classification Model.” Biomedical Signal Processing and Control 100 (PC): 106994. https://doi.org/10.1016/j.bspc.2024.106994.

16.   Liu, Tao, Yuanyuan Zhao, Yi Sun, Junchan Wang, Zhaosheng Yao, Chen Chen, Xiaochun Zhong, Shengping Liu, Chengming Sun, and Tao Li. 2024. “High-Throughput Identification of Fusarium Head Blight Resistance in Wheat Varieties Using Field Robot-Assisted Imaging and Deep Learning Techniques.” Journal of Cleaner Production 480: 144024. https://doi.org/https://doi.org/10.1016/j.jclepro.2024.144024.

17.   Luo, Xiao, Priyanka Gandhi, Zuoyi Zhang, Wei Shao, Zhi Han, Vasu Chandrasekaran, Vladimir Turzhitsky, et al. 2021. “Applying Interpretable Deep Learning Models to Identify Chronic Cough Patients Using EHR Data.” Computer Methods and Programs in Biomedicine 210: 106395. https://doi.org/https://doi.org/10.1016/j.cmpb.2021.106395.

18.   Maheswari, S, S Suresh, and S Ahamed Ali. 2024. “A Systematic Literature Review on Machine Learning and Deep Learning-Based Covid-19 Detection Frameworks Using X-Ray Images.” Applied Soft Computing 166: 112137. https://doi.org/https://doi.org/10.1016/j.asoc.2024.112137.

19.   UNICEF, “Multiple Indicator Cluster Surveys (MICS),” United Nations Children’s Fund, 2023. [Online]. Available: https://mics.unicef.org/