Predictive modeling for infectious diarrheal disease in pediatric populations: A systematic review

Abstract

Introduction

Diarrhea is still a significant global public health problem. There are currently no systematic evaluation of the modeling areas and approaches to predict diarrheal illness outcomes. This paper reviews existing research efforts in predictive modeling of infectious diarrheal illness in pediatric populations.

Methods

We conducted a systematic review via a PubMed search for the period 1990–2021. A comprehensive search query was developed through an iterative process and literature on predictive modeling of diarrhea was retrieved. The following filters were applied to the search results: human subjects, English language, and children (birth to 18 years). We carried out a narrative synthesis of the included publications.

Results

Our literature search returned 2671 articles. After manual evaluation, 38 of these articles were included in this review. The most common research topic among the studies were disease forecasts 14 (36.8%), vaccine‐related predictions 9 (23.7%), and disease/pathogen detection 5 (13.2%). Majority of these studies were published between 2011 and 2020, 28 (73.7%). The most common technique used in the modeling was machine learning 12 (31.6%) with various algorithms used for the prediction tasks. With change in the landscape of diarrheal etiology after rotavirus vaccine introduction, many open areas (disease forecasts, disease detection, and strain dynamics) remain for pathogen‐specific predictive models among etiological agents that have emerged as important. Additionally, the outcomes of diarrheal illness remain under researched. We also observed lack of consistency in the reporting of results of prediction models despite the available guidelines highlighting the need for common data standards and adherence to guidelines on reporting of predictive models for biomedical research.

Conclusions

Our review identified knowledge gaps and opportunities in predictive modeling for diarrheal illness, and limitations in existing attempts whilst advancing some precursory thoughts on how to address them, aiming to invigorate future research efforts in this sphere.

1. INTRODUCTION

Diarrhea is a global public health problem causing approximately 1.7 billion episodes in children annually, ¹ , ² with a significant proportion of diarrheal morbidity and mortality occurring in low‐income countries due to resource and infrastructural challenges to the health system. ³ , ⁴ Diarrheal disease, if not well managed, has a plethora of poor outcomes; malnutrition, longer duration diarrhea, dehydration, and even death. ⁵ , ⁶ Diarrhea is the second leading cause of mortality in children aged <5 years globally, causing about 1.5 million deaths a year, translating into nearly one in every nine child deaths. ⁷ , ⁸

The goal of predictive modeling in healthcare is to identify the likelihood of health events in patients or the population and such efforts guide healthcare providers and policy makers in making preventive strategies and local interventions, which ultimately reduce morbidity and mortality as well as the associated economic and social burden. ⁹ , ¹⁰ The early identification of diarrheal disease and its progression, in the backdrop of limited diagnostic capabilities in the developing world and budgetary constraints, is a key step in assessing and treating diarrheal illness in a cost‐effective manner, while improving quality of care and averting poor outcomes. Predictive models that are data‐driven can synthesize clinical, administrative, and socio‐economic data and complement clinician judgment while providing additional insights. Additionally, the prediction of disease incidence, seasonality, and outbreaks could help public health authorities to understand transmission dynamics and seasonal patterns in advance and aid them in planning and selecting the main response actions thereby enhancing the efficacy and timeliness of local interventions and preventive strategies translating to reduced morbidity and mortality. ¹¹ , ¹² , ¹³

Despite the advancement of predictive modeling in healthcare, there is currently no systematic evaluation, to the best of our knowledge, of the modeling areas and approaches in diarrheal disease. This paper, therefore, focuses on giving a synopsis of existing attempts to predict infectious diarrheal disease in children aged <18 years and their shortcomings. Additionally, we seek to identify topical and methodological gaps in knowledge that can be built upon to ameliorate predictive modeling for infectious diarrheal disease in pediatric populations.

2. QUESTION(S) OF INTEREST

What are the topical and methodological knowledge gaps in in predictive modeling for diarrheal disease in pediatric populations?

3. METHODS

Our study was conducted in line with the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines. ¹⁴ The systematic review protocol for this study was developed and iteratively refined with inputs from all study co‐authors. The protocol was registered with the international prospective register of systematic reviews (PROSPERO) (registration number: CRD42021241479): an online repository of systematic review protocols, maintained by the Center for Reviews and Dissemination at the University of York.

We developed a search strategy, limited to PubMed, to retrieve publications on predictive modeling of diarrheal illness between January 1, 1990 and December 31, 2021. The initial search strategy used was: Diarrhea AND (Predict OR Forecast). Two independent reviewers (Billy Ogwel and Gabriel Otieno) evaluated the title and abstract of each retrieved publication to determine the probable pertinence. Pertinence was assessed based on pre‐defined inclusion criteria ensuring that the publication's primary focus addressed some aspect of predictive modeling of diarrheal illness. We included all primary epidemiological study designs conducted in any country. We excluded literature that did not undergo a peer review: gray literature, conference proceedings, dissertations, and posters. The full texts of potentially relevant articles were further examined to make a decision on their inclusion.

Citations of included publications were also reviewed and if a citation was found pertinent but missing from the original search results, the search query was expanded by adding a keyword phrase extracted from the citation so as to include the relevant citation and other similar articles. Additionally, specific keyword phrases were used to narrow our search results by excluding irrelevant articles. The above process was iterated until a comprehensive search was developed.

The final search query used was: (Diarrh* OR gastroenteritis) AND (Predict* OR Forecast OR “risk scoring”) NOT (necrotizing OR appendicitis OR cancer OR ulcer OR IBD OR “Inflammatory bowel disease” OR colitis OR “Eosinophilic Esophagitis” OR Crohn). The following filters were applied to the search results: human subjects, English language, and children (birth to 18 years). The final literature review was done on articles that met the pre‐defined inclusion criteria (prediction of any aspect of infectious diarrhea in pediatric populations). Risk of bias and quality of the studies were not evaluated as this paper was descriptive in nature and no inferences were made based on the validity of the estimates of the performance metrics of the included studies. Difference of opinion about inclusion of articles were resolved by discussions among the two reviewers and where necessary a third reviewer (Bryan O. Nyawanda). One reviewer (Billy Ogwel) extracted the following information from the included articles: aspect of diarrheal illness, country, data used in modeling, modeling technique, time period, and performance of the predictive models. In terms of data analysis, we carried out a narrative synthesis of included publications.

4. RESULTS

A total of 2671 articles from the PubMed search were reviewed. Fifty nine of these appeared to be pertinent after review of titles and abstracts, and underwent full‐text review, where after 38 were established to be relevant and are discussed in this paper (Figure 1). Majority of the studies were conducted in Asia 17 (44.7%) ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ , ²² , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ , ²⁹ , ³⁰ , ³¹ with 8 (47.1%) of them conducted in Bangladesh alone. Eight studies (21.1%), 4 (10.5%), and 2 (5.3%) reported data from Africa, ³² , ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ Americas, ⁴⁰ , ⁴¹ , ⁴² , ⁴³ and Europe, ⁴⁴ , ⁴⁵ respectively. Seven studies (17.4%) were multi‐site ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹ , ⁵² of which 4 (57.1%) were reported from the Global Enteric Multicenter Study (GEMS). ⁵³ The most common research topics among the studies were disease forecasts 14 (36.8%), vaccine‐related 9 (23.7%), and disease/pathogen detection 5 (13.2%). Majority of these studies were published between 2011 and 2020, 28 (73.7%). The most common technique used in the modeling was machine learning 12 (31.6%) with various algorithms used for the prediction tasks (Table 1).

FIGURE 1.

Flowchart of included studies of predictive modeling in diarrheal disease.

TABLE 1.

Characteristics of included studies in predictive modeling for diarrheal disease.

	N = 38
Characteristics	n (%)
Region
Asia	17 (44.7)
Africa	8 (21.1)
Americas	4 (10.5)
Europe	2 (5.3)
Multi‐site	7 (18.4)
Topics
Disease forecasts	14 (36.8)
Vaccine‐related	9 (23.7)
Disease/pathogen detection	5 (13.2)
Outcomes	4 (10.5)
Strain dynamics	3 (7.9)
Determinants of diarrheal disease burden	2 (5.3)
Seasonality	1 (2.6)
Year of publication
1990–2000	1 (2.6)
2001–2010	7 (18.4)
2011–2020	28 (73.7)
2021	2 (5.3)
Modeling techniques
Machine learning algorithms	12 (31.6)
SIS‐/SIRS‐like compartmental models a	6 (15.8)
Dynamic, deterministic compartmental models	5 (13.2)
Logistic regression	4 (10.5)
Auto regression Model	3 (7.9)
Multiplicative Holt‐Winters method	1 (2.6)
Linear regression	1 (2.6)
Fourier analysis	1 (2.6)
Fitness models	1 (2.6)
Data assimilation: ensemble Kalfman filter	1 (2.6)
Gravity models	1 (2.6)
Multiple regression models	1 (2.6)
Spatially‐explicit stochastic model	1 (2.6)

^a SIS‐ (susceptible‐infectious‐susceptible)/SIRS‐like (susceptible‐infectious‐recovered‐susceptible) compartmental models.

4.1. Predicting disease forecasts

Much of the existing work on predictive modeling of diarrhea (14 [36.8%]) is focused on predicting diarrheal disease forecasts with five studies focusing on all‐cause diarrhea, ¹⁵ , ¹⁶ , ¹⁷ , ³² , ³³ while nine were specific to cholera. ²³ , ²⁴ , ²⁵ , ²⁹ , ³⁰ , ³¹ , ⁴⁰ , ⁴¹ , ⁴² The prediction of rotavirus‐specific incidence have been incorporated in vaccine impact predictions discussed under the sub‐topic vaccine‐related predictions.

The studies that predicted forecasts of all‐cause diarrhea used data collected from China, ¹⁵ , ¹⁷ Indonesia, ¹⁶ Mali, ³² and Botswana. ³³ The modeling techniques used in these predictions were Machine learning (Multiple Linear Regression; Random Forest [RF] Regression; Support Vector (SVM) Regression; Gradient Boosting Regression; Convolutional Neural Network; Neural Network (NN) Regression), Auto regression models (autoregressive integrated moving average [ARIMA/X]; seasonal‐auto‐regressive‐integrated‐moving‐average [SARIMA/X]), compartmental susceptible‐infected‐recovered‐susceptible (SIRS) model, multiplicative Holt‐Winters method, and parsimonious model (PM; Table 2). The data used in the above modeling included: morbidity, meteorological/climate and search indices data. The results from the studies varied: Fang et al. ¹⁵ reported that the RF model outperformed the ARIMA/X models with a mean absolute percentage error (MAPE) of approximately 20.0%; Medina et al. ³² realized a MAPE circa 25.0% using the multiplicative Holt‐Winters method; Pangestu et al. ¹⁶ reported an accuracy of 78.6% on the SARIMA model; Wang et al. ¹⁷ found the PM model to outperform all the other models in three metrics; the SIRS model built by Heaney et al. ³³ had a mean Root Mean Square Error (RMSE) and correlation of 0.79 and 0.99, respectively between the observations and simulations across all wet season outbreaks, while across dry season outbreaks, the mean RMSE and correlation were 1.33 and 0.99, respectively (Table 3).

TABLE 2.

Categorization and current status of research topics on predictive modeling for diarrheal disease.

Primary Category	Sub‐category	Modeling methods for existing predictive models
Disease forecast	All‐cause diarrhea	Random Forest, autoregressive integrated moving average (ARIMA/X), seasonal‐auto‐regressive‐integrated‐moving‐average (SARIMA/X), multiplicative Holt‐Winters method, compartmental susceptible‐infected‐recovered‐susceptible (SIRS) model, Parsimony Model, gravity models, Multiple Linear Regression, Random Forest Regression, Support Vector Regression, Gradient Boosting Regression, Extreme Gradient Boosting Regression, Convolutional Neural Network, Neural Network Regression
	Cholera	SIRS‐like models, data assimilation: ensemble Kalfman filter, individual‐based spatially‐explicit stochastic model, logistic regression, SARIMA, model, auto regression model, multiple regression models
Disease/pathogen detection	All‐cause diarrhea	Naïve Bayes, linear discriminant analysis, quadratic discriminant analysis, support vector machine, Artificial Neural Network
	Viral etiology	Random Forest, logistic regression
	Bacterial etiology	Random Forest, logistic regression
	Rotavirus	Classification trees
Strain dynamics	Rotavirus	Fourier analysis
	Norovirus	Fitness models
	Shigella	Logistic regression, Neural Network, support vector machines
Outcomes	Dehydration	Logistic regression/recursive partitioning model
	Malnutrition	Linear regression
	Hospitalization	None
	Prolonged/persistent diarrhea	None
	Mortality	None
Seasonality		Principal‐Component Analysis, K‐means clustering, classification and regression trees
Vaccine	Vaccine impact	SIS‐ (susceptible‐infectious‐susceptible), SIRS‐like compartmental models, ensemble models, dynamic, deterministic compartmental model, periodic regression models, age‐structured compartmental mode
	Vaccine cost‐effectiveness	Dynamic model
	Vaccine hesitancy	Logistic regression, Random Forest, and Neural Networks
Determinants of diarrheal disease burden		Classification and Regression Trees (CART)

TABLE 3.

Comprehensive list of included studies in predictive modeling for diarrheal disease.

Authors	Year of publication	Time period of the study	Country	Aspect of diarrheal disease	Technique used for predictive modeling	Data used (studied variables)	Performance of model
Pascual et al.	2008	1966–2005	Bangladesh	Cholera incidence/outbreaks	TSIR or TSIRS model	Laboratory‐confirmed cholera infection data, year and month of infection, climate data	The authors observed that lack of extreme events between 2001 and 2005 would have been anticipated with 75% confidence half a year ahead with a model fitted to data up to 2000.
Pasetto et al.	2017	2010	Haiti	Cholera incidence/outbreaks	Data assimilation: ensemble Kalfman filter	Laboratory‐confirmed cholera infection data, year and month of infection, rainfall data	The authors showed that the assimilation procedure with the sequential update of the parameters outperformed calibration schemes based on Markov chain Monte Carlo. Moreover, in a forecasting mode, the model predicted the spatial incidence of cholera at least 1 month ahead.
Bertuzzo et al.	2016	2010–2017	Haiti	Cholera incidence/outbreaks	Individual‐based spatially‐explicit stochastic model	Epidemiological dynamics and health‐care practice data	The model captured the timing and the magnitude of the peaks correctly (Nash–Sutcliffe index = 0.79). The authors showed that the probability that the epidemic would go extinct before the end of 2016 was of the order of 1%.
Jutla et al.	2015	2002–2010	Bangladesh	Cholera incidence/outbreaks	Logistic regression	River discharge data, terrestrial water storage (TWS) data, cholera prevalence data	The authors observed that TWS representing had an asymmetrical, strong association with cholera prevalence in the spring (τ = −0.53; P < 0.001) and autumn (τ = 0.45; P < 0.001) up to 6 months in advance.
Daisy et al.	2020	2000–2013	Bangladesh	Cholera incidence/outbreaks	Seasonal‐auto‐regressive‐integrated‐moving‐average (SARIMA) model	Cholera incidence and climatic variables	Root Mean Square Error (RMSE) = 14.7; mean absolute error (MAE) = 11.
Bengtsson et al.	2015	2010	Haiti	Cholera incidence/outbreaks	Gravity models	Case data, mobility data	Area under the Curve (AUC) = 0.79 for mobile phone‐based model
Matsuda et al.	2008	1983–2002	Bangladesh	Cholera incidence/outbreaks	Auto‐regression model	Cholera patient data, climate data	The authors reported a Pearson's correlation coefficient of 0.95 between the monthly number of patients predicted by the model and the actual monthly number of patients.
Koepke et al.	2016	2000–2007; 2010–2013	Bangladesh	Cholera incidence/outbreaks	SIRS model	Cholera case data, environmental variables	The authors showed that their model successfully predicted an increase in the number of infected individuals in the population weeks before the observed number of cholera cases increased.
Jutla et al.	2013	1998–2010	Bangladesh	Cholera incidence/outbreaks	Multiple regression models	Cholera incidence, sewage discharge data, satellite environmental determinants	Accuracy = 75%
Levine et al.	2015	2014	Bangladesh	Dehydration	Logistic regression/recursive partitioning model	Historical, demographic, clinical, and nutritional data	The authors reported an AUC of 0.79 (95% confidence interval [CI]  =  0.74–0.84) for severe dehydration and 0.78 (95% CI  =  0.74–0.81) for some (any) dehydration for the new DHAKA Dehydration Score. Additionally, their score had a 90% agreement between independent raters, with a Cohen's Kappa of 0.75 (95% CI  =  0.66–0.85) among children with a repeat clinical exam.
Levine et al.	2013	2010–2012	Rwanda	Dehydration	Logistic regression	Demographic and clinical data	The authors reported AUCs of 0.72 (95% CI =  0.60–0.85), 0.73 (95% CI =  0.62–0.84), and 0.80 (95% CI =  0.71–0.89) for the WHO severe dehydration scale, CDC scale, and Clinical Dehydration Scale, respectively, in the full cohort. They also showed that only the Clinical Dehydration Scale was a significant predictor of severe disease when used in infants, with an AUC of 0.77 (95% CI =  0.61–0.93).
Zodpey et al.	1999	1996–1997	India	Dehydration	Logistic regression	Demographic and clinical data	The authors reported sensitivity, specificity, positive predictive value, Cohen's kappa, and overall predictive accuracy of 0.81, 0.81, 0.81, 0.61, and 0.86, respectively.
Alexander and Blackburn	2013	2006–2009	Botswana	Determinants of diarrheal disease burden	Cluster analysis/classification and regression trees	Hospital surveillance data
Green et al.	2009	200–2007	Global	Determinants of diarrheal disease burden	Classification and Regression Trees (CART)	WASH, government spending, literacy levels	Mean squared prediction error (MSE) of 0.225
Fang et al.	2020	2012–2016	China	Diarrhea incidence	Random Forest, autoregressive integrated moving average (ARIMA/X)	Morbidity and meteorological data	20% mean absolute percentage error (MAPE) with actual values; 30% MAPE between ARIMAX and ARIMA Model
Pangestu et al.	2020	2010–2019	Indonesia	Diarrhea incidence	Seasonal‐auto‐regressive‐integrated‐moving‐average (SARIMA/X)	Burden of disease estimates and climate data	Accuracy = 78.6%
Wang et al.	2020	2012–2016	China	Diarrhea Incidence	Parsimony Model (PM)/Multiple Linear Regression/Random Forest Regression/Support Vector Regression/Gradient Boosting Regression/Extreme Gradient Boosting Regression/Convolutional Neural Network/Neural Network Regression	Historical outpatient visit counts, meteorological factors (MF) and Baidu search indices (BSI)	The authors observed that the PM model obtained the best performance in terms of three metrics benefiting from MF and BSI data.
Medina et al.	2007	1996–2004	Mali	Diarrhea incidence/seasonality	Multiplicative Holt‐Winters method	Clinical data and climate data	MAPE circa 25%.
Heaney et al.	2020	2007–2017	Botswana	Diarrheal incidence/outbreaks	Compartmental susceptible‐infected‐recovered‐susceptible (SIRS) model	Hospital surveillance data	The authors reported that the average RMSE and correlation between the observations and simulations across all wet season outbreaks was 0.79 and 0.99, respectively. Similarly, they reported an average RMSE and correlation across dry season outbreaks as 1.33 and 0.99, respectively.
Maniruzzaman et al.	2020	2014	Bangladesh	Diarrheal infection	naïve Bayes/linear discriminant analysis/quadratic discriminant analysis/support vector machine	Demographic health survey	Support Vector Machine (SVM) with radial basis kernel yielded 65.61% accuracy, 66.27% sensitivity, and 52.28% specificity.
Abubakar and Olatunji	2019	2013	Nigeria	Diarrheal infection	Artificial Neural Network	Demographic and health survey data	High accuracy of 95.78 and 95.63% during training and testing phases
Brander et al.	2019	2007–2011	The Gambia, Mali, Mozambique, Kenya, Pakistan, India, Bangladesh	Malnutrition	Linear regression	Clinical, historical, anthropometric	AUC of 0.67 (95% CI = 0.64–0.69)
Suzuki et al.	2016	2006–2015	Japan	Norovirus strain dynamics	Fitness models	Sequence data, year and month of isolation	The authors showed that their model predicted GII.3 and GII.4 would contract, whereas GII.17 would expand and predominate in the 2015–2016 season.
Garbern et al.	2021		Mali and Bangladesh	Pathogen detection/clinical profiles/viral etiology	Random Forest/logistic regression	Clinical, historical, anthropometric and microbiologic data	AUC of 0.754 (0.665–0.843)
Brintz et al.	2020	2007–2011	The Gambia, Mali, Mozambique, Kenya, Pakistan, India, Bangladesh	Pathogen detection/clinical profiles/viral etiology/Bacterial etiology	Random Forest/logistic regression	Clinical, historical, anthropometric and microbiologic data	AUC = 0.825; specificity = 0.85; sensitivity = 0.59, negative predictive value (NPV) = 0.82; positive predictive value (PPV) = 0.64
Ayers et al.	2016	2007–2011	Kenya	Pathogen detection/clinical profiles/Rotavirus	Classification trees	Clinical, historical, anthropometric and microbiologic data	AUC = 0.816 on training: AUC = 0.6125 on test data
Pitzer et al.	2011	1985–2009*	Italy, Hungary, Spain, Japan, United States, Australia	Rotavirus strain dynamics	Fourier analysis	Laboratory‐confirmed rotavirus infection data, sequence data, vaccination	The authors showed that their model explained the coexistence and cyclical pattern in the distribution of genotypes observed in most developed countries: predominant rotavirus strains cycle with periods (T) ranging from 3 to 11 years
Chao et al.	2019	2007–2011	The Gambia, Mali, Mozambique, Kenya, Pakistan, India, Bangladesh	Seasonality	Principal‐Component Analysis/K‐means clustering	Microbiological and weather data	The authors observed that rotavirus was most prevalent during the drier “winter” months and out of phase with bacterial pathogens, which peaked during hotter and rainier times of year corresponding to “monsoon,” “rainy,” or “summer” seasons.
Adamker et al.	2018	2002–2015	Israel	Shigella species/Outcomes	Logistic Regression (LR), Neural Network (NN), and Support Vector Machines (SVM)	National Shigella data as collated by the Ministry of Health (MoH) Division of Epidemiology	Accuracy of 93.2% (Shigella species) and 94.9% (hospitalization)
Freiesleben de Blasio et al.	2014		Kazakhstan	Vaccine cost‐effectiveness	Dynamic model		The authors reported that a vaccination program with 90% coverage would prevent ≈880 rotavirus deaths and save an average of 54,784 life‐years for children <5 years of age. They also showed that Indirect protection accounted for 40% and 60% reduction in severe and mild rotavirus gastroenteritis, respectively
Bar‐Lev et al.	2021	2014–2018	Israel	Vaccine hesitancy	Logistic regression, Random Forest and Neural Networks	Demographic, clinical, socio‐economic data, vaccination, social media traffic	The authors observed that the performance of models for Rotavirus, Hepatitis A and Hepatitis B, were close to random (accuracy <0.63 and F1 < 0.65). Additionally, they reported a negative association between on‐line discussions and vaccination.
de Blasio et al.	2010	2005–2008	Kyrgyzstan	Vaccine impact	Deterministic age‐structured dynamic model	Key features of rotavirus epidemiology, rotavirus associated events (death, hospitalization, outpatient visits), vaccination	The authors reported that a routine rotavirus vaccination program at 95% coverage and 54% effectiveness against severe infection was estimated to lead to a 56% reduction in rotavirus‐associated deaths and a 50% reduction in hospital admissions, while outpatient visits and homecare episodes would decrease by 52% compared to baseline levels after 5 years of intervention.
Atchison et al.	2010	1998–2007	England and Wales	Vaccine impact/seasonality	Deterministic age‐structured dynamic model	Key features of rotavirus epidemiology, vaccination	The authors showed that their model reproduced the strong seasonal pattern and age distribution of rotavirus disease observed in England and Wales. Furthermore, they observed that their model predicted that vaccination would provide both direct and indirect protection within the population resulting in 61% reduction of rotavirus disease incidence.
Park et al.	2017	2009–2012	Niger	Vaccine Impact/transmission dynamics	Susceptible‐infected‐recovered (SIR)‐like compartmental models/Ensemble models	Clinic admissions data and healthcare seeking data	The authors reported that their model predicted the current burden of severe rotavirus disease to be 2.6%–3.7% of the population each year and that a two‐dose vaccine schedule achieving 70% coverage could reduce burden by 39%–42%.
Pitzer et al.	2012	1999–2009	England and Wales	Vaccine impact/transmission dynamics	SIS‐ (susceptible‐infectious‐susceptible)/SIRS‐like (susceptible‐infectious‐recovered‐susceptible) compartmental models	Laboratory‐confirmed rotavirus infection data	The authors showed that their models predicted that during the initial year after vaccine introduction, incidence of severe Rotavirus gastroenteritis (RVGE) would be reduced 1.8–2.9 times more than expected from the direct effects of the vaccine alone (28%–50% at 90% coverage), but over a 5‐year period following vaccine introduction severe RVGE would be reduced only by 1.1–1.7 times more than expected from the direct effects (54%–90% at 90% coverage). They also reported that projections for the long‐term reduction of severe RVGE ranged from a 55% reduction at full coverage to elimination with at least 80% coverage.
Effelterre et al.	2009		France, Germany, Italy, Spain and the United Kingdom	Vaccine impact/transmission dynamics	Dynamic, deterministic compartmental model	Burden of disease estimates: hospitalizations, emergency‐room visits and primary‐care visits	The authors reported that with vaccination coverage rates of 70%, 90%, and 95% their model predicted that, in addition to the direct effect of vaccination, herd protection induced a reduction in RV‐related gastroenteritis (GE) incidence of 25%, 22%, and 20%, respectively, for RV‐GE of any severity, and of 19%, 15%, and 13%, respectively, for moderate‐to‐severe RV‐GE, 5 years after implementation of a vaccination program.
Asare et al.	2020	2007–2015	Ghana	Vaccine impact/transmission dynamics	SIRS‐like model	Epidemiological data and vaccination data	The authors showed that their model captured the spatio‐temporal variations in rotavirus incidence across the three sites and showed good agreement with the age distribution of observed cases
Olson et al.	2020	2002–2016	United States	Vaccine impact/transmission dynamics	Periodic regression models/ age‐structured compartmental mode	Case data, Emergency department (ED) visits data, hospitalization data, vaccination data	The authors reported that their published mechanistic model qualitatively predicted patterns more than 2 years in advance.

The studies that focused on the prediction of cholera forecasts utilized data from two countries: six from Bangladesh ²³ , ²⁴ , ²⁵ , ²⁹ , ³⁰ , ³¹ and three from Haiti. ⁴⁰ , ⁴¹ , ⁴² The modeling techniques employed by these studies were SIRS models, auto regression models, logistic regression (LR), multiple regression, gravity models, spatially‐explicit stochastic model, and data assimilation: ensemble Kalfman filter. The modeling was based on the following data: morbidity, climate, environmental, mobility, sewage discharge, terrestrial water storage, and satellite data. The results were diverse: Jutla et al. ³¹ reported an accuracy of 75.0% on the multiple regression model; Bengtsson et al. ⁴² reported an Area Under the Curve (AUC) of 79.0% on the mobile phone‐based model; Pascual et al. ²³ got an accuracy of 75.0% on their SIRs model; Pasetto et al. ⁴⁰ reported that the assimilation procedure outperformed other calibration schemes and it was able to forecast the spatial incidence of cholera at least 1 month ahead; Bertuzzo et al. ⁴¹ using a spatially‐explicit stochastic model observed an overall good agreement with the data, capturing the timing and the magnitude of the peaks correctly (Nash–Sutcliffe index = 0.79); Daisy et al. ²⁵ achieved the best fit in the model using climate data (RMSE = 14.7, Mean of Absolute value of Errors [MAE] = 11); Matsuda et al. ²⁹ achieved a correlation of 0.95 between simulations and actual number of disease.

4.2. Predicting vaccine‐related topics

Rotavirus is a leading cause of diarrheal morbidity and mortality. ¹ , ⁵⁴ Consequently, rotavirus vaccines have been incorporated into national immunization programs following recommendations by the World Health Organization (WHO). ⁵⁵ Nine studies focused on predictions around rotavirus vaccine with seven studies focusing on vaccine impact, ²⁸ , ³⁵ , ³⁸ , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁷ one on vaccine hesitancy, ²⁷ and one on vaccine cost‐effectiveness ¹⁹ (Table 3).

The seven studies on vaccine impact used data from Niger, ³⁵ Ghana, ³⁸ United States, ⁴³ Kyrgyzstan, ²⁸ England and Wales, ⁴⁴ , ⁴⁵ and a multi‐site study (France, Germany, Italy, Spain, and the United Kingdom). ⁴⁷ These studies built predictive models using SIRs models, periodic regression models and deterministic age‐structured dynamic model (Table 2). The models used the following data: Key features of rotavirus epidemiology, rotavirus‐associated events (death, hospitalization, outpatient visits), vaccination, and healthcare seeking data. The results were comparable: Park et al. ³⁵ predicted a 39%–42% reduction on burden at a 70.0% vaccine coverage; Pitzer et al. ⁴⁴ predicted that incidence of severe disease would be reduced 1.1‐1.7 times more than expected from the direct effects (54%–90.0% at 90.0% coverage) 5 years after roll‐out; Effelterre et al. ⁴⁷ reported that with vaccination coverage rates of 70.0%, 90.0%, and 95.0% the model predicted that, in addition to the direct effect of vaccination, herd protection induced a reduction in disease incidence (any severity) of 25.0%, 22.0%, and 20.0%, respectively, and of 19.0%, 15.0%, and 13.0%, respectively, for moderate‐to‐severe disease, 5 years after roll‐out; De Blasio et al. ²⁸ predicted a reduction of 56.0%, 50.0%,52.0% for deaths, admissions, and outpatient visits, respectively, 5 years after vaccine roll‐out at a 95.0% coverage and 54.0% effectiveness; Atchison et al. ⁴⁵ predicted a 61.0% reduction in incidence (Table 3).

Freiesleben de Blasio et al. ¹⁹ predicted cost‐effectiveness of rotavirus vaccine in Kazakhstan using a dynamic model. They reported that at a 90.0% coverage, ≈880 rotavirus deaths would be averted and an average of 54,784 life‐years for children <5 years of age would be saved in a 20‐year period. Additionally, they found 40.0% and 60.0% reduction in severe and mild rotavirus gastroenteritis due to indirect protection, respectively. Bar‐Lev et al. ²⁷ predicted vaccine hesitancy in Israel using machine learning algorithms (LR, RF, and NN). The performances of all algorithms were close to random in this study.

4.3. Predicting disease/pathogen detection

Five studies focused on disease/pathogen detection: two all‐cause diarrhea ¹⁸ , ³⁹ ; one viral‐only etiology ⁴⁹ ; one viral/bacterial etiology ⁴⁸ ; one rotavirus. ³⁷ Other than the rotavirus model, there are no other existing pathogen‐specific prediction models. Two studies focused on the prediction of diarrheal infection. Both studies used machine learning algorithms in the modeling of demographic health survey data in Bangladesh ¹⁸ and Nigeria, ³⁹ respectively. Maniruzzaman et al. ¹⁸ reported the SVM model with radial basis kernel outperformed other models yielding 65.6% accuracy, 66.3% sensitivity, and 52.3% specificity. Abubakar and Olatunji ³⁹ reported an accuracy of 95.6% on their artificial neural network model.

Brintz et al. ⁴⁸ used data from the GEMS study to predict both viral and bacterial etiologies in diarrhea utilizing RF and LR models with the following results: AUC = 82.5%; specificity = 85.0%; sensitivity = 59.0%; Negative Predictive Value (NPV) = 82.0%; Positive Predictive Value (PPV) = 64.0%. Garbern et al. ⁴⁹ predicted viral‐only etiology deploying machine learning algorithms (RF and LR regression) on data from Mali and Bangladesh yielding 75.4% AUC. Ayers ³⁷ used classification trees on the GEMS Kenya data to predict rotavirus infection realizing 61.3% AUC.

4.4. Predicting diarrheal outcomes

There were only four studies focused on diarrheal outcomes; three on dehydration ²⁰ , ²¹ , ³⁶ and one on linear growth faltering. ⁵¹ Despite risk factors for death, ⁵⁶ , ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰ hospitalization, ⁶¹ , ⁶² , ⁶³ , ⁶⁴ and prolonged/persistent diarrhea ⁶⁵ , ⁶⁶ , ⁶⁷ , ⁶⁸ being documented, there were no predictive models developed on these outcomes. The three publications on dehydration utilized data from Bangladesh, ²⁰ Rwanda, ³⁶ and India. ²¹ The primary modeling technique in all the studies was LR with historical, demographic, clinical, and nutritional data being used. The new DHAKA Dehydration Score developed by Levine et al. ²⁰ yielded AUC of 79.0% and 78.0%, for severe and some dehydration, respectively. Levine et al. ³⁶ also reported AUC of 72.0%, 73.0%, and 80.0% for the WHO severe dehydration scale, Centers for Disease Control and Prevention (CDC) scale, and Clinical Dehydration Scale, respectively. In the model developed by Zodpey et al., ²¹ the sensitivity, specificity, PPV, Cohen's kappa, and overall predictive accuracy were 81.0%, 81.0%, 81.0%, 61.0%, and 86.0%, respectively. Brander et al. ⁵¹ used the GEMS data to predict linear growth faltering based on a LR model that yielded 67.0% AUC.

4.5. Other research topics

Strain dynamics was addressed in three studies with rotavirus, ⁵² norovirus, ²² and Shigella ²⁶ being the only pathogens investigated. Pitzer et al. ⁵² used Fourier analysis to predict rotavirus strain dynamics based on multi‐site data (Italy, Hungary, Spain, Japan, United States, Australia). Their model explained the coexistence and cyclical patterns in the distribution of genotypes observed in most developed countries: predominant rotavirus strains cycle with periods ranging from 3 to 11 years. Suzuki et al. ²² predicted norovirus strain dynamics in Japan using fitness models. They reported that the model was effective in predicting the direction of change in the proportions of genotypes, it predicted that GII.3 and GII.4 would contract, whereas GII.17 would expand and predominate in the 2015‐2016 season. Adamker et al. ²⁶ predicted Shigella species among Shigellosis patients in Israel using machine learning algorithms (LR, NN, and SVM) achieving an accuracy of 93.2%.

There were two studies on determinants of diarrheal disease. Alexander and Burn ³⁴ used cluster analysis and Classification and Regression Tree (CART) models to evaluate patient attributes by outbreak in Botswana. They identified two main clusters associated with patient age while neither village nor outbreak had an influence on cluster. Furthermore, CART identified sex and hospitalization as predictors of diarrhea. Additionally, water shortages and water quality deficiencies were identified in both outbreaks. Green et al. ⁵⁰ also used a CART model to establish determinants of diarrheal illness at a country‐level. Improvements in rural sanitation was identified as the most important predictor for reducing diarrhea. They predicted that a 65% global reduction in unmet rural sanitation would save 1.2 million lives annually.

Finally, Chao et al. ⁴⁶ used machine learning algorithms (Principal‐Component Analysis and K‐means clustering) to characterize seasonality of different pathogens from GEMS data. The key findings from the study was that rotavirus was most prevalent during the dry “winter” months and out of phase with bacterial pathogens, which peaked during hotter and rainier times of year corresponding to “monsoon,” “rainy,” or “summer” seasons.

5. DISCUSSION

Notable effort has been made in predictive modeling for infectious diarrhea in pediatric populations. The key findings from our review are: (1) Diarrheal outcomes remain under researched; (2) with the shift in the landscape of diarrheal pathogens post rotavirus vaccine introduction, many pathogen‐specific areas remain open for exploration including disease forecasts, disease detection strain dynamics and vaccine‐related predictions; (3) the need for model ensembles to gauge and mitigate structural uncertainties in predictions problems; (4) need for more diverse, neoteric, and pertinent data to improve the precision and robustness of prediction models; (5) the lack of consistency on reporting of predictive models despite available guidelines highlighting the need for common data standards need for common data standards and adherence to guidelines on reporting of predictive models for biomedical research.

Despite the advances that have been made in predictive modeling, a number of topical areas remain open for further research. Notwithstanding the wide array of poor diarrheal outcomes and their public health impact, only four studies exist focusing on dehydration and linear growth faltering. Death, prolonged/persistent diarrhea, and malnutrition still remain uninvestigated. Furthermore, with the change in the landscape of diarrheal etiologies after rotavirus vaccine introduction, ⁶⁹ , ⁷⁰ other pathogens (Escherichia coli, Cryptosporidium, and Shigella) have emerged as important etiological agents of moderate‐to‐severe diarrhea in these settings ² potentially creating the need for pathogen‐specific models in the areas of disease forecasts, disease detection, and strain dynamics. Additionally, with pipeline vaccines advancing for other enteric pathogens like cholera, Shigella, Enterotoxigenic E. coli, ⁷¹ , ⁷² vaccine‐related predictions would be useful for these vaccines in understanding the comprehensive epidemiological impact of vaccination and providing insights into the prospective cost‐effectiveness of vaccination by forecasting vaccine‐induced changes in the epidemiology of diarrheal disease in the population over time. ⁷³

Although diverse and dynamic modeling approaches have been used so far, a number of methodological propositions could help to further improve the predictive modeling for diarrheal disease. A diversity of model structures should be considered in order to gauge structural uncertainties in prediction/forecasting. To enhance the characterization of forecasting uncertainty, model ensembles that aggregate a range of model structures and their individual uncertainty should be increasingly embraced. ⁷⁴ In addition, since model parameters in forecasting may be subject to uncertainties, approaches such as Bayesian methodologies, which combine uncertainties and expert knowledge through choice of prior probabilities, have become prominent and could be utilized. ⁷⁵ Furthermore, with the complexity and rise of data in healthcare, ⁷⁶ , ⁷⁷ we propose the increased adoption of machine learning algorithms in prediction tasks as they have been shown to provide additional information and understanding of the data beyond using standard statistical approaches. ³⁷ , ⁷⁸ Finally, different models reported different metrics for the same prediction task as well as incompleteness in the reporting of elements of the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement and there is need to define common data standards and adherence to developed guidelines on reporting of predictive models in biomedical research ⁷⁹ to make comparisons across multiple models possible.

With regards to data sources for disease forecasts, adding more diverse, recent, and pertinent data will improve the precision and robustness of such models. To better predict new diarrheal events and to inform spatial spread, incorporating population density and spatial data into the modeling is vital. ⁸⁰ This can gain ascendancy with the availability of novel universal geospatial datasets such as WorldPop and LandScan. Furthermore, understanding spread of diarrhea could be better understood by integrating remote sensing and satellite data, healthcare capacity, and human mobility data in the models. Additionally, we recommend use of data from multi‐site studies, where available, as they are larger, richer, and have geographical variance that could help to mitigate algorithmic bias that arises from sample and prejudice biases. ⁸¹ More recent data are also needed in the prediction of seasonality of pathogens since rotavirus vaccine has been shown to affect seasonality of rotavirus gastroenteritis ⁸² and climate change caused by increased global warming has also been shown to alter the distribution of infectious disease vectors and the seasonal distribution of some allergenic pollen species. ⁸³ , ⁸⁴ Finally, post‐vaccination data from diverse countries is needed to improve projections of long‐term impact of rotavirus vaccination by conducting further validation. ⁴⁴

6. LIMITATION

There were several limitations with this review. First, by limiting this systematic review to PubMed, we may have missed pertinent articles not indexed within PubMed. Second, by filtering‐out articles not written in English, other predictive models for diarrheal illness may have been missed. Additionally, this review did not conduct risk of bias and quality assessment of the studies, hence bias in primary studies may have been reported leading to inaccuracy of findings. However, we did not draw any inference based on the validity of the estimates of the performance metrics of the included studies. Furthermore, the comparison of statistical measures of predictions could not be directly made across models as the outcomes being predicted were different.

7. CONCLUSION

As the first systematic review on predictive modeling for diarrheal illness, we observed substantial effort to predict various aspects of diarrheal disease. However, many topical and methodological problems remain open and there is significant scope for improvement in the predictive modeling of diarrhea. Future research in predictive modeling for diarrheal illness should seek to address them to realize more comprehensive, robust, and precise models.

AUTHOR CONTRIBUTIONS

Billy Ogwel and Bryan O. Nyawanda conceived the study. Billy Ogwel, Bryan O. Nyawanda, Richard Omore, Gabriel Otieno, and Vincent Mzazi contributed to study design and implementation. Billy Ogwel conducted literature search and review, and drafted the manuscript, and all authors critically reviewed the manuscript for intellectual content and approved the final manuscript. Bryan O. Nyawanda and Gabriel Otieno also participated in literature search. All authors read and approved the final manuscript.

FUNDING INFORMATION

The authors used their own resources.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

ETHICS STATEMENT

Not applicable.

Ogwel B, Mzazi V, Nyawanda BO, Otieno G, Omore R. Predictive modeling for infectious diarrheal disease in pediatric populations: A systematic review. Learn Health Sys. 2024;8(1):e10382. doi: 10.1002/lrh2.10382