Abstract The early detection of diabetic retinopathy remains a critical challenge in medical diagnostics, with deep learning techniques in artificial intelligence offering promising solutions for identifying pathological patterns in retinal images. This study evaluates and compares the performance of three convolutional neural network (CNN) architectures ResNet-18, ResNet-50, and a custom, non-pretrained CNN using a dataset of retinal images classified into five categories. The findings reveal significant differences in the models’ ability to learn and generalize. The non-pretrained CNN consistently outperformed the pretrained ResNet-18 and ResNet-50 models, achieving an accuracy of 91 % and demonstrating notable classification stability. In contrast, ResNet-18 suffered severe performance degradation, with accuracy dropping from 70 % to 26 %, while ResNet-50 required extensive tuning to improve its outcomes. The non-pretrained CNN excelled in handling class imbalances and capturing complex diagnostic patterns, emphasizing the potential of tailored architectures for medical imaging tasks. These results underscore the importance of designing domain-specific architectures, demonstrating that model complexity does not necessarily guarantee better performance. Particularly in scenarios with limited datasets, well-designed custom models can surpass pre-trained architectures in diagnostic imaging applications.

Resumen La detección temprana de la retinopatía diabética representa un desafío crítico en el diagnóstico médico, donde el aprendizaje profundo dentro del campo de la inteligencia artificial emerge como una herramienta prometedora para optimizar la identificación de patrones patológicos en imágenes retinales. Este estudio evaluó comparativamente tres arquitecturas de redes neuronales convolucionales ResNet-18, ResNet-50 y una CNN personalizada o no-preentrenada para clasificar imágenes de retinopatía diabética en un conjunto de datos de imágenes agrupadas en cinco categorías, revelando diferencias significativas en su capacidad para aprender y generalizar. Los resultados demostraron que la arquitectura de red neuronal convolucional no-preentrenada superó consistentemente a los modelos preentrenados basados en ResNet-18 y ResNet-50, alcanzando una precisión del 91 % y una notable estabilidad en la clasificación. Mientras ResNet-18 mostró limitaciones severas, degradándose de un 70 % a un 26 % de precisión, y ResNet-50 requirió ajustes para mejorar su rendimiento, la CNN no preentrenada exhibió una capacidad sobresaliente para manejar el desbalance de clases y capturar patrones diagnósticos complejos. El estudio subraya la importancia de diseñar arquitecturas específicamente adaptadas a problemas médicos, destacando que la complejidad no garantiza necesariamente un mejor desempeño, y que un diseño cuidadoso puede superar modelos preentrenados en tareas de diagnóstico por imagen cuando la cantidad de datos con que se cuenta es limitada.

Abstract — The exponential increase in Internet traffic is mainly due to the proliferation of services such as audio and video streaming, the emergence of applications that require a lot of bandwidth to work optimally and generally the process of digitalization of services. In this context, bandwidth management plays a fundamental role, which translates into a better experience for users. Traffic congestion causes the exchange of information to become deficient, that is why techniques such as automatic bandwidth adjustment have been investigated, which manages the bandwidth according to the traffic demand, therefore in this document a study is made about the automatic bandwidth adjustment, the way in which Artificial Intelligence is integrated with computer networks, finally a comparison will be made of several machine learning methods, cataloged within supervised learning, carrying out several experiments determining that Random Forest is the most effective method to predict the automatic bandwidth adjustment, followed by Naive Bayes, Logistic Regression, and Support Vectorial Machine (SMV), on the other hand K -nearest neighbor (KNN) and neural network do not demonstrate considerable effectiveness, each experiment was carried out taking into account the Quality of Service (QoS).

Resumen El aumento exponencial del tráfico en Internet se debe principalmente a la proliferación de servicios como el streaming de audio y video, a la aparición de aplicaciones que requieren un alto ancho de banda para funcionar de manera óptima, y, en general, al proceso de digitalización de servicios. En este contexto, la gestión del ancho de banda desempeña un papel fundamental, ya que contribuye a una mejor experiencia para los usuarios. La congestión de tráfico genera que el intercambio de información sea ineficiente. Por ello, se han investigado técnicas como el ajuste automático de ancho de banda, que permite gestionar el ancho de banda según la demanda de tráfico. En este documento se realiza un estudio sobre el ajuste automático de ancho de banda, abordando cómo la Inteligencia Artificial se integra con las redes de computadoras. Además, se presenta una comparación de varios métodos de aprendizaje supervisado en machine learning. A través de diversos experimentos, se determinó que Random Forest es el método más efectivo para predecir el ajuste automático de ancho de banda, seguido por Naive Bayes, Regresión Logística y Support Vector Machine (SVM). En contraste, los métodos K-Vecinos Más Cercanos (KNN) y redes neuronales no demostraron una efectividad significativa. Cada experimento se realizó considerando la Calidad de Servicio (QoS).

Resumen La floricultura es un sector de creciente importancia a nivel mundial, contribuyendo a la generación de empleo, ingresos y a la promoción de la biodiversidad y la sostenibilidad. Este estudio tuvo como objetivo identificar los factores que influyen en la adopción de la floricultura como cultivo alternativo en la provincia de Leoncio Prado, Perú, y evaluar su viabilidad económica. Para ello, se encuestó a 269 agricultores, analizando actitudes, aptitud de la tierra, factores socioeconómicos y ambientales. Mediante análisis descriptivo, pruebas de chi-cuadrado y regresión logística (p < 0,1) se seleccionaron los factores influyentes. Además, se emplearon múltiples algoritmos de aprendizaje automático (Árboles de Decisión, Regresión Logística, KNN, SVM, Ensemble, Redes Neuronales, Naive Bayes) con validación cruzada (k = 5) y métricas AUC para modelar la intención de adopción. Se desarrollaron escenarios incrementando la predisposición a adoptar la floricultura, y se realizó un análisis económico de ocho especies tropicales (Ginger Rojo, Anturio, Bastón del Emperador, Heliconia, Gardenia, Pico de Loro, Heliconias Golden, Maracas). Los resultados revelan que la disposición a cambiar de cultivo, la participación en campañas de sensibilización, la destinación de áreas a la conservación y el control de costos son factores clave. El modelo de redes neuronales alcanzó un AUC de 83,3% y escenarios mejorados indican que la adopción podría incrementar hasta en 11,32%. El Ginger Rojo demostró alta rentabilidad (VAN S/10428; TIR 51%; PRI 0,7 años). En conclusión, la floricultura representa una alternativa económica y ambientalmente viable que contribuye a la diversificación agrícola y a la sostenibilidad.

Abstract Floriculture is a sector of growing global importance, contributing to employment generation, income creation, and the promotion of biodiversity and sustainability. This study aimed to identify the factors influencing the adoption of floriculture as an alternative crop in the province of Leoncio Prado, Peru, and to assess its economic viability. A total of 269 farmers were surveyed, analyzing attitudes, land suitability, and socioeconomic and environmental factors. Influential factors were identified using descriptive analysis, chi-square tests, and logistic regression (p < 0.1). Additionally, multiple machine learning algorithms (Decision Trees, Logistic Regression, KNN, SVM, Ensemble, Neural Networks, Naive Bayes) with cross-validation (k = 5) and AUC metrics were employed to model adoption intentions. Scenarios were developed to increase the willingness to adopt floriculture, and an economic analysis of eight tropical species (Red Ginger, Anthurium, Emperor's Staff, Heliconia, Gardenia, Parrot's Beak, Golden Heliconias, Maracas) was conducted. The results reveal that willingness to change crops, participation in awareness campaigns, allocation of areas for conservation, and cost control are key factors. The neural network model achieved an AUC of 83.3%, and improved scenarios indicate that adoption could increase by up to 11.32%. Red Ginger demonstrated high profitability (NPV S/10428; IRR 51%; PBP 0.7 years). In conclusion, floriculture represents an economically and environmentally viable alternative that contributes to agricultural diversification and sustainability.

ABSTRACT This experimental study investigates the axial compression effect of rectangular pultruded glass fiber-reinforced polymer (P-GFRP) tubular column sections, examining the impact of width-to-thickness ratio (B/t), aspect ratio (H/B), and column height on their structural performance. A total of 27 GFRP columns were subjected to axial compression tests to evaluate their ultimate load and initial stiffness. The columns exhibited a uniform failure pattern, characterized by crushing, mid-section fractures, and longitudinal splitting at the corners. The results revealed a negative correlation between the ultimate load and the aspect ratio, as well as the width-to-thickness ratio. This study utilized advanced machine learning algorithms, namely Response Surface Methodology (RSM) and Artificial Neural Network (ANN), to develop predictive models for the ultimate load of GFRP columns. The RSM model achieved an R2 value of 0.8347, demonstrating good accuracy in predicting ultimate load. The ANN model outperformed the RSM model, with an R criterion exceeding 0.68807 across training, testing, and validation phases, showing a stronger correlation between experimental and predicted outcomes. This research establishes a framework for forecasting the mechanical properties of column sections.

ABSTRACT This study presents the development of a methodology for designing neuro-adaptive robust controllers based on a reference model associated with an artificial neural network of radial basis functions (ANN-RBF) for solid fuel suborbital rockets. The modelling and neuro-adaptive robust control algorithms for these rockets are presented. Initially, the methodology is evaluated for a robust controller based on a reference model with ANN-RBF for altitude control. The main objective of the control is to suppress the effect of non-linear uncertainties inherent in the process. The method involves mathematical and computational modelling, together with the design of adaptive controllers for stability and performance analysis. The controllers considered include model reference adaptive control (MRAC) techniques and a model reference neuro-adaptive control (MRNAC) approach. The analysis, carried out using computer simulations, evaluates the behavior of each controller in relation to system stability and performance. The final objective is to select the most suitable controller for the suborbital rocket, taking into account the system constraints, robust performance requirements, robust stability, and optimal adaptability. This research promotes the development of adaptive controllers for suborbital rockets, with possible applications in scientific research and commercial launches.

ABSTRACT This study explores the fastening behavior of punctured double-double (DD) laminates, a superior alternative to traditional quadaxial laminates (QUAD) due to improved structural efficiency and lower maintenance. However, the effect of various cutout shapes and sizes on DD laminates’ fastening performance is still unexplored. This research examines optimal ply orientations, rotation angles, and fastening loads for DD laminates with circular, elliptical, and combined-shape cutouts to assess their impact on stability. A hybrid optimization method using an artificial neural network (ANN) and genetic algorithm (GA) is developed to predict maximum buckling loads, avoiding time-intensive finite element analysis (FEA). The ANN models, with R2 values of 0.994 to 0.999, show excellent performance. The best model for circular cutouts achieved R2 is 0.999 with a mean absolute error of 0.0059. Results indicate that elliptical and combined-shape cutouts significantly influence ply angles and buckling loads. Combined-shape cutouts offer superior stability as size increases, with buckling load improvements of 15% over circular cutouts. This study highlights the potential of ANN-GA techniques for optimizing DD laminate designs and improving structural performance.

ABSTRACT This study employs machine learning (ML) to analyze the melting and reconsolidation behaviors of iron, emphasizing the influence of cold reduction ratios and rolling sequences. Five samples with varied cold reduction ratios and rolling patterns were examined. Findings indicate that when the cold reduction ratio exceeds 65%, coordinated cold melting minimally impacts crystallographic consistency. Texture formation remains largely unaffected during cold melting and short-duration annealing. However, extended annealing prompts irregular grain growth, altering crystal orientation. Sheets rolled in alignment with their initial condition exhibit consistency patterns similar to conventionally cold-melted pure iron after prolonged annealing. Key parameters influencing material performance were evaluated, revealing annealing temperature as the most significant factor (5.94), followed by cold melting direction order (1.46), while the hanging period during annealing had minimal impact (1.02). ML models were employed to predict Goss angle expansion using cold-rolling and annealing parameters. This approach demonstrates the potential of ML to predict texture evolution in pure iron, offering valuable insights for optimizing industrial cold-rolling practices.

This study was conducted using the first-lactation records of 1092 Vrindavani crossbred cattle to compare the relative efficiency of an artificial neural network (ANN) versus multiple linear regression for predicting the first-lactation 305-day milk yield (FL305DMY). The two input sets used for predicting FL305DMY in the study were input set-1 : first four monthly test-day milk yields, age at first calving, and peak milk yield; and input set-2: first four monthly milk yields, age at first calving, and peak milk yield. The ANN was trained using a backpropagation algorithm based on Bayesian regularisation, and the algorithm was tested using four sets of training and test data at ratios of 66.67:33.33, 75:25, 80:20, and 90:10. The results revealed that the coefficient of determination showed no regular trend with decreasing the test dataset. Nevertheless, the observed values were highest for the 90:10 ratio of training-test data for both input sets, with the lowest root mean square error. The ANN model outperformed the multiple linear regression model when predicting FL305DMY, with an accuracy of 79.09% for input set-1 and 83.67% for input set-2, with the lowest root mean square error values for both input sets. Therefore, the ANN model can be used as an alternative technique to predict FL305DMY in Vrindavani cows.

ABSTRACT: The tomato (Solanum lycopersicum L.) plays a vital role in global agriculture and is a model organism in genetic studies. Visual classification of tomatoes for genetic improvement programs faces challenges due to variety diversity, uneven ripening, external damages, and evaluator subjectivity. Recent advances in the field of computational resources, such as image phenotyping have enabled pre- and post-harvest assessments that are both fast and precise. This study aimed to classify tomato fruits based on shape, group, color, and defects using Convolutional Neural Networks (CNNs). The performance of five architectures - VGG16, InceptionV3, ResNet50, EfficientNetB3, and InceptionResNetV2 was evaluated to identify and determine the most efficient one for this classification. The research considered ten hybrids and their five parental lines. The experiment was conducted in the field, and images of ripe fruits were acquired using a portable mini studio. The ExpImage package in R software was used for fruit individualization by image and to aid in creating a synthetic database for network training. Images were grouped according to their classifications in terms of shape, color, groups, and defects. The InceptionResNetV2 architecture was the most efficient, achieving metrics such as precision and recall exceeding 93 % for most analyzed variables, and shorter classification times. This study advances the understanding of CNN applications in agriculture and research and provides valuable guidelines for optimizing classification tasks in distinct types of fruits. ABSTRACT Solanum L. L studies diversity ripening damages subjectivity resources pre postharvest post harvest precise shape group color CNNs. CNNs . (CNNs) VGG16 VGG InceptionV3 InceptionV ResNet50 ResNet EfficientNetB3 EfficientNetB InceptionResNetV lines studio training groups 9 variables times (CNNs VGG1 ResNet5

Abstract Corrosion is a recurrent problem in the natural gas industry, due to the presence of corrosive gases such as CO2 and H2S that in the presence of water can attack steel and produce damages and leaks that generate accidents and contamination, that is why the prediction of the tendency of natural gas towards corrosion is essential to develop strategies to inhibit and mitigate this process. The objective of the study was to develop an Artificial Neural Network (ANN) trained with composition data of natural gas samples to predict the corrosive tendency, as an alternative method that can be used as an efficient predictive tool. The ANN was trained with a sample of 46 natural gases and 11 components for each one, in addition to pressure and temperature as operational conditions, representing a total of 598 training data of the ANN, also 8 additional samples were used for external validation of the model. The ANN that developed was within the principles of Bayesian statistics, with a Multilayer Perceptron architecture. It was obtained that the ANN is able to correctly classify 95.65% of the corrosive tendency of the samples, being those with non-corrosive tendency the ones that presented the lowest prediction percentage (75%). The prediction of the external samples followed the expected trend, with lower capacity to classify the non-corrosive ones, and the samples with possible corrosion and corrosive were predicted in 80%. It is concluded that ANN is an efficient tool for predicting the corrosive tendency of natural gas in the presence of CO2 and its effectiveness depends on using samples with all the necessary components to train it, and it can be improved by introducing other parameters such as the amount of water present.

Resumen La corrosión es un problema recurrente en la industria del gas natural, debido a la presencia de gases corrosivos como el CO2 y el H2S que en presencia de agua pueden atacar el acero y producir daños y fugas que generen accidentes y contaminación, es por ello que la predicción de la tendencia del gas natural hacia la corrosión es fundamental para desarrollar estrategias de inhibición y mitigación de este proceso. El objetivo del estudio fue desarrollar una Red Neuronal Artificial (RNA) entrenada con datos de composición de muestras de gas natural para predecir la tendencia corrosiva, como método alternativo que pueda ser utilizado como una herramienta predictiva eficiente. La RNA se entrenó con una muestra de 46 gases naturales y 11 componentes por cada una, además de la presión y la temperatura como condiciones operacionales, lo que representa un total de 598 datos de entrenamiento de la RNA, además se utilizaron 8 muestras adicionales para la validación externa del modelo. La RNA que desarrolló fue dentro de los principios de la estadística Bayesiana, con una arquitectura de Perceptrón Multicapa. Se obtuvo que la RNA es capaz de clasificar correctamente el 95,65% de la tendencia corrosiva de las muestras, siendo aquellas con tendencia no corrosiva las que presentaron el menor porcentaje de predicción (75%). La predicción de las muestras externas siguió la tendencia esperada, con menor capacidad para clasificar a las no corrosivas, y las muestras con posible corrosión y corrosivas fueron predichas en 80%. Se concluye que la RNA es una herramienta eficiente para la predicción de la tendencia corrosiva del gas natural en presencia de CO2 y su efectividad depende de que se utilicen muestras con todos los componentes necesarios para entrenar la misma, además puede ser mejorada al introducir otros parámetros como la cantidad de agua presente.

ABSTRACT The Theobroma cacao L. bean fermentation is an important post-harvest process for the development of its properties and aroma. Although cocoa fermentation is complex, farmers use empirical methods to determine its degree of fermentation. One of the traditional techniques used to recognize the quality of fermentation is the “Cut Test”, performed by a person manually. However, this type of techniques could have a computer-based alternative. Therefore, in this study, the use of convolutional neural networks (CNN) based on deep learning was analyzed to determine the degree of fermentation of cocoa beans. For this purpose, a model was developed whose performance was verified in terms of precision and confusion matrix. This model achieved a positive accuracy of 82 % and a confusion matrix with favorable numbers on the diagonal elements. These results show that CNN is a viable option for the classification of cocoa beans based on their fermentation.

RESUMEN La fermentación del grano de Theobroma cacao L. es un proceso postcosecha importante para el desarrollo de sus propiedades y del aroma. Aunque se trata de un proceso complejo, los agricultores utilizan métodos empíricos/tradicionales para determinar su grado de fermentación, como es la “Prueba de corte” realizada manualmente. No obstante, este tipo de técnicas podría tener una alternativa basada en computadores. Por ello, en este estudio se analizó el uso de redes neuronales convolucionales (CNN) basadas en el aprendizaje profundo, para determinar el grado de fermentación de los granos de cacao. Para el efecto, se desarrolló un modelo cuyo rendimiento fue comprobado en términos de precisión y matriz de confusión. Este modelo alcanzó una exactitud positiva del 82 % y una matriz de confusión con números favorables en los elementos diagonales. Estos resultados dan muestra de que CNN es una opción viable para la clasificación de granos de cacao, según su fermentación.

Abstract This study proposes an IoT architecture for monitoring emissions of polluting gases in vehicles, in response to the growing concern about air pollution and global warming. The architecture is based on a node equipped with DHT22, MQ9, and MQ135 sensors to capture temperature, humidity, and gas emissions, respectively. This node effectively communicates through the LTE network to send the data to the ThingSpeak platform. An analysis of CO2, CO, and CH4 pollution levels is conducted using the collected data. These data are validated through the technical review of a test vehicle. Subsequently, an Artificial Neural Network (ANN) is trained using a specific database of CO2 emissions from vehicles in Canada. As a result, a high R2 of 99.2 % is achieved, along with low values of RMSE and MSE, indicating that the model is making accurate predictions and fits well to the training data. The ANN aims to predict CO2 emissions and verify CO2 data from the IoT network. The architecture demonstrates its capability for real-time monitoring and its potential to contribute to pollution reduction.

Resumen Este estudio propone una arquitectura IoT para el monitoreo de emisiones de gases contaminantes en vehículos, en respuesta a la creciente preocupación por la contaminación del aire y el calentamiento global. La arquitectura se basa en un nodo equipado con sensores DHT22, MQ9 y MQ135 para capturar la temperatura, humedad y emisiones de gases, mismo que se comunica de manera efectiva a través de la red LTE para enviar los datos a la plataforma ThingSpeak. Se lleva a cabo un análisis de los niveles de contaminación de CO2, CO y CH4 mediante los datos recopilados. Estos datos se validan mediante la revisión técnica de un vehículo de prueba. Posterior, se entrena una red neuronal artificial (ANN) utilizando una base de datos específica de emisiones de CO2 de vehículos en Canadá, como resultado se obtiene un R2 alto de 99,2 % y los valores de RMSE y MSE bajos, esto indican que el modelo está haciendo predicciones precisas y se ajusta bien a los datos de entrenamiento. La ANN tiene como objetivo predecir las emisiones de CO2 y verificar los datos de CO2 provenientes de la red IoT. La arquitectura demuestra su capacidad para el monitoreo en tiempo real y su potencial para contribuir a la reducción de la contaminación.

ABSTRACT Cement is the fundamental binder of concrete, and its manufacture has a significant impact on the environment; therefore, it is necessary to look for eco-sustainable alternatives, including additions such as natural pozzolana, which affect the internal matrix of concrete and therefore the compressive strength and capillary absorption of concrete. In this context, prediction models for capillary absorption and compressive strength of concrete with pozzolana additions have been determined by applying linear multiple regression tools and artificial neural networks which will help reduce laboratory testing costs and times. For this purpose, 16 types of mixtures were designed with w/c ratios of 0.40, 0.45, 0.50 and 0. 55 and addition of 10, 15 and 20% of pozzolana; 160 cylindrical samples were manufactured and tested in laboratory, the values of capillary absorption and compressive strength at 28 and 56 days of curing were determined; the effect of each variable on the results obtained indicated that 15% pozzolana significantly improved the properties studied; using the data of the manufacturing variables of each design and the results of capillary absorption and compressive strength, prediction models were obtained for both properties; the best back propagation neural networks (BPNN) structure is [10,20,10,1], with R2 compression=0. 9486 and R2 capillary absorption=0.9756; while the models obtained with multiple linear regression obtained R2 compression = 0.9391 and R2 capillary absorption = 0.8693; both techniques showed a high reliability for the prediction of compressive strength and capillary absorption.

Abstract The ultrasonic wave propagation velocity (UPV) technique enables the evaluation of mechanical properties of concrete, including fiber reinforced concrete (FRC). Therefore, the objective of this study is to verify potential correlations between the mechanical properties of FRC and UPV, based on experimental data selected after a systematic literature review. Additionally, relationships between UPV and compressive strength were assessed using models proposed by researchers. An artificial neural network (ANN) technique was applied to analyze which properties, when associated with UPV, assist in estimating the compressive strength of FRC. It was observed that the literature-proposed models for estimating compressive strength using UPV proved to be ineffective. The ANN model resulted in a determination coefficient of 0.973, an accuracy of 3.18 MPa, utilizing UPV, cement consumption, and the water-to-cement ratio as input data, with three neurons in the hidden layer and one output data (compressive strength). (UPV . (FRC) Therefore review Additionally researchers (ANN literatureproposed ineffective 0973 0 973 0.973 318 3 18 3.1 MPa consumption watertocement water strength. strength) (FRC 097 97 0.97 31 1 3. 09 9 0.9 0.

Resumo A partir da técnica de velocidade de propagação de onda ultrassônica (VPU) é possível avaliar propriedades mecânicas de concretos, incluindo o concreto reforçado com fibras (CRF). Assim, o objetivo deste trabalho é verificar possíveis correlações entre as propriedades mecânicas do CRF e VPU, com base em dados experimentais de publicações selecionados após uma revisão sistemática da literatura. Também foram avaliadas relações entre a VPU e a resistência à compressão a partir de modelos propostos por pesquisadores. Aplicou-se a técnica de rede neural artificial (RNA) para analisar quais propriedades, quando associadas com a VPU, auxiliam na estimativa da resistência à compressão do CRF. Observou-se que os modelos propostos pela literatura para estimar a resistência à compressão por meio da VPU não foram eficazes. O modelo de RNA resultou em um coeficiente de determinação de 0,973, uma acurácia de 3,18 MPa e utilizou como dados de entrada a VPU, o consumo de cimento e a relação a/c, além de três neurônios na camada escondida e um dado de saída (resistência à compressão). (VPU concretos . (CRF) Assim pesquisadores Aplicouse Aplicou se (RNA Observouse Observou eficazes 0973 0 973 0,973 318 3 18 3,1 ac c a/c compressão. compressão) (CRF 097 97 0,97 31 1 3, 09 9 0,9 0,

BACKGROUND: Machine learning (ML) refers to computational algorithms designed to learn from patterns in data to provide insights or predictions related to that data OBJECTIVE: Multiple studies report the development of predictive models for triage or identification of critically ill children. In this study, we validate machine learning models developed in South Africa for the identification of critically ill children presenting to a tertiary hospital RESULTS: The validation sample comprised 267 patients. The event rate for the study outcome was 0.12. All models demonstrated good discrimination but weak calibration. Artificial neural network 1 (ANN1) had the highest area under the receiver operating characteristic curve (AUROC) with a value of 0.84. ANN2 had the highest area under the precision-recall curve (AUPRC) with a value of 0.65. Decision curve analysis demonstrated that all models were superior to standard strategies of treating all patients or treating no patients at a proposed threshold probability of 10%. Confidence intervals for model performance overlapped considerably. Post hoc model explanations demonstrated that models were logically coherent with clinical knowledge CONCLUSIONS: Internal validation of the predictive models correlated with model performance in the development study. The models were able to discriminate between critically ill children and non-critically ill children; however, the superiority of one model over the others could not be demonstrated in this study. Therefore, models such as these still require further refinement and external validation before implementation in clinical practice. Indeed, successful implementation of machine learning in practice within the South African setting will require the development of regulatory and infrastructural frameworks in conjunction with the adoption of alternative approaches to electronic data capture, such as the use of mobile devices