SUMMARY: In the present study, it is aimed to reveal the performances of different classification algorithms in sex determination from first cervical vertebra, that is, atlas measurements. The classification success of 4 different machine learning algorithms was comparatively examined for the purpose of sex determination by evaluating 22 atlas measurements on cone beam computed tomography (CBCT) images of 145 female and 145 male adults. Logistic regression (LR), classification and regression tree (CART), support vector machine (SVM) and neural network (NN) algorithms were used for sex diagnosis. Area under the ROC curve (AUC), classification accuracy (CA), F1-ratio, Precision and Recall indexes were used for model performances. Except for 2 measurements, there was a significant difference between men and women in terms of 20 other parameters (p<0.05). The adjusted effects of these parameters on sex determination were examined with multivariate models and algorithms, and the success of all 4 algorithms was quite good. The success of the NN algorithm (Accuracy 91.3 %; 0.87 Specificity, 0.85 Sensitivity) in correctly classifying male and female was the highest, followed by the LR algorithm (Accuracy 90.9 %; 0.86 Specificity, 0.83 Sensitivity). It was found that the machine learning algorithm applied to the variables of the atlas gave high accuracy regarding sex and the NN model was highly effective in sex determination. In addition, a large morphometric database of atlas was presented in our results.

RESUMEN: El objetivo de estudio fue revelar el desempeño de diferentes algoritmos de clasificación en la determinación del sexo a partir de la primera vértebra cervical, el atlas. El éxito de la clasificación de 4 algoritmos diferentes de aprendizaje automático se examinó comparativamente con el fin de determinar el sexo mediante la evaluación de 22 mediciones de atlas en imágenes de tomografía computarizada de haz cónico (CBCT) de 145 mujeres y 145 hombres adultos. Para el diagnóstico de sexo se utilizaron algoritmos de clasificación y árbol de regresión (CART), máquina de vectores de soporte (SVM) y redes neuronales (NN). Para el rendimiento del modelo se utilizaron el área bajo la curva ROC (AUC), la precisión de clasificación (CA), la relación F1, los índices de precisión y recuperación. Excepto en dos mediciones, hubo una diferencia significativa entre hombres y mujeres en otros 20 parámetros (p<0,05). Los efectos ajustados de estos parámetros en la determinación del sexo se examinaron con modelos y algoritmos multivariados, y el éxito de los cuatro algoritmos fue bastante bueno. El éxito del algoritmo NN (Precisión 91,3 %; 0,87 Especificidad, 0,85 Sensibilidad) en clasificar correctamente a hombres y mujeres fue el más alto, seguido por el algoritmo LR (Precisión 90,9 %; 0,86 Especificidad, 0,83 Sensibilidad). Se encontró que el algoritmo de aprendizaje automático aplicado a las variables del atlas dio una alta precisión con respecto al sexo, y el modelo NN fue altamente efectivo en la determinación del sexo. Además, en nuestros resultados se presentó una importante base de datos morfométricos del atlas.

RESUMEN: La preparación militar depende en gran medida de la aptitud física y la capacidad operativa de su personal. Este estudio tiene como objetivo mejorar la eficacia de los protocolos de evaluación y clasificación de la composición corporal dentro del contexto militar. Se realizó una evaluación integral de 4.370 militares chilenos en servicio activo, centrándose en las características antropométricas, incluido el tejido adiposo, el tejido muscular y los índices antropométricos. El estudio observó diferencias significativas en la composición corporal entre sexos, con los hombres exhibiendo niveles más bajos de porcentaje de grasa corporal (hombres: 26,28 %; mujeres: 34,62 %) pero niveles más altos de masa muscular (hombres: 42,0 %; mujeres: 36,0 %;), índice de músculo esquelético (hombres: 11,81; mujeres: 9,31) e índice de masa magra (hombres: 19,92; mujeres: 18,45) en comparación con las mujeres. Se observaron altos niveles de tejido muscular en ambos grupos. Al integrar estos hallazgos en un protocolo de evaluación estandarizado, se logró una clasificación más precisa del personal militar, superando los métodos tradicionales utilizados en poblaciones obesas sedentarias. El estudio aboga por la adopción futura de un modelo de evaluación basado en algoritmos de inteligencia artificial (IA), que consideren la naturaleza multifacética de la composición corporal y su impacto en la capacidad operativa. Un modelo de este tipo permitiría a las fuerzas militares optimizar la aptitud física y la preparación de su personal, mejorando así su eficacia en las operaciones de despliegue.

SUMMARY: Military readiness relies heavily on the physical fitness and operational capability of its personnel. This study aims to enhance the effectiveness of body composition assessment and classification protocols within the military context. A comprehensive evaluation of 4370 active-duty Chilean military personnel was conducted, focusing on anthropometric characteristics, including adipose tissue, muscle tissue, and anthropometric indices. The study observed significant differences in body composition between genders, with men exhibiting lower levels of body fat percentage (men: 26.28 %; women: 34.62 %) but higher levels of muscle mass (men: 42.0 %; women: 36.0 %;), skeletal muscle index (men:11.81; women: 9.31), and fat-free mass index (men: 19.92; women: 18.45) compared to women. High levels of muscle tissue were observed in both groups. By integrating these findings into a standardized assessment protocol, a more accurate classification of military personnel was achieved, surpassing traditional methods used in sedentary obese populations. The study advocates the future adoption of an assessment model based on artificial intelligence (AI) algorithms, which consider the multifaceted nature of body composition and its impact on operational capability. Such a model would enable military forces to optimize their personnel's physical fitness and readiness, thus enhancing their effectiveness in deployment operations.

RESUMEN: Dado que los algoritmos de aprendizaje automático dan resultados más fiables, en los últimos años han sido utilizados en el campo de la salud. Las variables orbitales dan resultados muy exitosos a la hora de clasificar correctamente el sexo. Esta investigación se ha centrado en la determinación del sexo utilizando determinadas variables obtenidas a partir de las imágenes orbitales de la tomografía computarizada (TC) mediante el uso de algoritmos de aprendizaje automático (AA). En este estudio se probaron 12 variables determinadas en 600 imágenes orbitales de 300 individuos (150 hombres y 150 mujeres) con diferentes AA. Se utilizaron algoritmos de AA de árbol de decisión (DT), K-Nearest Neighbour, regresión logística (RL), Random Forest (RF), análisis discriminante lineal (ADL) y Naive Bayes (NB) para el aprendizaje no supervisado. Los análisis estadísticos de las variables se realizaron con el programa Minitab® 21.2 (64 bits). La tasa de ACC de los algoritmos NB, DT, KNN y RL se encontró en % 83, mientras que la tasa de ACC de los algoritmos ADL y RFC se determinó en % 85. Según el análisis de Sharp, la variable con el mayor grado de efecto se encontró como BOW. El estudio determinó el sexo con alta precisión en las proporciones de 0,83 y 0,85 mediante el uso de las variables de las imágenes de TC orbitales, y se adquirieron los datos morfométricos relacionados de la población en cuestión, enfatizando la variación racial.

SUMMARY: Since machine learning algorithms give more reliable results, they have been used in the field of health in recent years. The orbital variables give very successful results in classifying sex correctly. This research has focused on sex determination using certain variables obtained from the orbital images of the computerized tomography (CT) by using machine learning algorithms (ML). In this study 12 variables determined on 600 orbital images of 300 individuals (150 men and 150 women) were tested with different ML. Decision tree (DT), K-Nearest Neighbour (KNN), Logistic Regression (LR), Random Forest (RF), Linear Discriminant Analysis (LDA), and Naive Bayes (NB) algorithms of ML were used for unsupervised learning. Statistical analyses of the variables were conducted with Minitab® 21.2 (64-bit) program. ACC rate of NB, DT, KNN, and LR algorithms was found as % 83 while the ACC rate of LDA and RFC algorithms was determined as % 85. According to Shap analysis, the variable with the highest degree of effect was found as BOW. The study has determined the sex with high accuracy at the ratios of 0.83 and 0.85 through using the variables of the orbital CT images, and the related morphometric data of the population under question was acquired, emphasizing the racial variation.

RESUMEN: El registro preciso de la biodiversidad tropical ha llevado a la necesidad de inventarios forestales extensos y precisos en zonas montañosas, donde las imágenes satelitales y la modelización de transferencia radiativa de los bosques han adquirido mucha relevancia. Los algoritmos de clasificación de imágenes de última generación (i.e. aprendizaje automático), aplicados con información auxiliar del relieve, se han revelado relativamente exitosos para el mapeo de tipos de bosques, sin embargo, la exactitud de la clasificación se mide solamente en promedio sobre áreas extensas, y poca atención se ha vertido en predecir las configuraciones de iluminación que en realidad causan alta incertidumbre en los resultados de la clasificación. En este artículo, se presenta una metodología basada en el simulador de imágenes Discrete Anisotropic Radiative Transfer (DART), para determinar si los tipos de bosque son distinguibles o no en configuraciones de terreno escarpado. Este trabajo describe la comparación de escenas forestales simuladas y reales en imágenes de alta resolución espacial (4 m). El método se aplicó para la estimación de la separabilidad espectral de tres tipos de bosques (encino, pino y selva alta perennifolia) en terrenos escarpados en México. Para ángulos de incidencia solar extremos (muy pequeños o muy grandes), el bosque de pino y la selva alta resultaron indistinguibles, y en contraste, fueron distinguibles en laderas cercanas al plano perpendicular solar. En consecuencia, para incluir un máximo de configuraciones favorables de laderas, recomendamos incorporar imágenes adquiridas en la mañana y en la tarde en el entrenamiento de los algoritmos de aprendizaje automático.

ABSTRACT: The precise account of biodiversity in the tropics has prompted the need for extended and accurate forest inventories in mountainous landscapes, where satellite imagery and radiative transfer modeling of forests have acquired much relevance. Last-generation classification algorithms (i.e. machine learning) using terrain information have proven useful for mapping forest types, however, the accuracy of classifications is measured on average over a large extent, and little emphasis has been placed on predicting which illumination configurations could in fact cause high uncertainty in the classification results. This paper presents a 3D modeling approach adapted to the Discrete Anisotropic Radiative Transfer (DART) image simulator, which indicates whether forest types are distinguishable in a set of steep terrain configurations. The approach also describes a comparison of simulated and real forest scenes on slopes at high (4 m) spatial resolution. This method was applied to estimate the spectral separability of three forest types (oak, pine, and high tropical forests) on steep terrain in Mexico. For extreme (low or high) solar incidence angles, the pine and high tropical forests were indistinguishable, and by contrast, they were distinguishable on slopes near to the solar perpendicular plane. As a consequence, to maximize favourable slope configurations, we recommend to incorporate images acquired in the morning and in the afternoon for machine learning classification algorithms.

The past decade has seen a renewed possibility of using machine learning algorithms to solve a large collection of problems in several fields. Data acquisition for mining operations has increased with the growth in sensor-based technologies, and therefore the amount of information available for mining applications has dramatically increased. Ore sorting equipment is available for separating ore from waste based on differences in physical properties detected by a real-time analyser. The separation efficiency depends on the contrast in these properties. In this study we investigate the application of machine learning models trained using data from the output of a dual-energy X-ray ore sorting apparatus at a gold mine. The particles were first hand-sorted into ore and gangue classes based on their mineralogical composition. Classification models were then used to help decide the balance between the number of true and false positives for ore in the concentrate, with a view to economic parameters, using their receiver operator characteristic (ROC) curves. The results showed AUC (area under the ROC curve) scores of up to 0.85 for the classification models and a maximum reward condition Fp r/Tp r around 0.5/0.9 for a simplified economic model.

Abstract: In many supervised learning problems, feature selection techniques are increasingly essential across various applications. Feature selection significantly influences the classification accuracy rate and the quality of SVM model by reducing the number of features, remove irrelevant and redundant features. In this paper, we evaluate the performance of twenty feature selection algorithms over four databases. The performance is conducted in term of: classification accuracy rate, Kuncheva’s Stability, Information Stability, SS Stability and SH Stability. To measure the feature selection algorithms, multiple datasets from the UCI Machine Learning Repository are utilized to assess both classification accuracy and stability variations.

Abstract: Computer vision, for decades, has been involved in solving problems in everyday life, under the implementation of different computational methods, that have evolved over time. Feature extraction, along with other computer techniques, is considered a way to develop computer vision systems; currently, plays an important role, considered a complex task, allowing to obtain essential descriptors of the segmented images, differentiating particular characteristics between different classes, even when they share similarity with each other, guaranteeing the delivery of information not redundant to classification algorithms. Likewise, in this work, a computer vision system has been developed for the recognition of foliar damage caused by diseases and pests in tomato plants. The methodology implemented is based on four modules: preprocessing, segmentation, feature extraction, and classification; in the first module, the image is preprocessed of a color space RGB to L*a*b*; in the second module, the area interest was segmented, under the implementation of the algorithm principal component analysis PCA; in the third module, features are extracted from the area of interest, obtaining texture descriptors with the Haralick algorithm, and chromatic features through Contrast descriptors, Hu moments, Gabor characteristics, Fourier descriptors, and discrete cosine transform DCT; in the fourth module, the performance of the classification algorithms were tested, with the characteristics obtained from the previous stage, considering: SVM, Backpropagation, Logistic Regression, KNN, and Random Forests.

Abstract: Currently, the diversity of sources generating data in a massive online manner cause data streams to become part of many real work applications. Learning from a data stream is a very challenging task due to the non-stationary nature of this type of data. Characteristics such as infinite length, concept drift, concept evolution and recurrent concepts are the most common problems that need to be addressed by data stream learning algorithms. In this work an algorithm for data stream classification based on an associative classifier is presented. This proposal combines a clustering algorithm and the Naïve Associative Classifier for Online Data (NACOD) to address this problem. A set of micro-clusters (MCs), a data structure that summarizes the information of the current data, is used instead of storing the whole data. The MCs are continually updated with the arriving data, either to create new MCs or to update existing ones. The added MCs helps to deal with concept drift. To assess the performance of the proposed model, experiments were carried out on 3 data sets commonly used to evaluate data stream classification algorithms: KDD Cup 1999, Forest Cover Type and Statlog (Shuttle). Our model achieved higher accuracies than those achieved with algorithms such as data stream version of Naïve Bayes and Hoeffding Tree, the average accuracies achieved were for KDD Cup 1999: 100 %, Statlog (Shuttle): 99.01 % and Forest Cover Type 70.44 %.

Abstract Background: Heart failure (HF) is a condition associated with high morbidity and mortality. Its early diagnosis in primary health care (PHC) represents a substantial challenge, considering its non-specific clinical manifestations and the limitations on timely access to diagnostic techniques. Objective: To evaluate the prevalence of HF, characterize subtypes and determine the prognosis of patients consulting in PHC for dyspnea Edema of the lower extremities. Methods: Prospective study in 340 patients who consulted in PHC, without previous diagnosis of HF. Clinical evaluation, electrocardiogram, NT-proBNP point-ofcare and echocardiography with telematic interpretation by cardiologists were performed. Using the HFA-PEFF and H2FPEF algorithms patients were classified as: 1) HF with reduced ejection fraction (HFREF); 2) HF with preserved ejection fraction (HFPEF) and 3) No HF. Actuarial survival analyses were performed. Results: We observed a prevalence of HFREF of 8%, high probability of HFPEF by HFA-PEFF in 42% and by H2FPEF in 8%. Intermediate probability of HFPEF, requiring complementary study, was observed in 47% of patients with HFA-PEFF and 76% of patients with H2FPEF (p<0.05). Overall survival at 36 months was 90±2% and cardiovascular survival at 36 months was 95±1%. Using HFA-PEFF, patients with HF presented lower overall survival compared to patients with no HF (HR 2.3, 95%CI 1.1-4.9; p=0.029). We did not observe mortality differences with H2FPEF. Conclusions: In patients consulting for dyspnea and/or lower extremity edema at PHC and undergoing evaluation with NT-proBNP and echocardiography, we observed a HF prevalence of 50%. HF classification through HFA-PEFF was associated with lower survival rates.

Resumen Introducción: La insuficiencia cardíaca (IC) tiene alta morbilidad y mortalidad. Su diagnóstico temprano en atención primaria de salud (APS) es un reto dada la baja especificidad de sus criterios clínicos y las limitaciones en acceso a técnicas diagnósticas. Objetivo: Analizar la prevalencia de IC, subtipos y pronóstico de pacientes con disnea y/o edema de extremidades inferiores que consultan en APS. Metodología: Se trata de un estudio prospectivo de 340 pacientes en APS, sin diagnóstico previo de IC. Se realizó una evaluación clínica, electrocardiograma, NT-proBNP “point-of-care”, ecocardiografía con interpretación telemática por cardiólogos. Utilizando los algoritmos HFA-PEFF y H2FPEF se clasificaron los pacientes como :1) IC con fracción de eyección (FE) reducida (ICFER); 2) IC con FE preservada (ICFEP) y 3) pacientes sin diagnóstico de IC. Se efectuó un análisis de sobrevida de los diferentes grupos. Resultados: La prevalencia de ICFER fue 8%, ICFEP por HFA-PEFF 42% y por H2FPEF 8%. Los algoritmos sugieren efectuar un estudio complementario en el 47% con HFA-PEFF y 76% con H2FPEF (p<0.05). La sobrevida global a 36 meses fue 90±2% y cardiovascular 95±1%. Usando HFA-PEFF, los pacientes con IC tuvieron menor sobrevida que aquellos sin IC (HR 2.3, IC95% 1.14.9; p=0.029). No hubo diferencias de mortalidad con H2FPEF. Conclusiones: En pacientes de APS que consultan por disnea y/o edema de extremidades inferiores sometidos a evaluación con NT-proBNP y ecocardiografía, se observó una prevalencia de IC de hasta 50%, 8% de ICFER y 42% de ICFEP. La caracterización de IC utilizando HFA-PEFF está asociada al pronóstico vital.

Abstract Tree diseases contribute to significant economic and food losses in the agricultural sector. Early detection of phytosanitary problems in trees with non-destructive methods is essential to guarantee sustainable orange production. This study presents the findings of a designed methodology conducted to identify diseased orange trees in an orchard situated in the citrus belt of Mexico, specifically in the Rioverde region of San Luis Potosi. To accomplish this, we captured images using a multispectral camera with very high spatial resolution, which was mounted on an unmanned aerial vehicle. These images were used to construct a georeferenced orthomosaic of the orchard. Six thematic classes were established to distinguish various health levels among the trees. We employed several supervised classification algorithms at the pixel level, including Random Forest (RF), K-Nearest Neighbor (KNN), Spectral Angle Mapper (SAM), Support Vector Machine (SVM), and Maximum Likelihood (ML). Considering the classification accuracy achieved by each algorithm, they can be ranked as follows: Maximum Likelihood (ML) with 88.10%, Support Vector Machine (SVM) with 77.38%, Spectral Angle Mapper (SAM) with 76.19%, K-Nearest Neighbor (KNN) with 64.68%, and Random Forest (RF) with 61.90%. These results successfully identified the phytosanitary status of all the trees in the orchard with an acceptable level of accuracy, providing valuable management information for the grower.

Resumen Las enfermedades de los árboles contribuyen a importantes pérdidas económicas y de alimentos en el sector agrícola. La detección temprana de problemas fitosanitarios en árboles con métodos no destructivos resulta fundamental para garantizar la producción sostenible de naranja. Este trabajo presenta los resultados de una metodología diseñada para la identificación de árboles de naranja enfermos en una huerta ubicada en el cinturón citrícola de México, particularmente en la región de Rioverde, San Luis Potosí. Para ello, se tomaron imágenes con una cámara multiespectral de muy alta resolución espacial instalada en un vehículo aéreo no tripulado con las que se construyó un ortomosaico georreferenciado. Se emplearon seis clases temáticas para identificar los diferentes niveles de sanidad. Se utilizaron diferentes algoritmos de clasificación supervisada a nivel píxel que incluyen Random Forest (RF), K-Nearest Neighbor (KNN), Spectral Angle Mapper (SAM), Support Vector Machine (SVM), y Maximum Likelihood (ML). Considerando la exactitud de clasificación obtenida por cada uno de los algoritmos, se pueden ordenar de la siguiente manera: Maximum Likelihood (ML) con un 88,10%, Support Vector Machine (SVM) con un 77,38%, Spectral Angle Mapper (SAM) con un 76,19%, K-Nearest Neighbor (KNN) con un 64,68% y Random Forest (RF) con un 61,90%. Los resultados permitieron identificar el estado fitosanitario de todos los árboles de la huerta, con una exactitud aceptable y representan información valiosa de manejo para el productor.

Resumen La visión por computadora (VC) puede ser un proceso que facilite algunas tareas en la gestión de inventarios, por medio de este proceso se puede realizar un análisis permanente de un inventario y así mantener registro de todos los movimientos realizados, entregando un reporte instantáneo cuando sea requerido. Esto supone una mejora en la seguridad, ya que al mantener un control estricto de los elementos existentes en el inventario se puede saber si un elemento pertenece o no a un inventario o cuando se retira o agrega un elemento, tras esta necesidad de control de inventario, surge la necesidad de diseñar un sistema inteligente que pueda facilitar el control de inventarios. Mediante la combinación de 2 frameworks, se realiza la creación de un algoritmo capaz de realizar la identificación y conteo de objetos, así como la identificación de la mano para determinar cuándo se realiza una manipulación humana al inventario. Para lograr este objetivo, se utilizaron dos algoritmos: MediaPipe y YOLOv5 combinado con el dataset de COCO, el primero se usó para la detección de manos y el segundo identifica y cuenta los objetos. Después de las pruebas realizadas al algoritmo se determina que el reconocimiento de manos de MediaPipe tuvo una precisión del 96% y la detección y clasificación de objetos usando YOLO fue de 43.7%. Teniendo como retos el algoritmo la superposición, la oclusión/auto oclusión de los objetos, o la pérdida de foco de los elementos debido al sensor.

Abstract Computer vision (CV) can be a process that facilitates some tasks in inventory management, through this process a permanent analysis of an inventory can be performed and thus keep record of all movements made, delivering an instant report when required. This means an improvement in security, since by keeping a strict control of the existing elements in the inventory it is possible to know if an element belongs or not to an inventory or when an element is removed or added, after this need for inventory control, the need arises to design an intelligent system that can facilitate inventory control. Through the combination of 2 frameworks, the creation of an algorithm capable of performing the identification and counting of objects, as well as the identification of the hand to determine when a human manipulation is performed to the inventory. To achieve this objective, two algorithms were used: MediaPipe and YOLOv5 combined with the COCO dataset, the first one was used for hand detection and the second one identifies and counts the objects. After testing the algorithm, it was determined that the hand recognition of MediaPipe had an accuracy of 96% and the detection and classification of objects using YOLO was 43.7%. Challenges for the algorithm were overlapping, occlusion/self-occlusion of objects, or loss of focus of items due to the sensor.

ABSTRACT Background: The growing number of echocardiographic tests and the need for strict adherence to international quantification guidelines have forced cardiologists to perform highly extended and repetitive tasks when interpreting and analyzing increasingly overwhelming amounts of data. Novel machine learning (ML) techniques, designed to identify images and perform measurements at relevant visits, are becoming more common to meet this obvious need for process automation. Objectives: Our objective was to evaluate an alternative model for the interpretation and analysis of echocardiographic tests mostly based on the use of ML software in order to identify and classify views and perform standardized measurements automatically. Methods: Images came from 2000 healthy subjects, 1800 of whom were used to develop ML algorithms and 200 for subsequent validation. First, a convolutional neural network was developed in order to identify 18 standard echocardiographic views and classify them based on 8 thematic groups (stacks). The results of automatic identification were compared to classification by experts. Later, ML algorithms were developed to automatically measure 16 Doppler scan parameters for regular clinical evaluation, which were compared to measurements by an expert reader. Finally, we compared the time required to complete the analysis of an echocardiographic test using conventional manual methods with the time needed when using the ML model to classify images and perform initial echocardiographic measurements. Inter- and intra-observer variability was also analyzed. Results: Automatic view classification was possible in less than 1 second per test, with a 90% accuracy for 2D images and a 94% accuracy for Doppler scan images. Stacking images had a 91% accuracy, and it was possible to complete the groups with any necessary images in 99% of cases. Expert agreement was outstanding, with discrepancies similar to those found between two human readers. Applying ML to echocardiographic imaging classification and measurement reduced time of analysis by 41% and showed lower variability than conventional reading methods. Conclusion: Application of ML techniques may significantly improve reproducibility and efficiency of echocardiographic interpretations and measurements. Using this type of technologies in clinical practice may lead to reduced costs and increased medical staff satisfaction.

RESUMEN Introducción: El número creciente de estudios ecocardiográficos y la necesidad de cumplir rigurosamente con las recomendaciones de guías internacionales de cuantificación, ha llevado a que los cardiólogos deban realizar tareas sumamente extensas y repetitivas, como parte de la interpretación y análisis de cantidades de información cada vez más abrumadoras. Novedosas técnicas de machine learning (ML), diseñadas para reconocer imágenes y realizar mediciones en las vistas adecuadas, están siendo cada vez más utilizadas para responder a esta necesidad evidente de automatización de procesos. Objetivos: Nuestro objetivo fue evaluar un modelo alternativo de interpretación y análisis de estudios ecocardiográficos, basado fundamentalmente en la utilización de software de ML, capaz de identificar y clasificar vistas y realizar mediciones estandarizadas de forma automática. Material y métodos: Se utilizaron imágenes obtenidas en 2000 sujetos normales, libres de enfermedad, de los cuales 1800 fueron utilizados para desarrollar los algoritmos de ML y 200 para su validación posterior. Primero, una red neuronal convolucional fue desarrollada para reconocer 18 vistas ecocardiográficas estándar y clasificarlas de acuerdo con 8 grupos (stacks) temáticos. Los resultados de la identificación automática fueron comparados con la clasificación realizada por expertos. Luego, algoritmos de ML fueron desarrollados para medir automáticamente 16 parámetros de eco Doppler de evaluación clínica habitual, los cuales fueron comparados con las mediciones realizadas por un lector experto. Finalmente, comparamos el tiempo necesario para completar el análisis de un estudio ecocardiográfico con la utilización de métodos manuales convencionales, con el tiempo necesario con el empleo del modelo que incorpora ML en la clasificación de imágenes y mediciones ecocardiográficas iniciales. La variabilidad inter e intraobservador también fue analizada. Resultados: La clasificación automática de vistas fue posible en menos de 1 segundo por estudio, con una precisión de 90 % en imágenes 2D y de 94 % en imágenes Doppler. La agrupación de imágenes en stacks tuvo una precisión de 91 %, y fue posible completar dichos grupos con las imágenes necesarias en 99% de los casos. La concordancia con expertos fue excelente, con diferencias similares a las observadas entre dos lectores humanos. La incorporación de ML en la clasificación y medición de imágenes ecocardiográficas redujo un 41 % el tiempo de análisis y demostró menor variabilidad que la metodología de interpretación convencional. Conclusión: La incorporación de técnicas de ML puede mejorar significativamente la reproducibilidad y eficiencia de las interpretaciones y mediciones ecocardiográficas. La implementación de este tipo de tecnologías en la práctica clínica podría resultar en reducción de costos y aumento en la satisfacción del personal médico.

Resumen La inteligencia artificial ha permitido desarrollar herramientas para el reconocimiento automático de frutos y hortalizas con mayor precisión y rapidez. El desarrollo de nuevos genotipos de árboles frutales requiere del uso de herramientas tecnológicas para la identificación de variedades con mayor robustez que los métodos convencionales. En esta investigación se aplicaron algoritmos de aprendizaje automático para identificar seis variedades de durazno (Prunus persica L.) CP-03-06, Oro Azteca, Oro San Juan, Cardenal, Colegio y Robin a partir de imágenes digitales de hojas. Los modelos máquina de soporte vectorial (SVM), bosque aleatorio (RF) y perceptrón multicapa (MLP) se entrenaron y evaluaron con base en tres descriptores cromáticos y 14 morfológicos extraídos de imágenes digitales. La evaluación del desempeño en predicción de los modelos se realizó con base en métricas globales y específicas para cada clase objetivo (variedad de durazno). Los cinco descriptores más importantes para identificar las variedades de durazno fueron los tres canales de color HSV (hue, saturation, value), la redondez y la excentricidad de las hojas. SVM obtuvo la mayor precisión global de clasificación con Acc de 98.7 % y F1macro de 98 %. SVM obtuvo el mayor puntaje F1 (99.2 %) para identificar la variedad de durazno CP-03-06 y el menor puntaje F1 (96.1 %) para identificar la variedad Cardenal. La utilización conjunta de descriptores cromáticos y morfológicos mejoró el desempeño de los algoritmos de aprendizaje para identificar las seis variedades de durazno. Los modelos SVM, RF y MLP obtuvieron un Acc de 98.7, 98.6 y 97 %, respectivamente. Este estudio muestra el potencial de las técnicas de aprendizaje automático para su aplicación en el reconocimiento de descriptores de interés en cultivos agrícolas y su aplicación a procesos automatizados en la agricultura.

Summary Artificial intelligence has allowed the development of tools for automatic recognition of fruits and vegetables with greater precision and speed. The development of new genotypes of fruit trees requires the use of technological tools to identify varieties with greater robustness than conventional methods. In this research, machine learning algorithms were applied to identify six peach varieties (Prunus persica L.) CP-03-06, Oro Azteca, Oro San Juan, Cardenal, Colegio and Robin from digital leaf images. The support vector machine (SVM), random forest (RF) and multilayer perceptron (MLP) models were trained and evaluated based on three chromatic and 14 morphological descriptors extracted from digital images. The evaluation of the prediction performance of the models was based on global metrics and specific for each target class (peach variety). The five most important descriptors to identify peach varieties were three HSV color channels (hue, saturation, value), roundness and eccentricity of the leaves. SVM achieved the highest overall classification accuracy with Acc of 98.7 % and F1macro of 98 %. SVM obtained the highest F1 score (99.2 %) to identify the peach variety CP-03-06 and the lowest F1 score (96.1 %) to identify the Cardenal variety. The joint use of chromatic and morphological descriptors improved the performance of learning algorithms to identify the six peach varieties. The SVM, RF and MLP models obtained an Acc of 98.7, 98.6 and 97 %, respectively. This study shows the potential of machine learning methods for their application in recognizing descriptors of interest in agricultural crops and their application to automated processes in agriculture.

RESUMO Objetivo: Esse estudo tem como objetivo criar um modelo de Machine Learning por um oftalmologista sem experiência em programação utilizando auto Machine Learning predizendo influxo de pacientes em serviço de emergência e casos de trauma. Métodos: Um dataset de 366,610 visitas em Hospital Universitário da Universidade Federal de São Paulo de 01 de janeiro de 2014 até 31 de dezembro de 2019 foi incluído no treinamento do modelo, incluindo visitas/dia e código internacional de doenças. O treinamento e predição foram realizados com o Amazon Forecast por dois oftalmologistas sem experiência com programação. Resultados: O período de previsão estimou um volume de 206,37 pacientes/dia em p90, 180,75 em p50, 140,35 em p10 e média de 7,42 casos de trauma/dia em p90, 3,99 em p50 e 0,56 em p10. Janeiro de 2020 teve um total de 6.604 pacientes e média de 206,37 pacientes/dia, 13,5% menos do que a predição em p50. O período teve um total de 199 casos de trauma e média de 6,21 casos/dia, 55,77% mais casos do que a predição em p50. Conclusão: O desenvolvimento de modelos era restrito a cientistas de dados com experiencia em programação, porém a transferência de ensino com a tecnologia de auto Machine Learning permite o desenvolvimento de algoritmos por qualquer pessoa sem experiencia em programação. Esse estudo mostra um modelo com valores preditos próximos ao que ocorreram em janeiro de 2020. Fatores que podem ter influenciados no resultado foram feriados e tamanho do banco de dados. Esse é o primeiro estudo que aplicada auto Machine Learning em predição de visitas hospitalares com resultados próximos aos que ocorreram. Objetivo Métodos 366610 366 610 366,61 0 201 3 visitasdia dia doenças Resultados 20637 206 37 206,3 pacientesdia p90 p 18075 180 75 180,7 14035 140 35 140,3 p1 742 7 42 7,4 traumadia 399 99 3,9 p5 056 56 0,5 202 6604 6 604 6.60 135 13 5 13,5 19 621 21 6,2 casosdia casos/dia 5577 55 77 55,77 Conclusão 36661 36 61 366,6 20 2063 206, p9 1807 18 180, 1403 14 140, 74 4 7, 39 9 3, 05 0, 660 60 6.6 1 13, 62 2 6, 557 55,7 3666 366, 66 6. 55,

ABSTRACT Purpose: The emergency medical service is a fundamental part of healthcare, albeit crowded emergency rooms lead to delayed and low-quality assistance in actual urgent cases. Machine-learning algorithms can provide a smart and effective estimation of emergency patients’ volume, which was previously restricted to artificial intelligence (AI) experts in coding and computer science but is now feasible by anyone without any coding experience through auto machine learning. This study aimed to create a machine-learning model designed by an ophthalmologist without any coding experience using AutoML to predict the influx in the emergency department and trauma cases. Methods: A dataset of 356,611 visits at Hospital da Universidade Federal de São Paulo from January 01, 2014 to December 31, 2019 was included in the model training, which included visits/day and the international classification disease code. The training and prediction were made with the Amazon Forecast by 2 ophthalmologists with no prior coding experience. Results: The forecast period predicted a mean emergency patient volume of 216.27/day in p90, 180.75/day in p50, and 140.35/day in p10, and a mean of 7.42 trauma cases/ day in p90, 3.99/day in p50, and 0.56/day in p10. In January of 2020, there were a total of 6,604 patient visits and a mean of 206.37 patients/day, which is 13.5% less than the p50 prediction. This period involved a total of 199 trauma cases and a mean of 6.21 cases/day, which is 55.77% more traumas than that by the p50 prediction. Conclusions: The development of models was previously restricted to data scientists’ experts in coding and computer science, but transfer learning autoML has enabled AI development by any person with no code experience mandatory. This study model showed a close value to the actual 2020 January visits, and the only factors that may have influenced the results between the two approaches are holidays and dataset size. This is the first study to apply AutoML in hospital visits forecast, showing a close prediction of the actual hospital influx. Purpose healthcare lowquality low quality Machinelearning Machine patients (AI machinelearning Methods 356611 356 611 356,61 01 201 31 visitsday Results 21627day 216 27 p90 p 18075day 180 75 14035day 140 35 p10 742 7 42 7.4 399day 3 99 056day 0 56 6604 6 604 6,60 20637 206 37 206.3 patientsday patients/day 135 13 5 13.5 p5 19 621 21 6.2 casesday cases/day 5577 55 77 55.77 Conclusions scientists mandatory 202 size 35661 61 356,6 20 p9 18 14 p1 74 4 7. 9 660 60 6,6 2063 206. 1 13. 62 6. 557 55.7 3566 356, 66 6, 55.

It is important to identify the types of tobacco diseases accurately and take effective control measures in time to improve the efficiency of tobacco planting. In this paper, a hand-held near-infrared spectrometer was used to collect the spectral data of different types of tobacco disease samples. The training models were established via convolutional neural network algorithm. Meanwhile, the traditional classification algorithms support vector machine and back propagation neural network were also compared. The results showed that the prediction accuracy of convolutional neural network algorithm was the highest and the overall performance of the model was the best. The rapid detection method based on a hand-held near-infrared spectrometer and convolutional neural network algorithm could identify tobacco leaf disease species efficiently, non-destructively, quickly and accurately, which provided a new technical reference for tobacco leaf disease species detection and identification. planting paper handheld hand held nearinfrared near infrared samples Meanwhile compared best efficiently nondestructively, nondestructively non destructively, destructively non-destructively identification