Applied Informatics

This study addresses the implementation of artificial intelligence methods for assisted quantification of biomarkers in sporadic colorectal cancer . Accurate artificial intelligence models are crucial for medical applications such as diagnosis and prognosis of this major disease, which represents a growing cause of death worldwide and in Argentina. In low and middle-income countries such as our own, access to automatic image analysis platforms is limited due to their high cost. To address this, we are developing a workflow that explores automatic segmentation and a simplified immunological score to evaluate antibody expression in tumoral tissue images of sporadic colorectal cancer. The processes explored in the workflow include multiclass semantic segmentation with U-Net neural network, correlation between biomarkers CD3, CD8, CD146 and PDGF-Rβ, and patient data analysis using machine learning models. Results indicated that the U-Net transfer learning model achieved the highest precision in segmenting the tumor core and invasive margin regions with intersection over union scores of 0.94 and 0.88 respectively. However reduced performance was noted in underrepresented regions, emphasizing the necessity for improved data balance. From the Pearson coefficient, a pattern of strong correlations was observed between the biomarkers and their expression within the tumor core. Random Forest analysis of the patient data evidenced the three most important features: TMN 18.48%, CD146 16.20% and CD8 12.24%. Our goal is to establish an analysis workflow that streamlines research efforts by facilitating the characterization of new potential biomarkers and, upon validation and implementation, enhances diagnostic processes for pathologists.

Ecuadorian cuisine is very varied, with typical dishes from various regions often using fresh herbs. Among them, coriander and parsley are frequently used but often confused because of their similar appearance. This confusion can result in inexperienced or visually impaired people using the wrong herb, thus altering the flavor and unique qualities of the dishes. To address this problem, in this work, we present an image classification model capable of distinguishing between coriander and parsley leaves. Three architectures (MobileNetV2, EfficientNetB0 and InceptionV3) were evaluated to determine which offered the highest accuracy. MobileNetV2 proved to be the most efficient model, demonstrating superior stability and achieving greater than 99% accuracy.

This study explores the challenges of identifying forest areas in the “Carolina” Park in Quito, Ecuador, using Convolutional Neural Networks (CNN) and aerial imagery to support sustainable urban expansion plans. A dataset was constructed using 32\(\,\times \,\)32\(\,\times \,\)3 pixel patches extracted from 230 aerial images obtained from several videos captured by drones over the park. Three CNN models (ResNet50, EfficientNetB0 and VGG16) were trained to differentiate between forested, non-forested and hybrid areas. The methodology involved manual classification of 2100 patches into these three categories. The results showed that ResNet50 performed the best overall, with an accuracy of 76.66% ± 8%, followed closely by VGG16, while EfficientNetB0 showed inferior performance on this specific dataset. Qualitative analysis of predictions on test images confirmed the effective identification of forest areas. These findings suggest that ResNet50 may be a suitable model for this task, demonstrating a high ability to learn and recognize patterns in forested areas through patch-based analysis, even with relatively small datasets derived from aerial drone imagery.

Assessing autonomic function using heart rate variability (HRV) in critically ill patients can provide valuable information about prognosis and treatment options. We implemented a novel convolutional neural network (CNN) algorithm to categorize the status of the autonomic nervous system (ANS) into three classes: basal state (BS) without any system dominance, sympathetic nervous system dominance state (SDS), and parasympathetic nervous system dominance state (PDS). For the BS class, we achieved an accuracy of 89%, a sensitivity of 92%, and a specificity of 89%. For the SDS class, we obtained an accuracy of 98%, a sensitivity of 95%, and a specificity of 98%. And for the PDS class, an accuracy of 87%, a sensitivity of 83%, and a specificity of 91% was obtained. We also developed a model to differentiate ANS activity between survivors and deceased during the first 24 h of hospitalization, which was evaluated using data from the MIMIC-III database. Results showed that in the deceased group the PDS decreased by 9.70% (p < 0.05) in contrast to a 8.02% increase in the SDS class (p < 0.05). This work provides a promising tool for implementation in the ICU, complementing fast and reliable decision-making by physicians during hospitalization.

This study aims to enhance the accuracy and efficiency of oral cancer diagnosis through the application of deep learning techniques in medical image analysis. The research employs convolutional neural networks (CNNs), specifically ResNet and DenseNet architectures, for the classification of oral cancer images into malignant and benign categories. Data preprocessing involves resizing, normalization, and augmentation to optimize model performance. Evaluation metrics including accuracy, loss, specificity, and sensitivity demonstrate varying performance across different CNN models. DenseNet architectures consistently outperform ResNet and conventional CNNs in terms of accuracy and sensitivity metrics. The results showed that DenseNet consistently outperformed ResNet, achieving higher accuracy and sensitivity, which are crucial for early cancer detection. The findings underscore the transformative potential of deep learning in augmenting clinical decision-making for oral cancer detection. Integration of these advanced technologies into healthcare workflows could significantly improve early detection rates and treatment outcomes, paving the way for personalized medicine approaches in oncology.

In the current era of software engineering, where Machine Learning (ML) plays a crucial role in technological innovation, effective implementation of development and operations practices is essential. The DevSecOps (Development, Security, Operations) approach has gained popularity due to its ability to integrate security and quality into every stage of the software development lifecycle. However, in the specific context of Machine Learning, the need arises for a specialized approach that takes into account the peculiarities of the models and algorithms used. Machine Learning Operations (MLOps), despite their relative novelty, seek to establish a framework to characterize the ML development life cycle, decouple it from software development, and ensure quality attributes such as scalability, maintainability, and security. This paper focuses on the exploration and application of Machine Learning Operations (MLOps) in the specific context of developing and provisioning machine learning models.

The COVID-19 pandemic has catalyzed a substantial increase in the availability of online academic resources. Users and students frequently encounter challenges in identifying the most pertinent content amidst this vast information space. Consequently, recommendation systems have become indispensable for curating content by leveraging data from prior interactions and resource characteristics. While significant research has been conducted in this domain, existing methodologies are often constrained to specific information types, resulting in the underutilization of diverse data sources. In contrast, fields such as entertainment and e-commerce have successfully employed deep learning models to integrate various data forms, including videos, user-item interactions, and thematic content. However, the application of these advanced techniques has yet to be explored in academic contexts. This study explores the integration of unstructured data, such as descriptions, and structured data, such as knowledge graphs, in the context of academic recommendation systems. We utilize two distinct datasets: MOOCCubeX, an open MOOC dataset, and DBLPv12, a citation network dataset, to evaluate the impact of diverse information sources on recommendation performance. Our findings indicate that deep learning models incorporating heterogeneous data types significantly enhance the accuracy and relevance of recommendations, thereby demonstrating the potential of these approaches in academic resource recommendation.

Automatic Music Transcription (AMT) entails creating an algorithm that converts an acoustic signal from an audio file into the corresponding sheet music representation. This paper uses deep learning methods and models AMT as a translation problem, comparing the effectiveness of an instance-based translation approach using an MLP to a sequence-based approach using an RNN. The models were trained on the EsAc dataset and evaluated using MUSTER metrics. The results show that the instance-based model better classifies the correct pitch. However, the sequence-based approach outperforms the instance-based approach on all other aspects of the MUSTER metrics, producing a 98% accuracy.

The use of Wireless Body Area Networks (WBAN) in healthcare is transforming how patients and doctors communicate and how patient data is collected. WBAN allows for the real-time transmission of patient information to healthcare providers, reducing the need for constant in-person monitoring and hospital stays for data collection. However, WBAN technology is vulnerable to significant cybersecurity risks, such as eavesdropping and man-in-the-middle attacks, which can compromise the integrity and confidentiality of user data. Common threats include Denial of Service (DoS), Spoofing, and Replay attacks. To mitigate these risks, advanced security protocols, latency assessments, and intrusion detection systems are necessary. Machine learning and artificial intelligence (AI) in Cyber-Physical Systems (CPS) for healthcare offer promising solutions for continuous monitoring and threat prediction. AI systems can enhance security by detecting patterns indicative of cyber threats and deploying adaptive defenses. This study emphasizes the importance of robust security architectures and innovative AI applications to protect patient data in WBAN and CPS, ensuring secure and reliable healthcare services.

Natural disasters pose a continuous threat to populations worldwide, with hydrometeorological disasters standing out due to their unpredictability, rapid onset, and significant destructive capacity. Consequently, countries invest substantial resources in implementing pre- and post-disaster measures, including prevention, and reconstruction. Developing countries, however, face stronger budgetary constraints and continuously depend on international support, having a limited and fluctuating ability to implement optimal disaster response strategies, affecting their response capabilities and timeliness, both of which are crucial for saving lives. This paper addressed these challenges by training and implementing YOLO (You Only Look Once) Convolutional Neural Network version 8 models, using high-resolution satellite images from the Maxar GeoEye-1 satellite. Additionally, this paper tested the viability of publicly publishing the model to be accessed via a REST API, providing a low-cost tool for quickly identifying damaged buildings following natural disasters, and enhancing the efficiency and effectiveness of disaster response efforts.

Breast cancer is a significant global health challenge, as it is the main cause of cancer-related deaths among women. It is a heterogeneous disease, manifesting in diverse clinical outcomes influenced by its intricate molecular characteristics. Accurate survival prediction for breast cancer patients is essential for guiding treatment strategies and improving their outcomes, which drives the exploration of various -omics datasets through advanced machine learning techniques.

This study aims to enhance the accuracy of breast cancer survival prediction by introducing novel feature extraction methods combined with rigorous feature selection processes, leveraging multi-omics data for a more comprehensive analysis. Specifically, we evaluate the predictive power of various -omics datasets, including RNA, microRNA, and protein expression levels, DNA copy number variations, somatic mutation positions, and methylation levels. These datasets are used to create and compare single- and multi-omic models capable of predicting 5-year survival rates for breast cancer. Additionally, we assess the impact of different feature selection and data aggregation techniques on patient classification accuracy, as measured by the area under the ROC curve (AUC).

Through the analysis of multi-omic data from 178 patients, we identified 35 key molecular predictors, which enabled the development of a robust 5-year survival prediction model, achieving an AUC of 0.89 in nested cross-validation. Our findings highlight the potential of integrating multi-omics data with advanced machine learning techniques to significantly improve breast cancer survival predictions. Moreover, we demonstrate that mRNA levels, gene copy number changes, and methylation levels are among the most effective predictors for this purpose.

Cloud computing is a well-established technology that has been already widely used due to its extensive benefits. However, with many systems still relying on traditional architectures, existing literature has focused on aiding in their migration process. Nonetheless, comprehensive studies integrating both white and grey literature to assist professionals and researchers in understanding strategies for migrating legacy systems to the cloud are lacking. We addressed this gap by identifying challenges and opportunities related to migrating from on-premises to a cloud environment. Following this, we first conducted a systematic literature mapping to summarize the knowledge regarding migrating legacy systems to the cloud. Then, we performed an exploratory analysis of discussions on Stack Overflow and other question-and-answer (Q&A) communities within the Stack Exchange network to gather professionals’ perspectives on this topic and compare these perspectives with the knowledge found in the literature. Finally, we developed a Proof-of-Concept (PoC) of a decision support tool using a Large Language Model (LLM) that provides targeted responses to questions about migrating legacy systems to the cloud, enhanced by the Retrieval-Augmented Generation (RAG) method.

Malnutrition remains a critical issue affecting children worldwide, leading to severe health, developmental, and mortality outcomes. Despite significant global efforts to combat this problem, corruption frequently undermines the effectiveness of monetary aid for nutritional programs. Funds often fail to reach their intended recipients due to misappropriation, fraud, and inefficiencies, exacerbating the plight of vulnerable children. However, advances in technology present new avenues to address these challenges. Digital tools, mobile applications for monitoring nutritional intake, and data analytics for optimizing resource allocation can enhance accountability and efficiency. Here, we present a scalable model of permanent monitoring, data collection, and real-time reporting that has the potential to mitigate the impact of corruption, ensuring that aid reaches those in need and improving the nutritional status of children globally. The presented proof of concept displays a framework in a remote region where child malnutrition is endemic, and the proposed technological platform supports the operation with real-time reporting and data analytics on demand.

The paper describes a longitudinal research conducted in a Health Service Pro-vider Institution in a rural area of Cundinamarca, between 2017 and 2022. The methodology used was VEDECO or Connected Planning. Demographic trends and access to health services were examined, using descriptive and diagnostic analyses to understand correlations between variables. The population served showed an expansive population pyramid with a high pro-portion of children and young people, although with a decrease in young adults, possibly due to migration to urban areas. Changes in the demographic structure were observed, such as an increase in the population over 50 years of age in 2020, related to the COVID-19 pandemic. The affiliation regime showed that the majority of users were from the subsidized regime, followed by the contributory regime. Outpatient consultation was the most used service, especially for chronic diseases such as hypertension and dia-betes. In the emergency department, consultations for trauma and poisoning pre-dominated, with a strong association with the female gender, suggesting possible implications of gender violence. Geographic analysis revealed that proximity to the health center influences the frequency of service use, with visits decreasing as the distance of residence from the center increases. This underscores the importance of geographic access in the utilization of health services. In summary, the study provided insights into the demographic structure, service use patterns, and geographic factors that affect health care in a rural community, informing recommendations for improving access and quality of health care.

Cities grow exponentially at the same time as the demand for food, needing to increase the production of supplies and its costs due to farmers having to apply chemicals and do genetic modifications to harvest to respond to this demand. There are some technologies to get information from the harvest through devices that can capture hyperspectral images, allowing them to see things we cannot with our vision. They are used to get information from crops, like plant health and vitality, plant stress, disease and pest detection, soil cover, and crop maturity. However, due to the cost of hyperspectral cameras and their lack of spatial resolution, this paper aims to explore models for getting those images from RGB images obtained with more common and cheaper cameras.

The development of AI-driven geolocation technology is important for applications in business, sustainable development, education, and other fields. This study introduces a geolocation framework utilizing social media data from platforms like Twitter, Instagram, and YouTube. Despite the challenge that only 1% to 3% of tweets are georeferenced, our framework employs Boolean queries, geocoding tools, and stochastic models to accurately infer user locations. This automated process handles data extraction, validation, preparation, and model training, ensuring efficiency and consistency. The model achieves high accuracy, particularly in predicting major U.S. cities, and offers a scalable solution for converting ambiguous social media data into actionable insights, supporting informed decision-making across various sectors.

This study evaluates the performance of various recommendation system algorithms using anonymized real-World data sets from a commercial group in the retail sector located in the northern coast of Colombia. We evaluate several Recommender System algorithms, including Singular Value Decomposition (SVD), Non-Negative Matrix Factorization (NNMF), and collaborative filtering techniques, coupled with different data normalization techniques. The results indicate that the performance of these algorithms varies significantly depending on the data volume and normalization methods used. The study highlights the potential of Recommendation Algorithms for the Colombian Retail Sector to increase sales and improve inventory turnover. Future research should explore memory-based techniques and the integration of additional data sources, such as social media and web scraping, to further enhance recommendation accuracy and relevance.

Insurance claim fraud is an escalating concern for Colombia’s insurance sector, particularly the Mandatory Traffic Accident Insurance, or SOAT. This study addresses this issue by applying supervised machine-learning techniques to enhance fraud detection and prevention. A comprehensive analysis of a historical dataset provided by Valuative SAS, which includes over one million claim records, evaluated multiple classification models, including Support Vector Machines, Random Forest, XG-Boost, and neural networks. The results demonstrate that the selected models can identify fraud patterns with high precision, offering significant potential to reduce financial losses and increase the sustainability of the SOAT insurance system in Colombia. This work proposes a replicable and scalable methodology to combat insurance fraud at both national and international levels.

Educational games have gained prominence by offering dynamic and engaging ways to impart knowledge, providing an alternative to traditional teaching methods. In this context, Role-Playing Games (RPGs) feature rich scenarios and immersive narratives that capture students’ attention. This is especially relevant in programming education, where maintaining student engagement can be challenging. However, there is a lack of such educational games targeted at high school students. In light of this, this article explores the implementation of the educational RPG “Code Legends,” developed as a complementary tool for teaching programming in high school. To achieve this goal, the game divides the learning process into journeys containing various missions, corresponding to different programming concepts. Each mission consists of challenges that students must overcome to progress in the game, ensuring gradual and systematic learning. The methodology involved a literature review to build a solid theoretical foundation and develop a conceptual model, which was implemented on the Classcraft platform. Additionally, a pilot test was conducted with volunteer students, followed by a feedback form to collect their impressions and suggestions about the game. The study results show that “Code Legends” is a promising approach to interactive programming education. Based on the feedback received, areas for improvement were identified, but the overall experience suggests that using educational RPGs can be an effective strategy to complement programming education in high school.

This study addresses global challenges posed by cognitive aging and disorders like dementia, emphasizing the need for early detection of mild cognitive impairment (MCI) to enable timely interventions. An innovative approach was proposed using a Unity-based video game incorporating mini-games adapted from the Montreal Cognitive Assessment (MoCA) to assess cognitive functions. The Huddle methodology guided the project through pre-production, production, and postmortem phases. Functionality tests demonstrated effective communication of game objectives, highlighting accessibility for non-gamers and achieving high user satisfaction with visual elements. Players found it easy to interrupt gameplay, indicating a well-designed user experience. Analysis of performance in MoCA-adapted mini-games revealed competence in attention, memory, and abstraction, though perceptions of game difficulty varied among participants. Overall, the video game showed promise in enabling effective early detection of MCI. While further research is necessary to validate its clinical efficacy, preliminary findings suggest that video games can serve as engaging and effective tools for early cognitive assessment. The study's outcomes, including clear communication of objectives, accessibility enhancements, and competent cognitive function evaluation, support the feasibility and potential utility of this innovative approach.