Monserrate Intriago-Pazmiño

Tax regulation in Ecuador changes constantly, and the resulting information is scattered across laws, regulations, resolutions, forms, and annexes with strong ties across meaning and time. We compare four retrieval-augmented generation (RAG) strategies for question answering over Ecuadorian tax legislation (vector-only, GraphRAG, hybrid, and agentic retrieval) in a system that supports text, speech, and image inputs and combines a tax knowledge graph, dense vector search, and three agents that refine, validate, and plan each query’s strategy. Following a Design Science Research approach informed by a SALSA literature review, the system was tested on a curated 51-document corpus using 100 text-only tax queries across the four retrieval modes and two cloud generator models (gemma4:31b-cloud and gpt-oss:20b-cloud), for 800 RAGAS-judged executions, plus a Likert-scale review by five domain specialists. Speech- and image-input performance was not evaluated in this benchmark. Graph-only retrieval was weakest on all five RAGAS metrics (Context Precision = 0.02), while hybrid gave the best balance (Faithfulness = 0.67, Answer Relevancy = 0.64, Context Precision = 0.52, Context Recall = 0.76) at the lowest latency (2.66 s). Faithfulness depended more on the generator model than on retrieval mode (0.78 vs. 0.47). No strategy won on every metric, so architecture choice in dynamic regulatory domains should follow operational priority rather than a single universally optimal setup.

Background: Preeclampsia remains a major cause of maternal and perinatal morbidity and mortality, and machine learning (ML) and artificial intelligence (AI) models have increasingly been proposed for risk prediction, diagnosis, monitoring, and prognosis. However, the extent to which key methodological details are visible from abstracts and bibliographic metadata remains unclear. Methods: We conducted an abstract-level systematic mapping review of ML/AI studies for preeclampsia (PE), hypertensive disorders of pregnancy (HDP), and directly related complications. Records were identified from ACM Digital Library, Elsevier/ScienceDirect, IEEE Xplore, PubMed, Scopus, SpringerLink, and snowballing. After DOI-based deduplication, abstract-availability checks, technical metadata filtering, and a two-stage eligibility audit, 29 primary ML/AI studies were retained for the final synthesis. Two authors independently re-audited the 145 technically retained records for eligibility. Initial agreement was 131/145 records (90.3%); for the binary include/exclude decision, Cohen’s kappa was approximately 0.764, indicating substantial agreement. Extraction was limited to title, abstract, year, DOI, and bibliographic metadata using a predefined RQ-aligned schema. Therefore, all frequencies represent information explicitly extractable at the abstract level and should be interpreted as lower-bound reporting estimates rather than full-text prevalence estimates. Results: Among the 29 eligible studies, tree-based models were the most frequently identifiable family (17/29), followed by linear/generalized linear models (15/29), support vector machines (9/29), and neural networks (8/29). The most frequently identifiable input modalities were EHR/clinical variables (21/29), demographics (10/29), laboratory variables (7/29), vital signs/blood pressure (7/29), and ultrasound/Doppler information (6/29). AUC was mentioned in 20/29 abstracts, but a numeric AUC value was extractable into the structured field in 14/29. These values were summarized descriptively and were not pooled because outcomes, prediction horizons, populations, predictors, and validation strategies were heterogeneous. External validation was explicitly identifiable in 8/29 abstracts, explainability-related methods in 5/29, and repository or dataset-source information in 1/29. Conclusions: The abstract-level evidence shows an active but heterogeneous ML/AI literature for PE/HDP prediction and related outcomes. However, abstracts often provide insufficient detail for robust comparative assessment of model validity, calibration, clinical readiness, reproducibility, and implementation. Future work should prioritize full-text evidence synthesis, standardized reporting, external and temporal validation, calibration, equity assessment, clinically meaningful explainability, and reusable data/model resources.

Background: Preeclampsia (PE) is a leading cause of maternal and perinatal morbidity, affecting 2–8% of pregnancies worldwide. In Ecuadorian public hospitals, clinical data are often fragmented, limiting the construction of complete datasets for prediction modeling. This study aimed to develop a Machine Learning (ML) framework to predict PE using demographic data from the Isidro Ayora Gynecology and Obstetrics Hospital (HGOIA), Quito, Ecuador. Methods: We proposed a custom automated Machine Learning (C-AutoML) workflow to evaluate six models (Logistic Regression, LightGBM, XGBoost, CatBoost, Multi-Layer Perceptron, and Random Forest) and compared its performance against a weighted ensemble built with Amazon SageMaker Canvas, optimizing for the F1-score. Model interpretability was analyzed using Shapley Additive Explanations (SHAP). Due to the scarcity of clinical variables at HGOIA, a synthetic dataset based on FullPIERS predictors was also generated and modeled. Results: On the hospital dataset, the SageMaker ensemble achieved an F1-score of 0.805, accuracy of 0.934, and AUC of 0.956, while the C-AutoML best single model (LightGBM) yielded comparable results (F1-score: 0.802, accuracy: 0.938, AUC: 0.944). SHAP analysis identified patient age as the most influential feature. On the synthetic dataset, Logistic Regression achieved perfect classification (F1-score: 1.0). Conclusions: The proposed C-AutoML framework matched the performance of the SageMaker weighted ensemble while relying on a single, interpretable model. Despite the current dependence on demographic data, the methodology shows promise for supporting PE diagnosis in resource-limited settings; future integration of clinical variables is expected to further enhance predictive power for risk stratification.

Automatic weather recognition using images has important applications in land and air traffic control, autonomous vehicles, road safety, and crop management. Despite its relevance, this field remains underexplored due to the difficulty of extracting robust features for weather conditions that often overlap visually, and the high cost of building large, labeled datasets. To address these challenges, we propose a Multistep Visual Pretraining (MVP) approach, a resource-efficient training pipeline designed to maximize performance with limited supervision. MVP leverages self-supervised pretraining on a large-scale unlabeled weather dataset, followed by supervised fine-tuning on a small labeled dataset, while reusing ImageNet pretrained models to minimize computational requirements. Unlike traditional strategies that demand massive manually labeled data and extensive GPU resources, MVP achieves competitive performance with a compact ResNet-50 architecture, only 1200 labeled images for fine-tuning, and a single mid-range GPU. Our method reached 94% accuracy in multi-class weather recognition, demonstrating that robust, real-world performance can be achieved without the prohibitive costs of large-scale annotation or high-end computing infrastructure. This study extends our previous work, “Weather Recognition Using Self-Supervised Deep Learning”, going from binary to multi-class classification while reducing data and resource dependency, thus offering a scalable and practical solution for real-world deployment.

This work addresses the development of a hybrid method for feature selection and a strategy to classify quite large datasets of handwritten characters. The divide and conquer paradigm, generally, is used to divide a big problem into minor problems. This research applied this concept to recognize handwritten uppercase letters and numbers. As a result, a big problem is split into two nodes or subproblems, one for numbers and one for letters. Then, letters are divided into two nodes representing the straight and curved ones. The division can be called the binary decision tree and allows to obtain a subset with the minimal features of each node called reduct. Here, an improvement of reducts is proposed using the ant colony algorithm as the embedded method. The application of these methods had the following result and conclusions. For each node, subsets of fewer features were obtained with high performance in the classification, considering the morphology of each letter. It is crucial to highlight that the distribution of the samples affects the performance of the classifier and the strategy improves the performance of the reduct.