Producción científica y docente
Producción científica reciente
SPE-ECG: An Ontology-Guided Knowledge Representation Framework for Enhancing LLM-Based Code Generation
Large Language Models (LLMs) can translate natural-language instructions into code, but when the input is unstructured clinical protocol text, the generated code may omit doses, monitoring intervals, or safety thresholds that the text states. Such omissions matter because they can propagate into software that supports clinical monitoring. We propose SPE-ECG (Structured Prompt Engineering for Enhanced Code Generation), a transparent, rule-based pipeline that converts unstructured domain text into a JSON prompt using text cleaning, part-of-speech tagging, bigram matching, and hand-built domain dictionaries, without fine-tuning any model. In a case study of 43 rheumatoid arthritis DMARD monitoring protocols, we compared raw and structured prompts on three LLMs (GPT-3.5, LLaMA 2, and Gemma), scoring the generated Python code using version 3.15.1 with BLEU against the raw-prompt code of the other two models. Mean BLEU rose for all three models (relative gains of 55.1%, 32.5%, and 19.9%); the paired differences were significant for GPT-3.5 and LLaMA 2 but not, after correction for multiple comparisons, for Gemma. These results suggest that lightweight input structuring can change how closely different models’ outputs agree, and that it could serve as a low-cost, auditable preparation step for protocol-to-code tools built on free or open models. BLEU does not measure correctness, no ablation or state-of-the-art baseline was run, and clinical validation is required before any use.
Cross-Dataset Generalization of Deep Learning for Retinal Vessel Segmentation in Fundus Images: External Evaluation from DRIVE to CHASE_DB1
Automated retinal vessel segmentation can support quantitative analysis of fundus photographs and facilitate measurement of vascular characteristics associated with ocular and systemic disease. Although deep-learning methods achieve high segmentation performance on benchmark datasets, their reliability may decrease when applied to images obtained under different acquisition conditions. This study evaluated the cross-dataset generalizability of a U-Net segmentation architecture incorporating an ImageNet-pretrained ResNet34 encoder. Sixteen images from the DRIVE dataset were used for model training and four were retained for in-domain validation. Images were resized to 256 × 256 pixels and processed using a consistent normalization pipeline, while spatial and photometric augmentation was applied during training. Without target-domain retraining or domain adaptation, the trained model was subsequently evaluated on the independent CHASE_DB1 dataset. Segmentation performance was assessed using the Dice coefficient, intersection over union (IoU), and sensitivity, together with qualitative examination of predicted vessel maps. On the held-out DRIVE images, the model achieved a Dice coefficient of 0.7263, IoU of 0.5703, and sensitivity of 0.7201. On CHASE_DB1, performance decreased to 0.6269, 0.4573, and 0.5502, respectively. The disproportionately larger reduction in sensitivity indicated an increased frequency of missed vessel pixels under cross-dataset evaluation. Qualitative examination showed greater fragmentation and omission of fine and peripheral vascular structures in CHASE_DB1 images. These findings demonstrate that within-dataset performance may overestimate the robustness of retinal vessel segmentation models when transferred to a different imaging domain. External evaluation should therefore form an important component of retinal imaging artificial intelligence research. Future work should investigate higher-resolution processing, vessel-width-specific evaluation, multi-domain training, domain-adaptation strategies, and larger multicentre datasets to improve model generalizability.
Fine-Grained Human Activity Recognition in Domestic Settings Using Ultra-Wideband Radar: A Comparative Study of Deep Learning Models
Human activity recognition with ultra-wideband (UWB) radar offers contactless, privacy-preserving monitoring of daily living, but distinguishing fine-grained activities performed in the same posture remains difficult. We recorded nine daily activities from 30 participants with a single PulsON 410 radar (3.1–5.3 GHz); seven are performed seated in the same chair and differ only by subtle hand and arm motions. Each 15 s recording is a 120×1440 channel–impulse–response matrix, preprocessed by background subtraction, two-dimensional wavelet denoising and standardization. Six deep learning models (Transformer, CNN-LSTM, 1D-CNN, LSTM, GRU and MLP) and a soft-voting ensemble were compared with participant-independent splits (22 training, 3 validation and 5 test participants) and five-fold cross-validation. The Transformer reached the highest cross-validation accuracy (97.74±0.30% ). On the held-out participants, the Transformer (97.16%), 1D-CNN (96.93%) and GRU (96.45%) performed similarly, ahead of the CNN-LSTM, LSTM and MLP (87.47%), and a three-model ensemble reached 98.35%. Two additional participant splits and a retrained re-run confirmed that the MLP is clearly the weakest model, but the order of the other models changed by several points between runs; in the re-run, an ensemble of all six models did not exceed the Transformer (97.74%), and a range–Doppler baseline reached only 50.63%. Eating versus drinking tea and mobile phone use versus reading or sitting idle were the most confused pairs. Given the small test cohort and the controlled setting, the results are indicative rather than definitive.
Cost-sensitive random forest with adaptive threshold optimization for imbalanced financial fraud detection
Background: Financial transaction fraud detection is a challenging binary classification problem characterized by severe class imbalance, rare fraudulent cases, and asymmetric misclassification costs. In real-world banking systems, failing to detect fraudulent transactions can lead to substantial financial losses, whereas false alarms mainly increase operational review costs. Therefore, fraud detection models should not be optimized only for overall accuracy, as accuracy-oriented classifiers may favor the majority non-fraud class and fail to identify minority fraud cases effectively. To address this issue, this study proposes a cost-sensitive ML framework for financial transaction fraud detection under class imbalance and asymmetric financial risk.Methods: In this study, a cost-sensitive RF framework with adaptive probability threshold optimization is proposed for binary fraud classification. Misclassification costs are incorporated during model training through class-weighted learning, allowing the classifier to assign greater importance to minority fraud cases. In addition, instead of using the default probability threshold of 0.5, the decision threshold is optimized using validation data to minimize the expected misclassification cost and improve minority-class detection. The proposed framework was evaluated on two publicly available financial fraud datasets with different transaction characteristics: the Credit Card Fraud Detection dataset and the Financial Fraud Prediction dataset. Its performance was compared with conventional ML baselines using fraud-oriented evaluation metrics, including precision, recall, F1-score, false-negative rate, false-positive rate, and expected misclassification cost.Results: The proposed framework achieved strong and consistent performance across both datasets. On the Credit Card Fraud Detection dataset, it achieved a precision of 94.4%, recall of 85.7%, and F1-score of 89.8%. On the Financial Fraud Prediction dataset, it achieved a precision of 93.1%, a recall of 83.8%, and an F1-score of 88.2%. Compared with conventional fixed-threshold ML models, the adaptive threshold strategy reduced false-negative predictions while maintaining high precision. These results indicate that the proposed cost-sensitive framework provides a more balanced decision strategy for fraud detection under highly imbalanced transaction distributions.Conclusion: The findings demonstrate that combining class-weighted learning with adaptive threshold optimization can improve fraud detection performance under asymmetric misclassification costs. The proposed framework offers a practical and interpretable decision-support approach for banking fraud detection systems, particularly in settings where reducing undetected fraudulent transactions is more important than maximizing overall accuracy alone.
Machine Learning-Based Assessment of Land-Use Change, Forest Recovery, and Landscape Connectivity in Islamabad
LULCC is a major driver of habitat fragmentation, biodiversity loss, and declining landscape connectivity, particularly in rapidly urbanizing regions. Although Islamabad has experienced substantial urban expansion and forest-cover change, long-term city-scale assessments linking land-cover dynamics with landscape connectivity remain limited. This study examined spatiotemporal LULCC in Islamabad from 1991 to 2021 and assessed whether recent forest recovery improved landscape structural connectivity. Landsat images acquired in 1991, 2001, 2011, and 2021 were classified into five land-cover categories: water, forest, built-up area, bare land, and agricultural land. Classification was performed using the Random Forest (RF) algorithm in Google Earth Engine (GEE). Landscape composition and spatial configuration were quantified using FRAGSTATS 4.3, while forest fragmentation was evaluated using the Landscape Fragmentation Tool v2.0 (LFT) with a 100 m edge threshold. The classifications achieved overall accuracies above 90%, with Kappa coefficients (K) greater than 0.85. Built-up area increased from 76.31 km2, representing 7.55% of the study area, in 1991, to 258.62 km2, or 25.60%, in 2021, demonstrating rapid urban expansion and associated habitat conversion. Forest cover increased to 340.86 km2 in 2001, declined to 271.96 km2 in 2011, and subsequently recovered to 409.22 km2 in 2021. Despite this increase in forest extent, fragmentation metrics indicated persistent spatial subdivision and limited structural connectivity. High patch density (PD), reduced landscape aggregation, and changes in the largest patch index (LPI) indicated persistent spatial subdivision and limited habitat continuity. These findings highlight the value of integrating RF-based land-cover classification, multitemporal remote sensing, and landscape metrics for urban environmental monitoring. The findings suggest that future land-use planning should consider landscape connectivity, protection of existing forest patches, and spatially coordinated restoration alongside continued reforestation.