Technological Institute of the Philippines

The proposed work presents a hybrid deep learning model for predicting water quality in Taal Lake by incorporating physicochemical factors meteorological and volcanic activity information. A multivariate time series dataset was created from 2013 to 2025 by collecting data from BFAR, PAGASA, and PHIVOLCS through data harmonization, gap filling, alignment, and normalization. Five different models were compared under equal conditions: CNN, LSTM, Hybrid CNN-LSTM, CNN-LSTM with single attention, and the proposed CNN-LSTM with dual attention. The proposed model performed the best with an MAE of 0.91, RMSE of $1.32, R^{2}$ of 0.926, and accuracy of 94.1 % based on NRMSE, which is 35.3 % better than the baseline models. The dual attention mechanism enhances interpretability by pointing to important factors such as temperature variations, rainfall-induced nutrient loading, and volcanic activity. A Streamlit-based application was built to enable real-time visualization. The proposed model is accurate, interpretable, and applicable to environmental monitoring tasks.

This paper aims to propose a hybrid deep learning framework that combines the architecture of convolutional and transformer-based to enhance the segmentation performance in 3D medical imaging. This study use the 3D U-Net is utilized for local feature extraction and the Swin UNETR to captures global contextual information through transformer-based attention mechanisms. A hybrid model is then developed using a learnable weighted fusion of the pretrained 3D U-Net and Swin UNETR outputs. Moreover, during training, a 3D PatchGAN discriminator is incorporated to perform adversarial learning, thereby enhancing the realism of segmentation and the accuracy of boundaries. The performance was evaluated using the Dice score, Intersection over Union (IoU), mean Intersection over Union (mIoU), Precision, and Recall. The results showed that the standalone 3D U-Net obtained a Dice score of 0.3654, IoU of 0.5048, mIoU of 0.2745, Precision of 0.4841, and Recall of 0.4252, and Swin UNETR achieved a Dice score of 0.3440, IoU of 0.4993, mIoU of 0.2622, Precision of 0.3006, and Recall of 0.4019. In the proposed hybrid model, the result is a Dice score of 0.3519, IoU of 0.5064, mIoU of 0.2687, Precision of 0.5861, and Recall of 0.3845. In contrast, the Swin UNETR showed comparatively lower performance across most metrics. These findings represent that while the 3D U-Net provides the most balanced performance, the hybrid model offers advantages in precision-oriented tasks. The choice of model therefore depends on the specific application requirements, particularly the trade-off between precision and recall.

Course completion is an important element of student success in higher education, but identifying at-risk students who will complete a course can be an ongoing challenge. This paper describes an explanation-based real-time decision-support system that uses students' test performance and engagement data to predict course completion outcomes and target interventions to academic success. The proposed system uses an XGBoost (Extreme Gradient Boosting) model as a basis for modeling complex, non-linear patterns in course behavior across subjects and has very good predictive power, with an accuracy of 0.9450 and an ROC-AUC of 0.9913. Although the model has strong predictive power for student course completion, the lower recall scores among noncompleters indicate that greater detection of at-risk students is needed. In order to promote transparency and usability, SHAP (Shapley Additive Explanations) is used to provide global and local interpretability, with results presented through role-based dashboards for students, faculty, and administrators. It's based on the project approach to an early-warning system; however, certain powerful features are made available at the end of the semester, and it serves as a progressive risk-monitoring system. It employs NSTP (National Service Training Program) data, and future research may expand the system across several courses to include early-semester and engagement features and to assess the program's lasting effects on student success and retention.

The Philippines' vulnerability to flooding necessitates reliable, localized early-warning systems (EWS). This study presents the empirical validation of a low-cost, bridge-mounted, solar-powered EWS prototype featuring integrated audible (siren) and digital (SMS) alerts. Building upon a prior feasibility study, this paper provides quantifiable performance measures and numerical evidence using nondestructive testing and internal validation protocols. Key performance indicators measured were sensor consistency, alert latency, power autonomy, and environmental robustness. Results from more than 50 trials showed high sensor consistency (activation SD greater than 0.15 cm). A debris simulation quantified the float sensor's vulnerability, yielding a 3 percent False Negative Rate and 5 percent False Positive Rate. The alert system's responsiveness was adequate, with a 95th percentile (P95) end-to-end SMS latency of 39.125 seconds, which is well within the 60 -second benchmark. A simulated blackout test confirmed 86.5 hours of continuous operational autonomy, exceeding the 72 -hour target. The prototype's physical design passed internal simulation protocols for IPX5 (water) and IPX6 (dust) ingress, a 20 kg static load test, and electrical integrity checks, with the audible alarm exceeding ISO 7731 guidelines. The study concludes that the developed prototype is a functionally reliable, responsive, and resilient EWS under the tested conditions. The methodology provides a practical framework for validating single prototypes, thereby bridging the gap between conceptual design and empirically assessed technology. Recommendations include long-term field validation and investigation of non-contact sensors.

Accurate moisture content prediction during substrate preparation is critical for optimizing yields and minimizing contamination in mushroom spawn cultivation. To address this challenge, this study presents a direct-wired, multi-layer embedded sensing system coupled with a Bidirectional Long Short-Term Memory (Bi-LSTM) neural network, enabling the prediction of moisture content at three distinct substrate depths (bottom, middle, top) within rice-grain medium. The hardware architecture utilizes a Raspberry Pi Pico for high-frequency (3.4 s) USB-based data acquisition. Building on this, the Bi-LSTM leverages two stacked layers (64 and 32 units) to process sequences bidirectionally, thereby capturing complex, non-monotonic temporal dependencies. Importantly, empirical data from six independent trials comprising 2,044 valid records revealed a consistent vertical moisture gradient driven by capillary action, with the bottom layer exhibiting maximum retention. To operationalize these readings for decision-making, a correlation-optimized Substrate Readiness Index (SRI) (SRI= 0.40 x MCB + 0.35 x MCM + 0.25 x MCT) was formulated, effectively differentiating acceptable batches by a margin of 5.81 percentage points (Wilcoxon p < 0.001, σ = 1.17%). Model generalization was rigorously evaluated using Leave-One-Session-Out Cross-Validation (LOSO-CV) across eight algorithms. Notably, while traditional models failed to generalize (yielding negative R2), the proposed Bi-LSTM significantly outperformed all baselines, achieving an RMSE = 0.480±0.507%, R2 = 0.972±0.036, and Within-±2% Accuracy (W2A) = 98.1±2.8%. Furthermore, statistical superiority over traditional algorithms and Transformer-Encoders was confirmed via Friedman's (χ2 = 42.67, p < 0.001) and Holm-Bonferroni-corrected pairwise tests. These findings demonstrate predictive accuracy equivalent to commercial sensors, but with enhanced multi-layer spatial resolution.