Altınbaş University Institutional Repository
Not a member yet
5805 research outputs found
Sort by
From Raw to Refined: A Data Preprocessing Pipeline for Robust Prediction
No full textPrecise air quality prediction is an essential criterion for smart cities. Prediction accuracy relies heavily on the quality of input data. However, raw environmental datasets often suffer from missing values and outliers, leading to reduced model performance. This paper presents a systematic preprocessing framework applied to an air quality dataset from Seoul, South Korea, focusing on key pollutants including SO2, NO2, O3, CO, PM10, and PM2.5. First, the missing values were identified and addressed using method chaining operations in Python, ensuring efficient and reproducible cleaning. Outlier detection was performed using multiple strategies, tailored for each pollutant according to its nature and value range. Results show significant variance reduction and improved data consistency, enabling better downstream machine learning modeling. The preprocessing approach demonstrated here can be generalized to similar urban air quality datasets, highlighting its effectiveness for environmental monitoring and prediction, calling for a more sustainable way of living that supports smart cities management
Alzheimer's Disease Prediction Using MRI-PET Fusion Through Multi-Modal Deep Learning With Late Fusion Techniques
No full textThis research presents an advanced deep learning framework for Alzheimer's Disease prediction through multimodal MRI-PET fusion, integrating modality-specific encoders and late-stage attention-guided fusion. Utilizing dual-stream 3D CNN architectures - 3D-ResNet for structural MRI and a custom PET-CNN for metabolic imaging - the model applies gated attention to maintain modality integrity and enhance semantic alignment at the decision level. Evaluation across large-scale neuroimaging cohorts including ADNI, AIBL, and OASIS-3 yielded classification accuracies up to 98.7%, with peak F1-scores of 0.894 and AUC values ranging from 0.92 to 0.96, confirming reliable stage differentiation among CN, MCI, and AD groups. Against unimodal and early fusion baselines, the proposed system achieved a 5.1-6.2% accuracy increase and reduced false positive rates by 25-28%, particularly in early MCI detection scenarios. Through domain-adaptive regularization and cohort-invariant training, the framework generalizes robustly across external datasets without retraining. The low-latency inference pipeline ( ≤ 160ms per subject) enables clinical integration, while its modular encoder architecture supports scalable extension to new imaging modalities and neurodegenerative markers with minimal reconfiguration
Integrative machine learning approaches for enhanced cardiovascular disease prediction: a comparative analysis of XGBoost and ANFIS algorithms
No full textCardiovascular diseases (CVDs) remain the leading cause of mortality worldwide, highlighting the urgent need for advanced diagnostic tools to improve early detection and patient outcomes. This study evaluates the predictive performance of two machine learning models-Extreme Gradient Boosting (XGBoost) and the Adaptive Neuro-Fuzzy Inference System (ANFIS)-across five datasets from the UCI Machine Learning Repository: Cleveland, Hungary, Switzerland, Long Beach VA, and Statlog Heart. Comprehensive preprocessing steps-including imputation, standardization, one-hot encoding, and SMOTEENN-were applied to ensure data consistency and address class imbalance. XGBoost achieved perfect accuracy (100%) on the Switzerland and Statlog datasets, reflecting its strength in structured data environments and consistent predictive performance. Conversely, ANFIS outperformed XGBoost on the Cleveland dataset, demonstrating its effectiveness in modeling complex, nonlinear relationships. Performance evaluation metrics included accuracy, precision, recall, F1 score, F2 score, and ROC-AUC. XGBoost consistently delivered high precision and recall, which are essential for minimizing false positives and negatives in clinical settings. ANFIS yielded high F2 scores, indicating a stronger emphasis on reducing false negatives-a critical concern in CVD diagnosis. This comparative analysis suggests that while XGBoost is well suited for scalable, high-throughput diagnostic applications, ANFIS offers greater interpretability and is more effective in nuanced clinical scenarios. These findings underscore the potential of integrating advanced machine learning models into cardiovascular disease prediction frameworks to enhance diagnostic accuracy and support real-world healthcare decision-making
Machine learning techniques for differentiating psychrophilic and non-psychrophilic bacterial α/β hydrolase enzymes
No full textPsychrophilic enzymes represent a category of macromolecules that have acquired specific properties that enable these enzymes to perform their catalytic activity at low temperatures with high efficiency. One of the factors contributing to their adaptation is increased active site flexibility. Psychrophilic enzymes are of significant industrial interest due to their applications in food production, environmental remediation, pharmaceuticals, textiles, and detergents. Despite growing interest, the molecular mechanisms underlying the adaptation of psychrophilic enzymes to low temperatures remain largely unexplored. This study aims to investigate the differences between psychrophilic and non-psychrophilic bacterial α/β hydrolase enzymes. 464 psychrophilic and 562 non-psychrophilic α/β hydrolase enzymes were retrieved from the UniProt database. Further classification of these enzymes based on amino acid composition was performed using a set of machine learning algorithms such as Decision Tree (DT), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM). Ten variables, including the contents of Ala, Gly, Ser, Thr, charged, aliphatic, aromatic, and hydrophobic amino acids, as well as the aliphatic index and the grand average of hydropathy (GRAVY), were analyzed. The Random Forest algorithm achieved the highest classification rate with an accuracy of 77%. Further analyses showed that the amino acid threonine and serine played the most important role in determining psychrophilic traits. This suggests that these amino acids play a significant role in enhancing the enzyme's hydrogen-bonding capacity, thereby contributing to its structural flexibility and stability under cold conditions. This study confirms that some amino acids, especially serine and threonine, are generally involved in the cold adaptation of psychrophilic α/β hydrolase enzymes and may provide an interesting platform from a biotechnological point of view
Speech Recognition of High Impact Model Using Deep Learning Technique: A Review
No full textConference name : 7th International Congress on Human-Computer Interaction, Optimization and Robotic Applications, ICHORA 2025
Conference city : Ankara
Conference date : 23 May 2025 - 24 May 2025
Conference code : 209351Machine learning has been the subject of enormous study in speech processing, particularly in speech recognition, for the last decades. On the other hand, deep learning's potential use in speech recognition has been the subject of much study in recent years. New evidence suggests that deep learning has far-reaching applications across many domains and has significantly contributed to AI. Several applications involving voice have demonstrated encouraging outcomes when using deep learning models. There has been a recent growth of attention-based approaches and models that apply transfer learning to enormous datasets, which offers added motivation for ASR. Focussing on several deep-learning models, it provides a summary and comparison of the state-of-the-art approaches used in this field of study. Additionally, we have evaluated the models on speech datasets to learn how they function on various datasets for practical application. Academics interested in open-source ASR could use this study as a jumping-off point for future research on issues like minimizing data dependence, increasing generalisability across languages with limited resources, speaker variability, noise conditions, and identifying and resolving obstacles to advancing existing research
The effect of unorthodox monetary policy on inflation in Türkiye: evidence from the synthetic control Method
No full textThis note examines the effect of unorthodox monetary policy on inflation in Türkiye after mid-2018. Using the Synthetic Control Method, we construct a synthetic counterpart to estimate the country’s inflation trajectory had it pursued an orthodox policy framework. We find that inflation could have remained at around 20% under a traditional monetary policy instead of rising until it peaked at 85% in 2023. The evidence reveals that repeated liquidity injections and interest rate cuts–despite rising inflation–led to severe price instability. Our findings highlight the economic risks associated with deviating from orthodox policy frameworks and emphasize the importance of stable, rule-based policies in maintaining price stability
A Novel Pancreatic Tumor Detection and Diagnosis Using Adaptive TransResUnet Aided Segmentation and ASPP with Multi-Scale EfficientNet-Based Classification
No full textCODEN : CYSYDA deadly disease with poor prognosis procedure available at present is the pancreatic tumor. Efficient detection is done using a Computer-Aided Diagnosis (CAD) system. The early detection of pancreatic tumors can enhance the survival rate. However, no sufficient works are dedicated to detect pancreatic tumors at its beginning stages. Hence, an advanced deep learning-oriented segmentation process to assist in the detection of pancreatic tumor is developed in this work. The necessary CT and MRI images are gathered from the utilization of IoT-based devices. Once the input image is gathered, the segmentation is carried out. An Adaptive TransResUnet (ATResUNet) is utilized for the segmentation procedure. The variables in the ATResUNet are tuned with the help of Improved African Vultures Optimization Algorithm (IAVOA). The segmented image is further considered to crop the Region of Interest (ROI). The cropped ROI is finally given as input to the suggested Atrous Spatial Pyramid Pooling-based Multi-scale EfficientNet with Attention Mechanism (ASPP-MENetAM) model. The detection of the pancreatic tumor is carried out using the ASPP-MENetAM framework. The detection outcome from the implemented ASPP-MENetAM is then compared with the results from other conventional pancreatic tumor detection models to assess the efficacy of the implemented detection system
Advancements in Concentrated Solar Power: A Review of Heat Transfer and Parabolic Trough Technologies
No full textThis review examines the advancements in concentrated solar power (CSP) technologies, focusing on their potential to meet energy demands sustainably while mitigating global warming. CSP systems, which harness solar energy through mirrors and lenses to generate high-temperature heat, offer a reliable alternative to fossil fuels. Key CSP technologies include Linear Fresnel Reflectors, Parabolic Dishes, Solar Towers, and Parabolic Trough Collectors (PTCs). Each system is assessed for its design, efficiency, and suitability in regions with high direct solar irradiance. Among these, PTCs are highlighted for their cost-effectiveness and thermal efficiency, with applications reaching temperatures up to 550°C, making them suitable for both small and large-scale implementations. The review also explores various heat transfer enhancement techniques, categorized into active, passive, and compound methods. Passive techniques, such as inserts, surface modifications, and nanofluids, are examined for their ability to increase heat transfer efficiency without external energy sources. Active methods like pumps, fans, and compound approaches are discussed for maximizing thermal performance. Advances in receiver design, including twisted tapes, wire coils, fins, and porous materials, are evaluated for their impact on heat transfer rates, thermal losses, and overall system efficiency. Additionally, the potential of nanofluids to improve thermal conductivity is explored. This comprehensive review underscores the importance of optimizing CSP systems to maximize efficiency, offering insights into innovations that could further enhance the adoption of solar thermal energy worldwide
Postmortem Inductively Coupled Plasma Mass Spectrometry Analysis Reveals Elevated Heavy Metal Concentrations in Coronary Arteries: A Comparative Autopsy Study Supporting a Toxic Inflammatory Hypothesis for Atherosclerosis
Get PDFIntroduction: A large number of studies have been carried out for the etiology of atherosclerosis and many risk factors have been identified, including environmental factors and heavy metals, which are related to the pathogenesis. This study aimed to determine the effects of heavy metals, which have activation and inhibition effects on various metabolic pathways, on atherosclerosis by examining coronary arteries obtained from autopsy series.
Methods: Coronary arteries of 28 autopsy cases were analyzed by inductively coupled plasma mass spectrometry method. Sixteen of the cases had coronary atherosclerotic plaques and 12 of the coronaries were normal. Twenty trace metal concentrations were examined from the samples obtained.
Results: Twenty-eight coronary artery samples (16 with atherosclerosis, 12 normal) were analyzed using ICP-MS. Levels of Mg, K, Ca, P, Fe, Zn, Al, S, As, Pt, Sb, Hg were significantly higher in atherosclerotic arteries (e.g., Ca: 51,384 vs. 1,723 ppm, p = 0.005; P: 30,791 vs. 3,443 ppm, p = 0.003; Hg: 3.2 vs. 0 ppm, p < 0.001). Elements such as lead, cobalt, and cadmium remained below detection limits in both groups.
Conclusion: Heavy metals through inflammation, oxidative stress, and disrupted antioxidant pathways are independent risk factors that increase the risk of atherosclerosis. These findings provide tissue-level evidence that heavy metal accumulation may contribute to atherosclerosis through oxidative stress, inflammation, and disruption of antioxidant defenses
Physicochemical attributes of irradiated proteins
No full textIn this book chapter, the effects of irradiation on the physicochemical properties of animal- and plant-based proteins were investigated. Studies in the literature showed that different methods such as γ-irradiation, electron beam irradiation, X-ray irradiation, and UV irradiation could be used to irradiate food proteins. Animal proteins studied were egg products such as liquid egg or dried egg or egg-based proteins such as globulin, ovalbumin, and lysozyme; meat-based proteins such as gelatin or myofibrillar proteins; and milk-based proteins such as casein, α-lactoalbumin, β-lactoglobulin. Irradiated plant-based proteins were mostly rice protein, watermelon seed kernel proteins, wheat germ protein hydrolysates, etc. The physicochemical properties affected by irradiation were solubility, foaming capacity, foaming stability, emulsifying activity, emulsion stability, gelling capacity, film-forming properties, viscosity, water holding capacity, as well as hydrophobicity. These properties were highly interrelated and affected by changes in protein structure, such as protein degradation, protein cross-linking, or amino acid side chain modifications, particularly carbonyl formation and oxidation of S-containing amino acids as a result of irradiation