YFYelda Fırat

Yelda Fırat

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Yelda Fırat has a total of 15 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Uncertainty and Geometric Image-Region Ranking in Artificial Intelligence-Based Panoramic Tooth Segmentation: A Multi-Source External Benchmark

    by Yelda Fırat

    Background/Objectives: This retrospective benchmark evaluated scores for selecting regions of panoramic tooth segmentations for review. Methods: Three-member U-Net ensembles trained on UFBA-425 were evaluated on 590 held-out images and subsequently on 329 photographed radiographs. The original mixed-granularity pipeline comparison used the calibrated residual ensemble on 100 pediatric images. Cells in a 2 × 4 grid received regional mutual information (MI) scores but shared one whole-image soft Dice confidence (SDC) value. A lower Dice-risk area under the risk–coverage curve (AURC) indicated better ranking against source masks. Interaction tests, quality subset analyses, and prediction-positive sensitivity analyses were post hoc. Image-key bootstrap intervals may be too narrow for patient-level inference because patient independence was unverified. Results: The primary MI-minus-SDC AURC difference was +0.007 (95% confidence interval −0.033 to +0.041), leaving the comparison inconclusive. Subsequent evaluations were exploratory. Analyses of reference-positive cells (containing reference tooth pixels) received co-principal emphasis after outcomes were known. MI had a higher Dice-risk AURC than the replicated SDC in the pediatric main subset (+0.238) and the local SDC in AKUDENTAL (+0.157 residual, +0.134 ResNet18). When 20% of grid cells were reviewed, the mean maximum pixel probability (MSP) captured more error pixels than MI in seven of ten source/model comparisons, including both AKUDENTAL ensembles. On photographed radiographs, the MSP led by 15.73 percentage points for the residual ensemble and MI by 3.66 for ResNet18. For ResNet18, equal weighting of false-positive and false-negative capture favored MSP by 6.59 points. Conclusions: The primary comparison established neither superiority nor equivalence. Exploratory capture varied by source, model, and error mix.

  • Open Access

    Transferability of Transcriptome-Based Gene-Dependency Models from Two-Dimensional Cancer Models to Patient-Derived Organoids: Blinded External Validation

    by Yelda Fırat, Yılmaz Kılıçaslan

    Transferability of transcriptome-based gene-dependency models from two-dimensional cancer models to organoids remains uncertain. Expression-only Elastic Net models for 20 preselected genes were selected using patient-grouped nested cross-validation in 681 models and evaluated in a locked 171-model holdout. Historical target selection used all 852 models, making holdout performance conditional internal confirmation rather than an unbiased generalization estimate. After refitting on all 852 models, frozen models underwent exploratory evaluation in 94 Broad organoids and blinded validation in 148 Sanger organoids from 144 donors. Predictions were sealed before Sanger outcome access. Primary target-wise Spearman correlations, ranking organoids per gene, averaged 0.4094, 0.1518, and 0.1015 in holdout, Broad, and Sanger, respectively. Sanger association exceeded the within-lineage permutation null (p = 0.00040), but improvement over lineage-only prediction remained uncertain (difference, 0.0462; 95% confidence interval, −0.0065–0.0986). CDK6, MDM4, and CCNE1 passed false discovery rate correction in Sanger. Retrospective within-organoid gene ranking modestly improved over lineage-only prediction in both cohorts; top-three overlap improved only in Sanger. Pancreatic results remained unvalidated projections. Findings support weak cross-domain rank transfer and limited within-organoid ranking within the selected panel, without establishing genome-wide transferability, causal explanations for attenuation, or prospective experimental or therapeutic utility.

  • Artificial Intelligence–Driven Life Cycle Assessment of Injectable Aesthetic Products: A Multi-Regional and Temporal Sustainability Analysis

    by Yelda Firat, Hüseyin Ali Sarikaya

    The environmental implications of injectable aesthetic products have not been thoroughly investigated, despite the rapid expansion of the aesthetic medicine sector. This study developed an innovative Artificial Intelligence (AI)-supported life cycle assessment (AI-LCA) framework, aligned with ISO 14040/14044 standards, to evaluate the carbon footprint of common products such as botulinum neurotoxins, hyaluronic acid fillers, and biostimulators. A comprehensive dataset was created covering four regions (European Union, United States, Asia-Pacific, and Türkiye) and five years from 2020 to 2024. The dataset was expanded from 300 to 9,960 records using Monte Carlo simulation, integration of regional environmental factors, and temporal expansion techniques. The model, trained using the Random Forest algorithm, demonstrated satisfactory predictive performance (R² = 0.954; root mean square error [RMSE] = 0.132; mean absolute error [MAE] = 0.075) and was compared with algorithms such as XGBoost and linear regression. Model stability was confirmed through 5-fold cross-validation, while generalizability was assessed using Leave-One-Region-Out Cross-Validation (mean R² = 0.769) and Temporal Hold-Out Validation (mean R² = 0.992). Bootstrap-derived 95% confidence intervals indicated narrow prediction uncertainty (relative uncertainty: 3.7–4.6%). SHapley Additive exPlanations (SHAP) analysis identified energy consumption and regional carbon intensity as the dominant predictors. Product type and cold chain requirements were found to be the main factors determining environmental impact; botulinum neurotoxins, which require refrigeration, exhibited an approximately 7.8-fold higher carbon footprint than non-refrigerated products. A reduction of approximately 8.5% in emissions was observed over five years, attributed to increased use of renewable energy, technological advances, and efficiency improvements in production processes. The AI-LCA framework provides a promising contribution to promoting sustainable aesthetic practices and increasing environmental awareness.

  • Exploratory analysis of pupillography-based machine learning for assessing autonomic dysfunction in multiple sclerosis

    by Neslihan Parmak Yener, Yelda Fırat, Meral Seferoğlu, Ahmet Metin Kargın, Yılmaz Kılıçaslan

  • Open Access

    Explainable machine learning-based identification of transcriptomic biomarkers in CD1c+ dendritic cells for non-infectious uveitis: an integrative analysis of bulk RNA-seq data

    by Yelda Fırat

    Introduction: Non-infectious uveitis (NIU) is a leading cause of intraocular inflammation, and the underlying immunological mechanisms remain incompletely understood. This study aims to classify NIU at the molecular level and interpret its biological mechanisms by applying machine learning and explainable artificial intelligence (XAI) methods to transcriptomic data from CD1c+ conventional dendritic cells type 2 (cDC2) isolated from the peripheral blood of NIU patients. Methods: Two independent cohorts (GSE194060 and GSE195501; n = 78) from the Gene Expression Omnibus (GEO) database were integrated using a strict, leakage-free pipeline. Batch correction and analysis of variance (ANOVA)-based feature selection, applied exclusively on the training set, identified the 20 most informative genes from 8,815 candidates. L2-regularized logistic regression, support vector machines, and random forest models were compared. The model’s biological basis was elucidated using Shapley Additive Explanations (SHAP) analysis and comprehensive pathway enrichment analyses. Results: The L2-regularized logistic regression model achieved the best performance with a nested cross-validation area under the curve (AUC) of 0.863, test AUC of 0.855, accuracy of 81.2%, recall of 90.9%, and F1-score of 0.870. Leave-One-Dataset-Out (LODO) validation supported cross-platform generalizability (mean AUC: 0.82). Permutation testing confirmed the model’s statistical significance (p = 0.005). SHAP analysis identified CD180 and TLR7 as the most important biomarkers. Pathway enrichment analysis revealed significant enrichment in chemokine receptor activity, MYD88-mediated signaling, Toll-like receptor cascades, and G protein-coupled receptor (GPCR) signaling pathways. Conclusion: These findings demonstrate the potential of XAI to elucidate disease mechanisms from transcriptomic data and present an interpretable 20-gene signature as a candidate diagnostic biomarker panel for non-infectious uveitis that warrants further clinical validation.


  • Open Access

    An explainable ensemble machine learning framework for predicting cartilage biomechanical degradation from transcriptomic profiles: toward smart orthopedic device design

    by Yelda Fırat

    Introduction: Osteoarthritis (OA) is a common joint disease characterized by progressive biomechanical deterioration of cartilage tissue. This study aims to develop an explainable ensemble machine learning model capable of predicting cartilage biomechanical deterioration from transcriptomic profiles and to present this model as a digital biomarker for intelligent orthopedic devices. Methods: Seven independent datasets (n = 203) obtained from the Gene Expression Omnibus (GEO) database were harmonized using the Joint ComBat algorithm. Differentially expressed gene (DEG) analysis performed on the training cohort (n = 122) identified 93 significant genes. An ensemble model combining Random Forest (RF), XGBoost, and Support Vector Machine (SVM) algorithms was trained using these genes. The disease probability generated by the model was redefined as a continuous Cartilage Degradation Index (CDI). The model’s performance and generalizability were tested on four independent validation cohorts including patients undergoing arthroscopic partial meniscectomy (APM, n = 44), chondrocytes subjected to in vitro mechanical stress (n = 18), and different geographic cohorts (n = 19). The biological basis of the model’s decisions was elucidated using SHAP (SHapley Additive exPlanations) analysis and comprehensive pathway enrichment analyses. Results: The ensemble model demonstrated a 5-fold cross-validation performance with an area under the curve (AUC) of 0.960. Permutation testing (p < 0.001) and bootstrap 95% confidence interval (CI) (AUC: 0.900–0.996) confirmed the model’s robustness. CDI modeled degradation across a biologically consistent spectrum: 0.03–0.09 in healthy cartilage, 0.30–0.42 in early-stage APM, and 0.79 in advanced-stage OA. CDI scores were significantly higher in cells exposed to abnormal mechanical stress (overstress) compared to the control group (0.60 vs. 0.52, AUC = 0.778). SHAP analysis highlighted COL15A1, CXCL14, CISH, FAP, and PIM2 as the most critical mechanosensitive genes. Pathway analysis confirmed that the model is based on mechanotransduction mechanisms such as focal adhesion, extracellular matrix (ECM) organization, and epithelial–mesenchymal transition. Conclusion: This study presents a transcriptomic CDI capable of measuring cartilage biomechanical degradation at the molecular level. The identified mechanosensitive gene profile holds strong potential as a digital biomarker for next-generation smart orthopedic implants and wearable biomechanical sensors.


  • Open Access

    Deep Learning-based Detection of Motor Biomarkers for Autism from Children's Video Recordings

    by Yelda Fırat, Yılmaz Kılıçaslan, Hüseyin Ali Sarıkaya, Murat Kaan Yılmaz

    Autism Spectrum Disorder is a neurodevelopmental disorder with onset in early childhood and its diagnosis often requires clinical processes based on long, subjective observations. Although early diagnosis and intervention can significantly improve developmental outcomes, existing methods are limited in terms of scalability and objectivity. The aim of this study is to develop a hybrid deep learning model that detects Autism Spectrum Disorder with high accuracy by analyzing motor behaviors from videos of children recorded in their natural home environment. In this study, joint coordinates were extracted using the MediaPipe Pose model and spatial, temporal, frequency and coordination-based features were calculated from these data. The features were processed with a hybrid architecture integrating CNN, BiLSTM and attention mechanism. CNN captured spatial patterns, BiLSTM learned the dynamics over time, and the attention mechanism focused on critical movement segments. The model achieves over 97% accuracy on closed datasets and over 83% on public videos such as YouTube and TikTok. These results show that the method performs robustly under both controlled and real-world conditions. The study provides a scalable, objective and clinically applicable screening tool that overcomes the problems of artificial environments and limited data.

  • Open Access

    SHAP-based explainable AI framework for autism severity classification using 3D motor biomarkers

    by Yelda Fırat

    Introduction: Early Autism spectrum disorder (ASD) diagnosis is critical for intervention, yet current methods rely on subjective clinical observations. This study develops objective tools to classify ASD severity using 3D motor movement analysis, addressing motor abnormalities as core diagnostic features. Methods: A Random Forest (RF) model classified three severity levels using 463 motor features from 25 Kinect V2 joint points. Data from 109 children (50 typical, 50 moderate ASD, 9 severe ASD) were validated via 5-fold cross-validation and two held-out test sets (20% each). Shapley Additive Explanations (SHAP) analysis identified critical motor biomarkers. Results: The model achieved 84.6±10.9% accuracy (5-fold cross-validation) and 86.4% accuracy (internal and held-out test sets). For severe ASD, the model achieved 100% classification accuracy on synthetic test data (4/4 cases; 95% CI: 39.8%-100.0%). However, this result represents a methodological proof-of-concept rather than clinical validation, as severe ASD features were synthetically generated from moderate ASD data and the model has not been validated on real Kinect-derived severe ASD motor data. SHAP analysis identified wrist movements, knee trajectories, and elbow-to-foot distances as key motor biomarkers for severity classification. Discussion: This Kinect-based approach with RF and SHAP offers effective, interpretable ASD severity assessment for typical and moderate ASD classes, with promising methodological foundations for severe ASD pending validation on real data.


  • Open Access

    An explainable ensemble machine learning model using baseline blood transcriptomics to predict Parkinson's disease motor progression

    by Yelda Fırat

    Introduction: Predicting Parkinson's disease (PD) motor progression remains challenging despite advances in neuroimaging. Blood-based transcriptomic profiling offers a more accessible and cost-effective alternative. This study aimed to develop and validate a machine learning approach using blood-based transcriptomic data to predict 12-month motor severity in PD and to identify the transcriptomic features and biological pathways most strongly associated with progression. Methods: A Stacking Regressor ensemble model combining three gradient boosting algorithms (XGBoost, LightGBM, CatBoost) was developed using baseline Parkinson's Progression Markers Initiative (PPMI) data ( n = 390), integrating blood RNA sequencing (RNA-seq) and clinical features to predict 12-month UPDRS Part III scores. SHapley Additive exPlanations (SHAP) analysis was applied to identify key prognostic features, evaluating seven PD risk genes (SNCA, LRRK2, GBA, PRKN, PINK1, PARK7, VPS35) and pathway scores for mitochondrial dysfunction, neuroinflammation, and autophagy. Results: On an independent test set ( n = 78), the model achieved a Coefficient of Determination (R²) of 0.551 and Mean Absolute Error (MAE) of 6.01. SHAP analysis identified the baseline UPDRS × PINK1 interaction (UPDRS_BL × PINK1) as the most influential feature (mean |SHAP| = 0.283). Among transcriptomic features, VPS35 (mean |SHAP| = 0.010), GBA, and LRRK2 were most prominent. Mitochondrial dysfunction showed the highest pathway contribution (mean |SHAP| = 0.008). Discussion: The study establishes that machine learning integrating blood transcriptomics and clinical data effectively predicts motor progression in PD. Crucially, the interplay between initial clinical state and specific genetic backgrounds-particularly PINK1-is a more powerful prognostic indicator than any factor alone. This study provides systematic evidence that mitochondrial dysfunction is a dominant prognostic signal for disease progression, nominating key genes and pathways for future mechanistic and therapeutic investigation.


  • Open Access

    A Hybrid Deep Learning Model for Wheat Price Prediction: LSTM–Autoencoder Ensemble Approach with SHAP-Based Interpretability

    by Yelda Fırat, Hüseyin Ali Sarıkaya

    Accurate prediction of wheat prices is crucial for market participants and policymakers because volatility in agricultural markets affects food security and economic planning. This study proposes a hybrid deep-learning-based framework for daily wheat price prediction in Türkiye. The approach first applies an autoencoder to detect and remove anomalous price–quality records from a dataset of 38,019 market transactions collected between June 2022 and May 2023. A weighted ensemble combining Linear Regression, Random Forest, Support Vector Regression and an attention-based Long Short-Term Memory network is then trained on quality parameters and market attributes, with data split into training, validation and test sets. On the independent test set the ensemble achieved a coefficient of determination R2 = 0.9942 and a mean absolute error of 0.1646 TL, outperforming the constituent models. SHAP analysis identifies the price–quality ratio as the most influential feature, while the ablation analysis shows that some of the high accuracy derives from price-derived variables’ strong correlation with the target. Cross-validation confirms robustness and generalization. Overall, the framework provides an effective and interpretable tool for wheat price forecasting, though the short data collection period and single-product focus limit generalizability.