A Novel Hybrid Approach Using MRMR-based Feature Selection and Bayesian Optimized Random Forest Classification for Accurate Fabric Defect Detection

Chromatography is a key unit operation in the biopharmaceutical manufacturing process used for protein purification and polishing. Design and optimization of these processes are resource-intensive resulting from the complex combinatorial design space. This constraint combined with the wide diversity in therapeutic formats and increased pressure for timely delivery to the market necessitates an efficient, fast, robust and generalized framework for process design and optimization. Here we present Gaussian processes as a potent machine learning methodology for predictive modeling in the context of QSAR modeling used for resin and solvent condition selection. We highlight the on-par predictive power of Gaussian Processes with other reported machine learning algorithms. Furthermore, we demonstrate additional properties of Gaussian processes such as its ability to provide confidence estimates for its prediction that makes it suitable for model-assisted optimization. Finally, we demonstrate the possibility to derive feature importances from Gaussian processes, making these models as interpretable as ensembled tree-based methods such as random forests.

Understanding emotions in text is an important part of sentiment analysis, especially in areas like mental health monitoring, customer feedback analysis, hate speech and disaster management. Unlike basic sentiment analysis that classifies text as positive or negative, real human emotions are complex and often overlapping, requiring multi-label classification to accurately capture multiple emotional states within a single input. While transformer-based models have improved performance, challenges persist particularly in low-resource languages and culturally diverse contexts due to the scarcity of annotated data and the difficulty in generalizing in different languages. This study proposes Hyb-Stack, a hybrid stacked ensemble framework that integrates simple stacking and cross-validation stacking, combining predictions from four transformer-base models (BERT, DistilBERT, RoBERTa, and mBERT) using a Random Forest meta-classifier to enhance classification accuracy, adaptability, and cross-lingual generalisation. Hyb-Stack was evaluated on three datasets: a high-resource English corpus (SemEval-2018 Task 1 E-c) and two low-resource corpora, the Bahasa Indonesia hate speech and HaEmoC_V1, a newly constructed Hausa-language emotion corpus developed to address the lack of annotated data for this language. Experimental results demonstrate that mBERT outperforms individual base models, achieving F1-scores of 82.28 (HaEmoC_V1), 85.33 (Bahasa Indonesia), and 88.90 (SemEval-2018 English). Also, the EM-9 ensemble set (BERT + DistilBERT + mBERT) improves performance, yielding F1-scores of 89.48, 88.19, and 90.67 on the respective datasets, which surpasses both individual models and conventional ensemble strategies such as averaging and weighted averaging. These findings highlight the effectiveness of combining multiple transformers with an optimized decision layer to advance multi-label emotion classification on diverse linguistic contexts.

ABSTRACT This study presents a machine learning-based framework for predicting driver injury severity (DIS) in urban traffic crashes, using ten years (2011–2020) of police-reported data from Imphal, India. Focusing on morning and nighttime accidents, six ML models were trained and evaluated based on accuracy, precision – recall analysis, including AUC-PR (area under the Precision – Recall curve), recall (sensitivity) for the fatal class, and precision for the fatal class, under multiple train/test splits and cross-validation schemes. The best-performing models—Random Forest for morning crashes and LightGBM for night—were identified and further analyzed using SHAP-based sensitivity methods to determine key predictors. Results revealed that model performance and variable impact were sensitive to training ratio and cross-validation strategy. Two-wheeler involvement and narrow-road conditions were among the most significant factors. Overall, the study contributes to data-driven traffic injury modeling and highlights the potential of explainable artificial intelligence (XAI) as a decision-support tool for improving urban transport safety management.

Internet of Things (IoT) botnets are networks of infected smart devices controlled by attackers and posing a serious cybersecurity challenge. Developing detection approaches that maintain high accuracy while protecting privacy presents considerable challenges, particularly in large and heterogeneous IoT networks. This paper empirically compares three modelling approaches on Bot-IoT and N-BaIoT in binary and multiclass settings: handcrafted machine learning with random forest (RF), centralised deep learning (CDL) with DNN/LSTM/BiLSTM, and federated deep learning (FDL) with the same architectures. Model hyperparameters are selected via randomised search on stratified subsets and then fixed for final training. Results show near-perfect performance for all approaches in binary detection: on Bot-IoT, CDL-DNN attains perfect accuracy, and RF is virtually perfect (only four benign-to-attack false positives), while FDL models are similarly strong with only small false-positive and false-negative counts. On N-BaIoT, RF and CDL (especially LSTM) are near-perfect, and FDL is very close to CDL. For multiclass detection, CDL-DNN leads on Bot-IoT, RF remains near perfect with minimal cross-class confusion, and FDL trails slightly; on N-BaIoT, FDL-BiLSTM and RF are essentially perfect, with CDL-LSTM close behind. Overall, the findings validate RF as a competitive classical approach, show where centralised representation learning adds value, and demonstrate that federated training preserves most of the centralised accuracy while avoiding raw data centralization (data locality) for scalable deployment.

The proliferation of data-intensive IoT applications has created unprecedented demand for wireless spectrum, necessitating more efficient bandwidth management. Spectrum sensing allows unlicensed secondary users to dynamically access idle channels assigned to primary users. However, traditional sensing techniques are hindered by their sensitivity to noise and reliance on prior knowledge of primary user signals. This limitation has propelled research into machine learning (ML) and deep learning (DL) solutions, which operate without such constraints. This study presents a comprehensive performance assessment of prominent ML models: random forest (RF), K-nearest neighbor (KNN), and support vector machine (SVM) against DL architectures, namely a convolutional neural network (CNN) and an Autoencoder. Evaluated using a robust suite of metrics (probability of detection, false alarm, missed detection, accuracy, and F1-score), the results reveal the clear and consistent superiority of RF. Notably, RF achieved a probability of detection of 95.7%, accuracy of 97.17%, and an F1-score of 96.93%, while maintaining excellent performance in low signal-to-noise ratio (SNR) conditions, even surpassing existing hybrid DL models. These findings underscore RF’s exceptional noise resilience and establish it as an ideal, high-performance candidate for practical spectrum sensing in wireless networks.

The aim of the study: to develop a method for predicting the development of alcohol delirium in patients with initial manifestations of alcohol withdrawal syndrome. Materials and methods: 4 laboratory indicators were used to build prognostic models: the number of platelets in the blood of patients, levels of potassium, sodium and chlorine in the blood serum. The probabilistic model is based on the multilayer perceptron (MLP) algorithm, binary on the random forest (Random Forest, RF) algorithm. To train the models, an anonymized database of patients with alcohol withdrawal syndrome (498 people) was used, 295 patients from this sample had alcohol delirium, and 203 had uncomplicated withdrawal syndrome. Study results: Both models on the test sample showed good results: the average value of the forecast accuracy of the model based on the MLP algorithm - 84%, on RF - 83%. The MLP model was verified in an addiction hospital setting, with 84.4% accuracy of correct predictions

Depression (DEP) is a common yet underdiagnosed comorbidity in adults with type 2 diabetes mellitus (T2DM), worsening glycemic control and increasing complication risk. Practical, interpretable risk tools using routine patient data are limited. We conducted a cross-sectional analysis using data from adults with T2DM enrolled in the National Health and Nutrition Examination Survey between 2009 and 2023. DEP was classified based on a Patient Health Questionnaire-9 score of 10 or higher. Twenty-eight candidate predictors encompassing demographic characteristics, clinical and biochemical measurements, and lifestyle factors were initially included. Variable selection was performed using least absolute shrinkage and selection operator regression. Five machine learning algorithms - random forest, extreme gradient boosting (XGBoost), multilayer perceptron, logistic regression, and support vector machine - were trained and evaluated using 5-fold cross-validation. The best-performing model was interpreted through SHapley Additive exPlanations analysis to identify the most influential predictors. A streamlined version incorporating the top 10 predictors was further developed and implemented as a user-friendly web-based risk estimation tool. Among 2837 participants, 449 (15.8%) were identified as having comorbid DEP. The XGBoost model demonstrated the highest discriminative ability, with a validation area under the receiver operating characteristic curve of 0.888, accuracy of 0.834, F1-score of 0.715, sensitivity of 0.577, and specificity of 0.979, surpassing the performance of the other algorithms evaluated. SHapley Additive exPlanations analysis revealed gender, poverty-to-income ratio, sleep duration, smoking status, educational levels, race, age, high cholesterol, hypertension, and insulin use as the most influential predictors. A streamlined XGBoost model incorporating only these 10 variables achieved an area under the curve of 0.886, closely matching the predictive capability of the full model. The deployed web-based tool enables rapid and individualized estimation of DEP risk in patients with T2DM using routinely available clinical and demographic information. Explainable machine learning applied to nationally representative data can accurately identify adults with T2DM at heightened risk of DEP using a small set of noninvasive clinical features. The deployed tool offers a scalable, interpretable, and clinically actionable approach to support early detection and intervention, potentially improving mental health outcomes in this high-risk population.

ABSTRACT This study addresses the increasing issue of vacant houses in urban areas, particularly in South Korea, by introducing a predictive framework that integrates machine learning and spatial autocorrelation. Focusing on detached housing in Jinju-si, a mid-sized city experiencing population decline, we employ generalized additive models (GAM), random forest (RF), and support vector machines (SVM) to identify high-risk areas for vacancy. Among the models tested, GAM achieved the highest predictive accuracy (R2 = 0.62), outperforming OLS (R2 = 0.51), RF (R2 = 0.48), and SVM (R2 = 0.39). The analysis highlights key influencing factors such as building age, land price, and proximity to pollutants, and shows how incorporating spatially lagged variables improves prediction performance. Findings reveal that vacant houses tend to cluster in older neighborhoods and spread spatially, underscoring the need for early intervention. This study provides data-driven insights for urban regeneration policies targeting housing stability and vacancy mitigation.

Evaluating roundabout safety in low- and middle-income countries is challenging due to unreliable crash data. This study addresses the issue by analyzing safety perception data from 1,530 questionnaire respondents across two Indian cities. A novel framework was developed to compare advanced ordinal ensemble models (Ordered XGBoost, LightGBM, Random Forest) against conventional ordinal regression (Logit, Probit). The ensemble models proved vastly superior, achieving Quadratic Weighted Kappa (QWK) scores exceeding 0.94, while conventional methods scored below 0.53. The top-performing Ordered XGBoost model (QWK=0.97) was interpreted using the state-of-the-art explainable artificial intelligence (XAI) technique SHAP (SHapley Additive exPlanations). SHAP analysis quantified the influence of key factors on perceived risk, identifying personal attributes (occupation, accident/near-accident experience) and infrastructure deficits (inadequate lighting, missing navigational aids) as primary drivers. The findings offer SHAP-quantified insights for deploying targeted, evidence-based safety interventions, providing a blueprint for improving perceived safety in complex traffic environments where traditional analysis is infeasible.

Adulteration of olive oil significantly compromises the interests of both producers and consumers, making its authentication a crucial challenge in the food industry. This study explored the potential of combining Raman spectroscopy with machine learning for discriminating various blended samples and quantifying olive oil content in mixtures. Raman features, such as peak intensities at specific shifts, were extracted from the spectra and analyzed using hierarchical cluster analysis (HCA) and correlation analysis (CA) to identify significant variations corresponding to altered proportions of olive oil. Qualitative and quantitative analyses were performed to classify 10 oil types and predict compositional ratios in binary and ternary blends, comparing different chemometric techniques and input features. Among these, the random forest (RF) model yielded a high classification accuracy (98.9%) and strong predictive performance, with coefficients of determination (R2) of 0.985 and 0.926 on the binary and ternary samples, respectively. The Shapley additive explanations (SHAP) algorithm was subsequently employed to assess the contribution of key Raman features to the prediction accuracy of superior models. Overall, this novel analytical framework highlights Raman features and offers a promising solution for real-time quality monitoring of olive oil products.

Introduction Tannin content is a key factor influencing the taste of walnuts and serves as an important index for evaluating walnut quality. Rapid and accurate detection of tannin levels in walnut kernels is therefore significant for quality assessment and management. This study aims to develop an efficient method for predicting tannin content in walnut kernels using near-infrared (NIR) spectroscopy combined with machine learning techniques. Methods A total of 180 samples of ‘Wen 185’ walnut kernels were used as the research objects. The NIR reflectance spectra of the samples were measured within the range of 4000–10000 cm⁻¹. The spectral data were processed using mathematical transformations and continuous wavelet transform (CWT), both separately and in combination. Pearson correlation analysis was applied to extract characteristic bands related to tannin content. Based on these features, a random forest (RF) model was constructed to quantitatively predict tannin content. Additionally, the SHAP algorithm was employed to interpret and visualize the machine learning model. Results The results indicated that within the spectral range of 4000–10000 cm⁻¹, the NIR reflectance of walnut kernels increased with tannin content under different orchard management modes. Both first-order differential transformation and CWT, as well as their combination, significantly enhanced the correlation between spectral data and tannin content. The combination of first-order differential transformation and CWT notably improved the model's prediction performance. The optimal prediction model was achieved using the feature lg’(1/R)_CWT_28, with training set metrics of R² = 0.880, RMSE = 1.188, RPD = 2.904, and validation set metrics of R² = 0.831, RMSE = 1.620, RPD = 2.459. Discussion The study demonstrates that combining mathematical transformations with wavelet transform can effectively improve the prediction accuracy of models for tannin content in walnut kernels. The RF model based on processed spectral data showed strong performance, indicating its potential for rapid and non-destructive tannin quantification. The use of SHAP algorithm further enhances model interpretability. These findings provide a valuable reference for the accurate prediction of tannin content in walnut kernels and may support quality control in walnut production and processing.

Atlas cedar (Cedrus atlantica) forests in Morocco’s Middle Atlas are experiencing an accelerated decline due to combined climatic and human pressures. Building on previous work on forest transition modeling, this study presents a deep-learning–based framework designed to predict and monitor the ecological resilience of Atlas cedar ecosystems. Multi-sensor satellite images from Landsat-8 and Sentinel-2, combined with field inventory data from the Ain Leuh–Sidi M’Guild massif, were processed to evaluate vegetation health, canopy density, and regeneration potential from 2013 to 2024. A hybrid Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN–BiLSTM) model was built to capture both spatial and temporal patterns of forest loss and recovery. Spectral indices such as NDVI, NBR, NDMI, and SAVI were extracted and standardized, while terrain features (altitude, slope, aspect) and bioclimatic variables (temperature seasonality, precipitation during the driest quarter) were included in the model. The hybrid CNN–BiLSTM architecture achieved an overall prediction accuracy of 94.7%, surpassing traditional machine learning methods (Random Forest, SVM, and Gradient Boosting). The spatio-temporal projections reveal a notable decline (−62%) of high-density cedar stands in low-elevation areas, while upper-slope refugia show partial stability and higher regeneration likelihoods. These results demonstrate the potential of deep learning combined with high-resolution Earth observation data for real-time forest health monitoring and adaptive management. The developed framework provides an operational foundation for Morocco’s Forest Strategy 2020–2030, enabling proactive decision-making for climate-resilient reforestation and ecological restoration in Mediterranean mountain ecosystems.

Objectives: The incidence of stroke is higher among type 2 diabetes mellitus (T2DM) patients with a higher mortality rate. Prognostic scores for stroke patients can assist with treatment planning and counseling. The objective of this study was to create a machine-learning-based prognostic score to estimate in-hospital mortality in acute stroke with T2DM. Materials and Methods: This study used data from claims-based diabetes registry at Dr. Sardjito General Hospital, Yogyakarta, Indonesia, to identify patients diagnosed with acute stroke and T2DM between January 2016 and December 2020. Four machine learning algorithms were trained and evaluated based on standard performance metrics. Important features were selected from the best-performing model and implemented in a web-based in-hospital mortality prediction scoring system. Results: Of the 18,652 patients in the registry, the final analytic dataset comprised 749 patients (557 survivors and 192 non-survivors). The random forest showed superiority compared to other models. The six most important features were length of stay, sepsis, pneumonia, age, dyslipidemia, and hemiplegia. Using these features, the web-based system estimates the probability of in-hospital death for an individual patient. Conclusion: Machine learning analysis may support an in-hospital mortality prediction score for patients with acute stroke and T2DM patients by leveraging the key features identified by the random forest model.

ABSTRACT In this study, the effects of cutting parameters, namely cutting speed (v), feed rate (f), depth of cut (t), and machining diameter (d), on surface roughness in external turning of C45 steel were investigated using the Taguchi method. Taguchi analysis, Random Forest, and ANOVA were employed to identify the factors affecting surface roughness. The results revealed that feed rate had the most significant effect, followed by machining diameter, depth of cut, and cutting speed. Four regression models, including polynomial regression, Random Forest Regression (RFR), Artificial Neural Network (ANN), and Extreme Learning Machine (ELM), were developed to predict surface roughness based on cutting parameters. Among them, the ELM model demonstrated the highest prediction accuracy, characterized by a high coefficient of determination (R2) value and low mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean squared error (RMSE). Therefore, the ELM model is considered the most suitable for predicting surface roughness in precision external turning operations.

Introduction & purpose: Physical activity (PA) is vital for the health of individuals with physical disabilities (IWPDs), reducing their risk of multimorbidity (Brinkhof et al., 2016). As IWPDs are significantly less active than the general population (WHO, 2022), organised sport – like regular wheelchair sports groups offered by wheelchair clubs of the Swiss Paraplegic Association (SPA) – can serve as a PA promoting setting. While research, including multilevel approaches, has studied intention and adherence in leisure-time PA settings, only scarce research exists for settings like regular wheelchair sports groups (Martin Ginis et al., 2016; Schlesinger & Nagel, 2015; Sivaramakrishnan et al., 2023). Applying a socio-ecological approach, this study investigated the relevance of individual and club-level factors on adherence in wheelchair sport groups among all members (active vs. inactive) (RQ 1) and the influence of individual factors on intention to participate in a wheelchair sport group (non-participants, RQ2) among SPA members. Methods: This study employs an explorative cross-sectional quantitative research design. Data was collected from February until June 2025 including an online survey of SPA members (n=273), targeting a highly specific sample of IWPDs in Switzerland, and structured interviews with board members (n=26). Adherence to sport (yes/no) and intention to participate (0-100) were primary outcomes. For RQ1, a logistic multilevel regression model was employed. Due to a low ICC (.06) in the null model, a small, predetermined number of clusters (n=26 SPA wheelchair clubs) and highly unbalanced clusters, only L1 predictors were included in the analysis, retaining the random intercept to adhere to the multilevel data structure. For RQ2, random forests will be used, as successfully applied in other IWPDs research (e.g. Gross-Hemmi et al., 2021). Results: For RQ1, small cluster variance (tau^2_00 =.22) was found, indicating little difference between wheelchair clubs’ affiliation regarding the outcome variable. Using a backwards-selection approach, the best-fitting model identified standardized ORs (change per one SD) for motivation for sport (OR=3.71, 95%CI: [2.12, 6.42], p < .001), wellbeing (OR=2.32, 95%CI: [1.42, 3.74], p < .001), and age (OR=.57, 95%CI: [.40, .82], p < .01), negatively associated with the odds of adherence, as the strongest individual-level predictors. Furthermore, overall satisfaction with club characteristics (OR=1.70, 95%CI : [1.12, 2.56], p < .05), satisfaction with social life (OR= .65, 95%CI: [.43, .99], p < .05), and being female (OR=.53, 95%CI: [.26, 1.08], p ≤ .10) were also important predictors, with the latter two negatively associated with the odds of adherence. Predictors for intention (RQ2) from the random forest analysis will be detailed in the presentation. Discussion/conclusion: The findings underscore the high importance of psychological factors (e.g. motivation, wellbeing) as strong predictors of sports groups adherence, confirming existing literature and providing novel insights, notably the predictive role of satisfaction with club characteristics (Martin Ginis et al., 2016; Schlesinger & Nagel, 2015). The results suggest tailored interventions focusing on both individual factors and club support. While the self-report, cross-sectional design limits causal claims, these findings are valuable for guiding future research and informing strategy development by national federations (e.g. SPA).

Produced water management is one of the major environmental concerns in the Niger Delta, where the operation of oil production generates enormous volumes of effluents with complex chemical characteristics. In this study, hybrid, physics-based, and machine learning models were formulated and compared for the prediction of main produced water quality parameters of pH, Total Dissolved Solids (TDS), Oil and Grease (O&G), Heavy Metal Concentration (HMC), and Chemical Oxygen Demand (COD). Historical monitoring data from 2010 to 2023 were fitted using five types of models: Autoregressive Integrated Moving Average (ARIMA), ARIMA–Long Short-Term Memory (ARIMA–LSTM), Physics-Informed LSTM (PI–LSTM), Random Forest (RF), and a physics-based process model. Model performance was compared using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), coefficient of determination (R²), and probabilistic forecast intervals. Amongst models, the hybrid PI–LSTM consistently performed better than the rest in terms of prediction accuracy (RMSE = 12.6, MAE = 8.8, R² = 0.87) in terms of seasonal variability and long-term dependency capture for all parameters. The physics-based model provided interpretive insights into water–hydrocarbon interactions and production system dynamics. Overall, results indicate that the integration of physical principles into deep learning models enhances predictive performance and interpretability of water quality predictions generated. Results have significant implications for Niger Delta environmental monitoring, regulatory decision-making, and sustainable produced water management.

Natural hydrogen has gained attention as a low-carbon energy vector, and some reported surface expressions have been linked to subcircular patterns, or fairy circles (FC), that may be detectable in multispectral satellite imagery. The Carolina Bays region, on the eastern coast of the United States, was selected because it hosts abundant, well-mapped subcircular depressions. This study aims to comparatively evaluate machine learning algorithms for identifying subcircular structures using Landsat-8 data. We processed 105 Collection 2 Level 2 scenes, masking clouds and shadows using the Level 2 quality band. Pixel-level labels were determined using a well-curated public dataset, derived from a high-resolution LiDAR survey. Traditional models (logistic regression, random forest, and multilayer perceptron) achieved precision scores below 0.66 and enabled a variable-importance analysis, which identified Band 3 (green), Band 6 (SWIR1), and five Normalised Unit Indices as the most predictive features. Deep learning models improved detection, and a U-Net architecture allowed for pixel-level segmentation of FC-like structures, producing false positives mostly in cloudy or shadowed areas. Overall, the results suggest that FC detection from multispectral data alone remains challenging due to class overlap and cloud/shadow contamination. Future work could explore integrating additional non-spectral descriptors, such as morphometric variables, to reduce ambiguities.

Post-stroke seizures (PSS) manifests variably due to ischemic brain injury, yet its risk factors remain unclear. This study developed a machine learning (ML) model using clinical and laboratory data to predict PSS risk in acute ischemic stroke (AIS) patients post-thrombolysis, aiming to enhance risk assessment and clinical management. Retrospective analysis included 332 AIS patients treated between January 2020 and November 2024. Twenty-one variables (demographics, clinical parameters, lab biomarkers) were analyzed. Establish a diagnostic model with the occurrence of seizures after thrombolytic surgery as the classification variable. Missing data were handled via median/mean substitution, and class imbalance was corrected using synthetic minority oversampling technique. Feature selection combined expert consensus and Boruta algorithm. The dataset was split into training (70%) and testing (30%) cohorts. Seven ML models - logistic regression, Naïve Bayes, support vector machines, multilayer perceptron, AdaBoost, gradient boosting decision tree, and random forest (RF) - were evaluated using area under the curve (AUC), Brier score, accuracy, sensitivity, and specificity. Shapley Additive exPlanations (SHAP) analysis interpreted feature importance. PSS occurred in 39 patients (11.7%). Four predictors were identified: age, serum sodium, serum calcium, and fasting blood glucose. The RF model achieved optimal performance (AUC: 0.867, 95% CI: 0.793-0.930); accuracy: 0.803 (95% CI: 0.73-0.869); specificity: 0.810 (95% CI: 0.71-0.898), F1-score: 0.797 (95% CI: 0.711-0.87); positive predictive value: 0.797 (95% CI: 0.691-0.897), Kappa: 0.606 (95% CI: 0.441-0.739)). SHAP ranked fasting blood glucose as the strongest predictor, followed by serum sodium, serum calcium, and age. Lower electrolyte levels, elevated glucose, and younger age correlated with higher PSS risk. The model was deployed as a web-based clinical tool. The RF-based model effectively stratifies PSS risk in thrombolysis-treated AIS patients using accessible clinical variables. SHAP interpretability underscores fasting glucose, serum sodium/calcium, and age as pivotal predictors, offering actionable insights for prevention and personalized care. This tool may aid early intervention strategies to mitigate PSS burden.

Abstract Transverse cracking is a major distress mechanism in Continuously Reinforced Concrete Pavement (CRCP), affecting ride smoothness, service life, and maintenance strategies. This research introduces a hybrid predictive framework that couples Particle Swarm Optimization (PSO) with Gradient Boosting Machine (GBM) to enhance the accuracy of transverse crack prediction in CRCP. The analysis utilized 395 records from 33 pavement sections obtained from the Long-Term Pavement Performance (LTPP) program, encompassing structural, environmental, traffic, and performance-related parameters. PSO was applied to fine-tune critical GBM hyperparameters, namely the number of iterations, learning rate, and tree depth. The optimized PSO–GBM model demonstrated excellent performance, yielding an average RMSE of 1.62 and an R 2 of 0.99 under 5-fold cross-validation, surpassing benchmark models such as conventional GBM, Random Forest, Support Vector Regression (SVR), Artificial Neural Networks (ANN), and Linear Regression. Sensitivity analysis revealed that L3 thickness, L4 thickness, and Annual Average Daily Traffic (AADT) were the most significant contributors, consistent with engineering knowledge of crack development. Validation through residual distribution and equality line plots confirmed the robustness and stability of the proposed approach across varying severity levels.