Spatial-temporal gated transformer network for freeway secondary crash prediction considering the impact of class imbalance.

Soil and litter emission sources as important contributors to ozone production from volatile organic compounds in island tropical forests.

Aim/Purpose: The primary aim of this study is to address the persistent challenge of delayed autism spectrum disorder (ASD) diagnosis in toddlers. Early detection enables timely interventions that can improve developmental outcomes; however, conventional approaches rely on lengthy and resource-intensive behavioral assessments. We therefore introduce an interpretable AI screening framework designed to accelerate ASD triage while providing clinically understandable rationales to support decision-making. Background: Traditional ASD diagnosis depends on expert behavioral evaluation and parent reports which, despite their value, are time-consuming and capacity-limited, delaying access to early intervention. With ASD prevalence rising, scalable and effective approaches are urgently needed. This study proposes a robust AI framework for early ASD detection that integrates targeted preprocessing, feature selection, principled model optimization, and post-hoc explanations, aiming to improve diagnostic utility and clarity for end users in clinical and community settings. Methodology: We develop a unified, reproducible pipeline that combines data preprocessing, class balancing, feature selection, and Bayesian hyperparameter tuning. The pipeline also incorporates SHapley Additive exPlanations (SHAP) to provide model explanations. Six diverse machine learning models – Extreme Gradient Boosting (XGB), Histogram-based Gradient Boosting (HGB), Random Forest (RF), Naïve Bayes (NB), Mixture Discriminant Analysis (MDA), and Multi-layer Perceptron (MLP) – are evaluated to assess framework robustness rather than to crown a single best classifier. A cross-cultural dataset of toddlers aged 12–36 months (n=1,560) is constructed by merging two public sources containing Q-CHAT-10 items with demographic attributes. Preprocessing removes non-informative variables and encodes categorical features; Gaussian noise-based upsampling (GNUS) mitigates post-merge imbalance; RobustScaler stabilizes training. Gradient Boosting Feature Selection (GBFS) ranks and reduces features to enhance parsimony and interpretability. Performance is reported via accuracy, precision, recall, F1, and Matthews Correlation Coefficient (MCC). Model behavior is elucidated with SHAP to reveal feature contributions and decision pathways. Contribution: This work presents an interpretable AI framework for early ASD detection that couples performance with clinician-oriented explanation in a single pipeline. Rather than optimizing for accuracy alone, we emphasize synergy among preprocessing, balancing, feature selection, and explanation – the multimodel evaluation evidence adaptability across algorithmic families. GBFS and SHAP are used to ensure concise, explainable predictions. Notably, the framework achieved very strong internal validation results (high F1 and MCC across folds) with XGB, while SHAP-derived patterns aligned with clinical heuristics. Results are promising but preliminary, pending external, multi-site validation. Findings: GNUS and robust normalization improved generalization on the cross-cultural dataset. With GBFS-selected features, XGB achieved near-ceiling internal scores across key metrics, a trend observed – though to a lesser extent – in other models after comparable optimization. SHAP consistently highlighted behaviors such as gaze-following and social/emotional responsiveness among the most influential predictors, in line with clinical practice. Collectively, the findings indicate that interpretable ML can complement conventional screening, while warranting prospective and external validation to assess generalizability and potential dataset shift. Recommendations for Practitioners: Clinicians and community programs may consider adopting interpretable ML as a screening aid to prioritize referrals and shorten time-to-assessment. Attention to features repeatedly identified as influential can guide focused early interventions and resource allocation. Recommendation for Researchers: Future studies should test the framework on larger and more diverse cohorts to evaluate generalizability. Exploring ensembles and deeper architectures, as well as alternative preprocessing, resampling, and feature selection strategies, may further enhance performance, particularly for cases that are borderline. Impact on Society: Earlier, more reliable screening can improve access to services during critical neurodevelopmental windows. Integrating interpretable AI into practice may also strengthen clinician confidence in ML-assisted tools, supporting responsible, human-centered deployment and broader public health benefits. Future Research: Next steps include conducting real-world pilots across various clinical/community settings, integrating with complementary diagnostic tools to build multimodal platforms, and systematically evaluating balancing/optimization choices. These directions will help translate the framework into practical impact and inspire analogous applications in pediatric neurodevelopmental assessment.

Tree-based machine learning (ML) algorithms, such as Extra Trees (ET), Random Forest (RF), Gradient Boosting Machine (GBM), and XGBoost (XGB) are among the most widely used in early drug discovery, given their versatility and performance. However, models based on these algorithms often suffer from misclassification and reduced interpretability issues, which limit their applicability in practice. To address these challenges, several approaches have been proposed, including the use of SHapley Additive Explanations (SHAP). While SHAP values are commonly used to elucidate the importance of features driving models' predictions, they can also be employed in strategies to improve their prediction performance. Building on these premises, we propose a novel approach that integrates SHAP and features value analyses to reduce misclassification in model predictions. Specifically, we benchmarked classifiers based on ET, RF, GBM, and XGB algorithms using data sets of compounds with known antiproliferative activity against three prostate cancer (PC) cell lines (i.e., PC3, LNCaP, and DU-145). The best-performing models, based on RDKit and ECFP4 descriptors with GBM and XGB algorithms, achieved MCC values above 0.58 and F1-score above 0.8 across all data sets, demonstrating satisfactory accuracy and precision. Analyses of SHAP values revealed that many misclassified compounds possess feature values that fall within the range typically associated with the opposite class. Based on these findings, we developed a misclassification-detection framework using four filtering rules, which we termed "RAW", SHAP, "RAW OR SHAP", and "RAW AND SHAP". These filtering rules successfully identified several potentially misclassified predictions, with the "RAW OR SHAP" rule retrieving up to 21%, 23%, and 63% of misclassified compounds in the PC3, DU-145, and LNCaP test sets, respectively. The developed flagging rules enable the systematic exclusion of likely misclassified compounds, even across progressively higher prediction confidence levels, thus providing a valuable approach to improve classifier performance in virtual screening applications.

Sublobar resection is suitable for peripheral stage I lung adenocarcinoma (LUAD). However, if tumor spread through air spaces (STAS) present, the lobectomy will be considered for a survival benefit. Therefore, STAS status guide peripheral stage I LUAD surgical approach. This study aimed to identify radiological features associated with STAS in peripheral stage I LUAD and to develop a predictive machine learning (ML) model using radiomics to improve surgical decision-making for thoracic surgeons. We conducted a retrospective analysis of patients who underwent surgical treatment for lung tumors from January 2022 to December 2023, focusing on clinical peripheral stage I LUAD. High-resolution computed tomography (CT) scans were used to extract 1,581 radiomics features. Least absolute shrinkage and selection operator (LASSO) regression was applied to select the most relevant features for predicting STAS, reducing model overfitting and enhancing predictability. Ten ML algorithms were evaluated using performance metrics such as area under the receiver operating characteristic curve (AUROC), accuracy, recall, F1-score, and Matthews Correlation Coefficient (MCC) after a 10-fold cross-validation process. SHapley Additive exPlanations (SHAP) values were calculated to provide interpretability and illustrate the contribution of individual features to the model's predictions. Additionally, a user-friendly web application was developed to enable clinicians to use these predictive models in real-time for assessing the risk of STAS. The study identified significant associations between STAS and radiological features, including the longest diameter, consolidation-to-tumor ratio (CTR), and the presence of spiculation. The Random Forest (RF) model for 3-mm peritumoral extensions demonstrated strong predictive performance, with a Recall_Mean of 0.717, Accuracy_Mean of 0.891, F1-Score_Mean of 0.758, MCC_Mean of 0.708, and an AUROC_Mean of 0.944. SHAP analyses highlighted the influential radiomics features, enhancing our understanding of the model's decision-making process. The RF model, employing specific intratumoral and 3-mm peritumoral radiomics features, was highly effective in predicting STAS in peripheral stage I LUAD. This model is recommended for clinical use to optimize surgical strategies for LUAD patients, supported by a real-time web application for STAS risk assessment.

Anemia is one of the common adverse reactions after hip fracture surgery. The traditional method to solve anemia is allogeneic transfusion. However, the transfusion may lead to some complications such as septicemia and fever. So far, few studies have reported roles of machine learning in predicting whether blood transfusion is needed or not after hip fracture surgery. Therefore, the purpose of this study is to develop machine learning models to predict the likelihood of postoperative blood transfusion in patients undergoing hip fracture surgery. This study enrolled 1355 patients who underwent hip fracture surgery at the Affiliated Hospital of Qingdao University from January 2016 to December 2021. Among all patients, 210 cases received postoperative blood transfusion. All patients were randomly divided into a training group and a testing group at a ratio of 7:3. In the training group, univariate and multivariate logistic regression analyses were used to determine independent risk factors for the postoperative transfusion. Then, based on these independent risk factors, tenfold cross-validation method was utilized to develop five machine learning models, including logistic, multilayer perceptron (MLP), extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM). The receiver operating characteristic (ROC) curve, area under ROC curve (AUC), and Matthews correlation coefficient (MCC) were generated to evaluate the performance of the models. Calibration plot and decision curve analysis (DCA) were used to test the performance, stability, and clinical applicability of the models. The models were validated using the testing group; and the ROC curve, MCC, calibration plot, and DCA curves were also generated to validate the performance, stability, and clinical applicability of the models. To further verify the robustness of the model, we randomly grabbed 70% of the samples in the testing set, performed 1000 iterations, and calculated the AUC and confidence interval of the five models. Finally, we used SHapley Additive exPlanations (SHAP) to explain these models. Multivariate logistic regression analysis showed that there were 8 independent risk factors, including age, blood transfusion history, albumin (ALB), globulin (GLO), total bilirubin (TBIL), indirect bilirubin (IBIL), hemoglobin (HB), and blood loss > 200ml. We finally selected five independent risk factors including HB, GLO, age, IBIL, and blood loss > 200ml. Based on these five independent risk factors, we generated six characteristic variables, namely HB, HB × HB, HB × blood loss, GLO × HB, age, age × IBIL, and established five machine learning models using a tenfold cross-validation method. In the training group, the AUC values of logistic, RF, MLP, SVM, and XGB were 0.9320, 0.8911, 0.9327, 0.9225, and 0.8825, respectively, and the average AUC was 0.9122 ± 0.0212. The MCC values were 0.65, 0.77, 0.65, 0.66, and 0.68, respectively, and the calibration plot and DCA performed well. In the testing group the AUC values of logistic, RF, MLP, SVM, and XGB were 0.8483, 0.7978, 0.8576, 0.8598, and 0.8216, respectively. The average AUC was 0.8370 ± 0.0238, and the MCC values were 0.41, 0.35, 0.40, 0.41, and 0.41, respectively. The calibration plot and DCA in the testing group also showed good performance. The AUC values and confidence intervals of the 1000-iteration model were: logistic (AUC, min confidence interval [CI]-max confidence interval [CI] 0.848, 0.804-0.903), RF (AUC, minCI-maxCI 0.797, 0.734-0.857), MLP (AUC, minCI-maxCI 0.858, 0.812-0.902), SVM (AUC, minCI-maxCI 0.859, 0.819-0.910), and XGB (AUC, minCI-maxCI 0.821, 0.764-0.894). The model performed well. Finally, according to SHAP, among all five models, HB played the most important role in model prediction and interpretation. The five models we developed all performed well in predicting the likelihood of blood transfusion after hip fracture surgery. Therefore, we believed that the prediction model based on machine learning had great application prospects in clinical practice, which could help clinicians better predict the risk of blood transfusion after hip fracture surgery.

Machine learning-driven clinical decision support for low bone mineral density: A web-based prediction model with explainable AI integration.

Failures within water distribution networks (WDNs) lead to significant environmental and economic impacts. While existing research has established various predictive models for pipe failures, there remains a lack of studies focusing on the probability of leaks and bursts. Addressing this gap, the present study introduces a new approach that harnesses deep learning algorithms - Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and TabNet for failure prediction. The study enhances these base models by optimising their hyperparameters using Bayesian Optimisation (BO) and further refining the models through data scaling. The Copeland algorithm and SHapley Additive exPlanations (SHAP) are also applied for model ranking and interpretation, respectively. Applying this methodology to Hong Kong's WDN data, the study evaluates the models' predictive performance across several metrics, including accuracy, precision, recall, F1 score, Matthews Correlation Coefficient (MCC), and Cohen's Kappa. Results demonstrate that BO significantly enhances the models' predictive abilities, such that the TabNet model's F1 score for leak prediction increases by 36.2% on standardised data. The Copeland algorithm identifies CNN as the most effective model for predicting both leak and burst probabilities. As indicated by SHAP values, critical features influencing model predictions include pipe diameter, material, and age. The optimised CNN model has been deployed as user-friendly web applications for predicting the probability of leaks and bursts, enabling both single-pipe and batch predictions. This research provides crucial insights for WDN management, equipping water utilities with sophisticated tools to forecast the probability of pipe failure, enabling more effective mitigation of such failures.

Identifying cardiovascular disease risk in the U.S. population using environmental volatile organic compounds exposure: A machine learning predictive model based on the SHAP methodology

BackgroundIschemic heart disease (IHD) and type 2 diabetes mellitus (T2DM) are leading causes of disability-adjusted life years globally among adults aged 55 years and older. Although both diseases share common risk factors and pathophysiological pathways, previous research has predominantly addressed these conditions in isolation. The co-occurrence patterns and regional variations of IHD and T2DM burden remain poorly understood. We aimed to characterize the global co-occurrence patterns of IHD and T2DM from a spatial perspective and to identify the corresponding risk factors distinguishing different burden regions.MethodsUsing data from the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) 2021 database, we extracted age-standardized disability-adjusted life year (DALY) rates for IHD and T2DM among individuals aged 55 years and older from 204 countries and territories. Based on quartile distributions of global DALY rates for both diseases, we classified countries into four distinct burden regions: Low-Burden Regions (56 countries), T2DM-Dominant Regions (46 countries), IHD-Dominant Regions (46 countries), and Dual-Burden Regions (56 countries). We examined temporal trends from 1990-2021, computed population attributable fractions for major risk factors, and used machine learning-based SHAP (Shapley Additive Explanations) analysis to screen and quantify the effects of corresponding risk factors distinguishing regional classifications.ResultsDual-Burden Regions were distributed across multiple geographic areas including the Caribbean and Central America, Persian Gulf states, Balkan Peninsula, Southeast Asia, West Africa, Eastern Mediterranean, and Northern Europe. The spatial distribution revealed distinct geographic clustering, with higher IHD rates in Eastern Europe and Central Asia, and elevated T2DM rates in Pacific Island nations and parts of the Middle East. Countries and territories with the highest burden for both diseases included North African countries (eg, Morocco: IHD 25,193.1/100,000 and T2DM 32,197.24/100,000) and Pacific Island nations such as Fiji exhibiting IHD burden of 24,758.17 per 100,000 and T2DM burden of 32,197.24 per 100,000. Marshall Islands showed IHD burden of 25,107.72/100,000 and T2DM burden of 22,122.46/100,000, while Nauru demonstrated the highest IHD burden (39,483.92/100,000). High systolic blood pressure contributed most to IHD burden globally (49.79%), while high body-mass index dominated T2DM burden (51.89%). Environmental factors demonstrated clear regional gradients, with household air pollution ranging from 4·58% in Low-Burden to 14.43% in Dual-Burden Regions for IHD. High body-mass index contributed 51.89% to T2DM burden globally, with regional variation from 40.61% in IHD-Dominant to 51.36% in Low-Burden Regions. SHAP analysis identified sociodemographic index (SDI2021) as the primary factor distinguishing Low-Burden from Dual-Burden Regions for both IHD (mean |SHAP| = 1.245) and T2DM (mean |SHAP| = 1.317). Diet high in processed meat consistently showed strong discriminatory power across multiple regional comparisons for T2DM (SHAP values 0.923-1.721), while secondhand smoke emerged as a critical differentiator with SHAP values exceeding 1.0 across various regional distinctions. Diet low in vegetables served as a primary differentiator between Low-Burden and T2DM-Dominant Regions (mean |SHAP| = 1.188).ConclusionThe co-occurrence of IHD and T2DM exhibits pronounced global heterogeneity, with Pacific Island nations and multiple geographic regions including Gulf states, North Africa, and other areas bearing disproportionate dual-burden. Socioeconomic development level fundamentally characterizes dual-burden status, while dietary and environmental factors serve as key regional differentiators. Intervening in modifiable risk factors, particularly processed meat consumption, vegetable intake, and environmental exposures, can fundamentally reduce the global burden of these co-occurring diseases.

Postoperative pancreatic fistula (POPF) remains one of the most relevant complications following pancreaticoduodenectomy (PD), significantly impacting short-term outcomes and delaying adjuvant therapies. Current predictive models offer limited accuracy, often failing to incorporate early postoperative data. This retrospective study aimed to develop and validate machine learning (ML) models to predict the absence and severity of POPF using clinical, surgical, and early postoperative variables. Data from 216 patients undergoing PD were analyzed. A total of twenty-four machine learning (ML) algorithms were systematically evaluated using the Matthews Correlation Coefficient (MCC) and AUC-ROC metrics. Among these, the GradientBoostingClassifier consistently outperformed all other models, demonstrating the best predictive performance, particularly in identifying patients at low risk of postoperative pancreatic fistula (POPF) during the early postoperative period. To enhance transparency and interpretability, a SHAP (SHapley Additive exPlanations) analysis was applied, highlighting the key role of early postoperative biomarkers in the model predictions. The performance of the GradientBoostingClassifier was also directly compared to that of a traditional logistic regression model, confirming the superior predictive performance over conventional approaches. This study demonstrates that ML can effectively stratify POPF risk, potentially supporting early drain removal and optimizing postoperative management. While the model showed promising performance in a single-center cohort, external validation across different surgical settings will be essential to confirm its generalizability and clinical utility. The integration of ML into clinical workflows may represent a step forward in delivering personalized and dynamic care after pancreatic surgery.

The issue of weak interpretability in geological disaster susceptibility assessments using machine learning models has been a long-standing concern. Although SHAP (Shapley Additive Explanations) models have been extensively used in recent years to interpret the decision-making details of models, the specialized skills required and the non-intuitiveness of SHAP plots make their application challenging in practical decision-making environments. In response, our study introduces a map-based SHAP visualization framework to enhance the interpretability of susceptibility assessment results. Utilizing Optuna for hyperparameter tuning, we developed a high-performance XGBoost model to assess the susceptibility of the most impactful disaster in Hangzhou: urban road collapses. In addition to interpreting the contributions of evaluation factors through traditional SHAP summaries and bar plots, we displayed the SHAP values for each evaluation factor using map visualizations, and discussed the model’s sensitivity to different values. To validate the alignment between model predictions and physical collapse mechanisms, our study selected typical collapse cases, interpreted these cases combining map visualizations, SHAP force plots at collapse points, and the physical mechanisms of collapse. Our research improves the interpretability of susceptibility assessments with machine learning by using map visualizations, providing new insights into spatial effects and robust support for urban decision-making applications.

There is a growing need for labeled data, yet manual annotation is costly, error-prone, and often infeasible in privacy-sensitive, highly imbalanced domains such as fraud detection. We introduce a fully unsupervised framework that combines unsupervised SHapley Additive exPlanations (SHAP) feature selection with our novel unsupervised labeling method. We apply unsupervised SHAP to the Kaggle Credit Card Fraud Detection and Medicare Part D datasets to produce high-impact feature subsets, and then label the datasets with our unsupervised labeling approach. To effectively evaluate the labels generated by our novel methodology, we apply a baseline unsupervised learner, Isolation Forest (IF), to both the original datasets and their subsets. We calculate Matthew’s Correlation Coefficient (MCC), Jaccard Index (JI), Precision, Recall, and F1-score by comparing our generated labels against the ground truth labels. It is important to note, the ground truth labels were used solely for evaluation. Our empirical results surpass the results obtained with the full feature dataset and baseline. By improving label quality while reducing computational complexity and preserving privacy, our approach offers a practical solution for learning from unlabeled, severely imbalanced data.

Ameliorating microglia-mediated neuroinflammation is a crucial strategy in developing new drugs for neurodegenerative diseases. Plant compounds are an important screening target for the discovery of drugs for the treatment of neurodegenerative diseases. However, due to the spatial complexity of phytochemicals, it becomes particularly important to evaluate the effectiveness of compounds while avoiding the mixing of cytotoxic substances in the early stages of compound screening. Traditional high-throughput screening methods suffer from high cost and low efficiency. A computational model based on machine learning provides a novel avenue for cytotoxicity determination. In this study, a microglia cytotoxicity classifier was developed using a machine learning approach. First, we proposed a data splitting strategy based on the molecule murcko generic scaffold, under this condition, three machine learning approaches were coupled with three kinds of molecular representation methods to construct microglia cytotoxicity classifier, which were then compared and assessed by the predictive accuracy, balanced accuracy, F1-score, and Matthews Correlation Coefficient. Then, the recursive feature elimination integrated with support vector machine (RFE-SVC) dimension reduction method was introduced to molecular fingerprints with high dimensions to further improve the model performance. Among all the microglial cytotoxicity classifiers, the SVM coupled with ECFP4 fingerprint after feature selection (ECFP4-RFE-SVM) obtained the most accurate classification for the test set (ACC of 0.99, BA of 0.99, F1-score of 0.99, MCC of 0.97). Finally, the Shapley additive explanations (SHAP) method was used in interpreting the microglia cytotoxicity classifier and key substructure smart identified as structural alerts. Experimental results show that ECFP4-RFE-SVM have reliable classification capability for microglia cytotoxicity, and SHAP can not only provide a rational explanation for microglia cytotoxicity predictions, but also offer a guideline for subsequent molecular cytotoxicity modifications.

Background and Objectives: This study aims to evaluate the diagnostic potential of routinely available hematological parameters for acute myocardial infarction (AMI) by employing an Explainable Neural Network (ENN) model that combines high predictive accuracy with interpretability. Materials and Methods: A publicly available dataset comprising 981 individuals (477 AMI patients and 504 controls) was analyzed. A broad set of hematological features—including white blood cell subtypes, red cell indices, and platelet-based markers—was used to train an ENN model. Bootstrap resampling was applied to enhance model generalizability. The model’s performance was assessed using standard classification metrics such as accuracy, sensitivity, specificity, F1-score, and Matthews Correlation Coefficient (MCC). SHapley Additive exPlanations (SHAP) were employed to provide both global and individualized insights into feature contributions. Results: The study analyzed hematological and biochemical parameters of 981 individuals. The explainable neural network (ENN) model demonstrated excellent diagnostic performance, achieving an accuracy of 94.1%, balanced accuracy of 94.2%, F1-score of 93.9%, and MCC of 0.883. The AUC was 0.96, confirming strong discriminative ability. SHAP-based explainability analyses highlighted neutrophils (NEU), white blood cells (WBC), RDW-CV, basophils (BA), and lymphocytes (LY) as the most influential predictors. Individual- and class-level SHAP evaluations revealed that inflammatory and erythrocyte-related parameters played decisive roles in AMI classification, while distributional analyses showed narrower parameter ranges in healthy individuals and greater heterogeneity among patients. Conclusions: The findings suggest that cost-effective, non-invasive blood parameters can be effectively utilized within interpretable AI frameworks to enhance AMI diagnosis. The integration of ENN with SHAP provides a dual benefit of diagnostic power and transparent rationale, facilitating clinician trust and real-world applicability. This scalable, explainable model offers a clinically viable decision-support tool aligned with the principles of precision medicine and ethical AI.

This study introduces APD-FFNet (Automated Periodontitis Diagnosis-Feature Fusion Network), a novel, explainable deep learning architecture for automated periodontitis diagnosis using panoramic radiographs. A total of 337 panoramic radiographs, annotated by a periodontist, served as the dataset. APD-FFNet combines custom convolutional and transformer-based layers within a deep feature fusion framework that captures both local and global contextual features. Performance was evaluated using accuracy, the F1 score, the area under the receiver operating characteristic curve, the Jaccard similarity coefficient, and the Matthews correlation coefficient. McNemar's test confirmed statistical significance, and SHapley Additive exPlanations provided interpretability insights. APD-FFNet achieved 94% accuracy, a 93.88% F1 score, 93.47% area under the receiver operating characteristic curve, 88.47% Jaccard similarity coefficient, and 88.46% Matthews correlation coefficient, surpassing comparable approaches. McNemar's test validated these findings (P < .05). Explanations generated by SHapley Additive exPlanations highlighted important regions in each radiograph, supporting clinical applicability. By merging convolutional and transformer-based layers, APD-FFNet establishes a new benchmark in automated, interpretable periodontitis diagnosis, with low hyperparameter sensitivity facilitating its integration into regular dental practice. Its adaptable design suggests broader relevance to other medical imaging domains. This is the first feature fusion method specifically devised for periodontitis diagnosis, supported by an expert-curated dataset and advanced explainable artificial intelligence. Its robust accuracy, low hyperparameter sensitivity, and transparent outputs set a new standard for automated periodontal analysis.