Class Imbalance Research Articles

Data Mining for the Adjustment of Credit Scoring Models in Solidarity Economy Entities: A Methodology for Addressing Class Imbalances

This study addresses the quantification of credit risk in solidarity economy entities, proposing a new methodology to redefine the concept of a “default” in the frequent situations of extreme class imbalances. The objective is to develop and evaluate credit scoring models that enhance risk management by incorporating internal and external data to assess default risk. Data mining techniques are applied to address class imbalances, redefining the term “default” to include external credit information and increasing the representation of the minority class. The effectiveness of machine learning and statistical models is evaluated using class-balancing methods such as under-sampling, over-sampling, and the Synthetic Minority Over-sampling Technique (SMOTE). The evaluation is based on the Balanced Accuracy metric and the holding power of the performance, ensuring a consistent predictive power of the model while avoiding overfitting. While machine learning methods can improve credit scoring, logistic regression-based models remain effective, especially when combined with class-balancing techniques. It is concluded that a balanced sample in a class size is essential to improve predictive performance.

Achieving Excellence in Cyber Fraud Detection: A Hybrid ML+DL Ensemble Approach for Credit Cards

The rapid advancement of technology has increased the complexity of cyber fraud, presenting a growing challenge for the banking sector to efficiently detect fraudulent credit card transactions. Conventional detection approaches face challenges in adapting to the continuously evolving tactics of fraudsters. This study addresses these limitations by proposing an innovative hybrid model that integrates Machine Learning (ML) and Deep Learning (DL) techniques through a stacking ensemble and resampling strategies. The hybrid model leverages ML techniques including Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Logistic Regression (LR) alongside DL techniques such as Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory Network (BiLSTM) with attention mechanisms. By utilising the stacking ensemble method, the model consolidates predictions from multiple base models, resulting in improved predictive accuracy compared to individual models. The methodology incorporates robust data pre-processing techniques. Experimental evaluations demonstrate the superior performance of the hybrid ML+DL model, particularly in handling class imbalances and achieving a high F1 score, achieving an F1 score of 94.63%. This result underscores the effectiveness of the proposed model in delivering reliable cyber fraud detection, highlighting its potential to enhance financial transaction security.

A novel hybrid ViT-LSTM model with explainable AI for brain stroke detection and classification in CT images: A case study of Rajshahi region.

Computed tomography (CT) scans play a key role in the diagnosis of stroke, a leading cause of morbidity and mortality worldwide. However, interpreting these scans is often challenging, necessitating automated solutions for timely and accurate diagnosis. This research proposed a novel hybrid model that integrates a Vision Transformer (ViT) and a Long Short Term Memory (LSTM) to accurately detect and classify stroke characteristics using CT images. The ViT identifies essential features from CT images, while LSTM processes sequential information generated by the ViT, adept at capturing crucial temporal dependencies for understanding patterns and context in sequential data. Moreover, our approach addresses class imbalance issues in stroke datasets by utilizing advanced strategies to improve model robustness. To ensure clinical relevance, Explainable Artificial Intelligence (XAI) methods, including attention maps, SHAP, and LIME, were incorporated to provide reliable and interpretable predictions. The proposed model was evaluated using the primary BrSCTHD-2023 dataset, collected from Rajshahi Medical College Hospital, achieving top accuracies of 73.80%, 91.61%, 93.50%, and 94.55% with the SGD, RMSProp, Adam, and AdamW optimizers, respectively. To further validate and generalize the model, it was also tested on the Kaggle brain stroke dataset, where it achieved an impressive accuracy of 96.61%. The proposed ViT-LSTM model significantly outperformed traditional CNNs and ViT models, demonstrating superior diagnostic performance and generalizability. This study advances automated stroke diagnosis by combining deep learning innovations, domain expertise, and enhanced interpretability to support clinical decision-making, providing reliable diagnostic solutions.

Machine learning framework for breast cancer detection with feature selection with L2 ridge regularization: insights from multiple datasets

Aim: This study aims to investigate and apply effective machine learning techniques for the early detection and precise diagnosis of breast cancer. The analysis is conducted using various breast cancer datasets, including Breast Cancer Wisconsin, Breast Cancer Diagnosis, NKI Breast Cancer, and SEER Breast Cancer datasets. The primary focus is on identifying key features and utilizing preprocessing methods to enhance classification accuracy. Methods: The datasets undergo several preprocessing steps, such as label encoding for categorical variables, linear regression for handling missing values, and Robust scaler normalization for data standardization. To address class imbalance, Tomek Link SMOTE is employed to improve dataset representation. Significant features are selected through L2 Ridge regularization, helping to pinpoint the most important predictors of breast cancer. A range of machine learning models, including decision tree, random forest, support vector machine (SVM), neural network, K-nearest neighbor, naïve bayes, extreme gradient boost (XGBoost), and AdaBoost, are applied for classification tasks. The performance of these models is assessed using metrics such as accuracy, precision, recall, F1-score, and the Kappa statistic. Additionally, the models' effectiveness is further evaluated using the receiver operating characteristic curve and precision-recall curve. Results: The XGBoost model achieved the best performance on both the breast cancer Wisconsin and diagnosis datasets. The SVM model reached 100% accuracy on the NKI breast cancer dataset, while the random forest model performed optimally on the SEER breast cancer dataset. The feature selection process through L2 Ridge regularization was crucial in enhancing the performance of these models. Conclusions: This work emphasizes the critical role of machine learning in improving breast cancer detection. By applying a combination of preprocessing techniques and classification models, the study successfully identifies significant features and boosts model performance. These findings contribute to the development of more accurate diagnostic tools, ultimately enhancing patient outcomes.

MVGNN-PPIS: A novel multi-view graph neural network for protein-protein interaction sites prediction based on Alphafold3-predicted structures and transfer learning.

Protein-protein interactions (PPI) are crucial for understanding numerous biological processes and pathogenic mechanisms. Identifying interaction sites is essential for biomedical research and targeted drug development. Compared to experimental methods, accurate computational approaches for protein-protein interaction sites (PPIS) prediction can save significant time and costs. In this study, we propose a novel model named MVGNN-PPIS. To the best of our knowledge, it is the first to utilize predicted structures generated by AlphaFold3, and combined with transfer learning techniques, for predicting PPIS. This approach addresses the limitations of traditional methods that depend on native protein structures and multiple sequence alignments (MSA). Additionally, we introduced a multi-view graph framework based on two types of graph structures: the k-nearest neighbor graph and the adjacency matrix. By alternately employing a Graph Transformer and Graph Convolutional Networks (GCN) to aggregate node information, this framework effectively captures both local and global dependencies of each residue in the predicted structures, thereby significantly enhancing the model's sensitivity to binding sites. This framework further integrates direction, distances and angular information between the 3D coordinates of side-chain atom centroids to construct a relative coordinate system, generating enhanced edge features that ensure the model's equivariance to molecular translations and rotations in space. During training, the Focal Loss function is employed to effectively address the class imbalance in the dataset. Experimental results demonstrate that MVGNN outperforms the current state-of-the-art methods across multiple PPIS benchmark datasets. To further validate the model's generalization capability, we extended MVGNN to the domain of predicting protein-nucleic acid interaction sites, where it also achieved superior performance.

Development and Implementation of Automated Referral Triaging System for Spinal Cord Stimulation Procedure in Pain Medicine.

Effective referral triaging enhances patient service outcomes, experience and access to care especially for specialized procedures. This study presents the development and implementation of an automated triaging system to predict patients who would benefit from Spinal Cord Stimulation (SCS) procedure for their pain management. The proposed triage system aims to improve the triage process by reducing unnecessary appointments before SCS assessment, ensuring appropriate pain management care. It compares various machine learning techniques for the prediction while addressing the class imbalance and overlap challenges inherent in the data. Both data-level and algorithm-level approaches were explored. Two years of patient data was collected including patient characteristics, diagnosis history, pain symptoms, appointment history, medication history, and concepts from clinical notes extracted using Natural Language Processing. EasyEnsemble with Ada Boosting method, an algorithm-level approach, showed the most promising results. The tenfold validation indicated the average area under curve of 0.82, true positive rate (TPR) of 77.3%, and true negative rate (TNR) of 73.0%. The probability threshold was adjusted to 0.575 to meet practice expectation of 15% or less on false positive rate (FPR). The implementation pipeline for the selected model was designed to be applicable to real clinical settings. The one-year implementation results showed TPR of 64.7% and TNR of 87.2%, which reduced FPR by 12.8% while reduced TPR by 12.6%. The trade-off was acceptable to practice. The proposed triage system demonstrated promising accuracy, leading to the enhancement of scheduling systems, patient care, and the reduction of unnecessary appointments in a pain medicine setting.

Optimalisasi Akurasi Algoritma Naïve Bayes Dengan Metode Syntetic Minority Oversampling Technique (Smote) Pada Data Numerik

This research will classify numerical data, namely loan data taken from Kaggle. The data used amounted to 9578 datasets which included data classes with borrowers able to complete credit as many as 8045 records and loans that could not complete credit as many as 1533 records. From the amount of data there is an imbalance of classes so it is necessary to do balancing in order to get more accurate classification results. The purpose of this research is to improve the accuracy of the Naïve Bayes algorithm in classifying numerical data. Fraud in financial transactions is an example of a case of imbalanced data, where the number of legitimate transactions is much greater than those that are fraudulent. Optimizing accuracy in minority (fraud) classes is very important to avoid losses. The method used to improve the accuracy of the algorithm is the Synthetic Minority Oversampling Technique (SMOTE) by over sampling the minority of the dataset. In addition, it also uses the K-Fold Cross Validation method to evaluate the performance of the algorithm process used. Data preprocessing is done to clean the data from missing and invalid values and normalize the data so that all features are on the same scale and suitable for classification analysis. Based on the results of the analysis conducted, before the application of SMOTE the model's ability to recognize minority classes was 16.1%, while after the application of SMOTE the model's ability to recognize minority classes became 48.8%. besides that, before the application of SMOTE the model was able to predict the minority class correctly in 10 cases while after the application of SMOTE, the model was able to predict the minority class correctly in 102 cases. So it can be concluded that the SMOTE technique is able to improve the ability of the model

C MAL: cascaded network-guided class-balanced multi-prototype auxiliary learning for source-free domain adaptive medical image segmentation.

Source-free domain adaptation (SFDA) has become crucial in medical image analysis, enabling the adaptation of source models across diverse datasets without labeled target domain images. Self-training, a popular SFDA approach, iteratively refines self-generated pseudo-labels using unlabeled target domain data to adapt a pre-trained model from the source domain. However, it often faces model instability due to incorrect pseudo-label accumulation and foreground-background class imbalance. This paper presents a pioneering SFDA framework, named cascaded network-guided class-balanced multi-prototype auxiliary learning (C MAL), to enhance model stability. Firstly, we introduce the cascaded translation-segmentation network (CTS-Net), which employs iterative learning between translation and segmentation networks to generate accurate pseudo-labels. The CTS-Net employs a translation network to synthesize target-like images from unreliable predictions of the initial target domain images. The synthesized results refine segmentation network training, ensuring semantic alignment and minimizing visual disparities. Subsequently, reliable pseudo-labels guide the class-balanced multi-prototype auxiliary learning network (CMAL-Net) for effective model adaptation. CMAL-Net incorporates a new multi-prototype auxiliary learning strategy with a memory network to complement source domain data. We propose a class-balanced calibration loss and multi-prototype-guided symmetry cross-entropy loss to tackle class imbalance issue and enhance model adaptability to the target domain. Extensive experiments on four benchmark fundus image datasets validate the superiority of C MAL over state-of-the-art methods, especially in scenarios with significant domain shifts. Our code is available at https://github.com/yxk-art/C2MAL .

Visual impairment prevention by early detection of diabetic retinopathy based on stacked auto-encoder

Diabetic retinopathy (DR) presents a significant concern among diabetic patients, often leading to vision impairment or blindness if left untreated. Traditional diagnosis methods are prone to human error, necessitating accurate alternatives. While various computer-aided systems have been developed to assist in DR detection, there remains a need for accurate and efficient methods to classify its stages. In this study, we propose a novel approach utilizing enhanced stacked auto-encoders for the detection and classification of DR stages. The classification is performed across one healthy (normal) stage and four DR stages: mild, moderate, severe, and proliferative. Unlike traditional CNN approaches, our method offers improved reliability by reducing time complexity, minimizing errors, and enhancing noise reduction. Leveraging a comprehensive dataset from KAGGLE containing 35,126 retinal fundus images representing one healthy (normal) stage and four DR stages, our proposed model demonstrates superior accuracy compared to existing deep learning algorithms. Data augmentation techniques address class imbalance, while SAEs facilitate accurate classification through layer-wise unsupervised pre-training and supervised fine-tuning. We evaluate our model’s performance using rigorous quantitative measures, including accuracy, recall, precision, and F1-score, highlighting its effectiveness in early disease diagnosis and prevention of blindness. Experimental results across different training/testing ratios (50:50, 60:40, 70:30, and 75:25) showcase the model’s robustness. The highest accuracy achieved during training was 93%, while testing accuracy reached 88% on a training/testing ratio of 75:25. Comparative analysis underscores the model’s superiority over existing methods, positioning it as a promising tool for early-stage DR detection and blindness prevention.

Comparison of SVM, Naïve Bayes, and Logistic Regression Algorithms for Sentiment Analysis of Fraud and Bots in Purcashing Concert Ticket

Music concerts are highly anticipated entertainment events, but they are often subject to fraud and the use of bots in online ticket purchases, to the detriment of fans and organizers. Fans may lose confidence in the ticket system and reduce interest in the event. For organizers, it can reduce the event's reputation and finances. This research aims to analyze public sentiment regarding this issue by comparing three classification algorithms: Support Vector Machine (SVM), Naïve Bayes, and Logistic Regression. Data taken from Twitter which contains comments related to fraud and bots. The methods used include data crawling, preprocessing, sentiment labeling, and model evaluation. Preprocessing includes data cleaning, case folding, tokenizing, stopwords, and stemming. Sentiment labeling is done manually or by human annotators. The results showed that SVM had the best accuracy of 91.27%, followed by Logistic Regression (90.03%) and Naïve Bayes (77.70%). Applying SMOTE to overcome class imbalance and improve the performance of negative sentiment models. This research emphasizes the importance of choosing the right algorithm and using SMOTE to improve the accuracy of sentiment analysis regarding fraud and bots in concert ticket purchases. The research results can be applied to improve bot usage detection systems and provide insight for organizers.

Normalization-Guided and Gradient-Weighted Unsupervised Domain Adaptation Network for Transfer Diagnosis of Rolling Bearing Faults Under Class Imbalance

Normalization-Guided and Gradient-Weighted Unsupervised Domain Adaptation Network for Transfer Diagnosis of Rolling Bearing Faults Under Class Imbalance

Comparison of Naïve Bayes, Random Forest, and Logistic Regression Algorithms for Sentiment Analysis Online Gambling

This study aims to compare the performance of Naïve Bayes, Random Forest, and Logistic Regression algorithms for sentiment analysis on the topic of online gambling. The dataset consisted of 4592 entries after preprocessing and applying the SMOTE technique to address class imbalance. The evaluation results show that Random Forest achieved the best performance with an accuracy of 78%, followed by Naïve Bayes and Logistic Regression, both achieving 77%. Random Forest excelled in classifying positive and negative sentiments, while Naïve Bayes demonstrated a significant improvement in recall for neutral sentiment, increasing from 0.45 to 0.82 after the SMOTE application. Logistic Regression showed less optimal performance, particularly for neutral sentiment. This study provides essential guidance for selecting the best algorithms for sentiment analysis in specific domains such as online gambling and highlights the importance of SMOTE in handling imbalanced datasets. The findings of this study can be used by practitioners and policymakers to make more informed decisions in regulating online gambling.

Multiclass Synthetic Accessibility Prediction.

Evaluating synthetic accessibility of in silico molecules is an integral component of the drug discovery process. While the application of machine learning models to predict whether small molecules are easy or hard to synthesize has gained attention recently, predetermined thresholds and data set imbalances present challenges for these binary classification approaches. In this study, we introduce a novel multiclass fold-ensembled classification approach to predict the minimum number of steps needed to synthesize a small molecule. By ensembling the base models trained on multiple stratified subsampled folds, this approach effectively mitigates the impact of class imbalance through probability aggregation or voting aggregation strategies. Additionally, we propose fuzzy evaluation metrics that account for practical tolerances in predictions, providing a more flexible and realistic assessment of model performance. Through experimentation on two reaction benchmark data sets, we demonstrate the effectiveness of our model in a multiclass synthetic accessibility prediction task and the superiority of our proposed method over six existing models in binary synthetic accessibility prediction tasks.

Improving Multiclass Rainfall Prediction with Multilayer Perceptron and SMOTE: Addressing Class Imbalance Challenges

Rainfall is a key climate element that affects weather patterns and human activities, especially in agriculture and daily life. Therefore, accurately classifying rainfall is crucial for predicting future rainfall amounts. This study uses the Multilayer Perceptron (MLP) classification method, a neural network algorithm, to classify rainfall. The dataset, sourced from the BMKG website, has a class imbalance, requiring using the SMOTE (Synthetic Minority Over-sampling Technique) technique. The research compares the performance of MLP with and without SMOTE. The results show that the best model was achieved with SMOTE. MLP without SMOTE achieved an accuracy of 75%, sensitivity of 40.34%, specificity of 86.15%, and an AUC of 63.25%. In comparison, MLP with SMOTE achieved an accuracy of 71.27%, sensitivity of 71.14%, specificity of 90.30%, and an AUC of 80.72%. Although accuracy decreased, the overall evaluation, particularly the AUC, improved significantly. Therefore, the SMOTE technique effectively addresses the class imbalance issue in rainfall classification.

Predicting financial distress in high-dimensional imbalanced datasets: a multi-heterogeneous self-paced ensemble learning framework

Financial distress prediction (FDP) is a critical area of study for researchers, industry stakeholders, and regulatory authorities. However, FDP tasks present several challenges, including high-dimensional datasets, class imbalances, and the complexity of parameter optimization. These issues often hinder the predictive model’s ability to accurately identify companies at high risk of financial distress. To mitigate these challenges, we introduce FinMHSPE—a novel multi-heterogeneous self-paced ensemble (MHSPE) FDP learning framework. The proposed model uses pairwise comparisons of data from multiple time frames combined with the maximum relevance and minimum redundancy method to select an optimal subset of features, effectively resolving the high dimensionality issue. Furthermore, the proposed framework incorporates the MHSPE model to iteratively identify the most informative majority class data samples, effectively addressing the class imbalance issue. To optimize the model’s parameters, we leverage the particle swarm optimization algorithm. The robustness of our proposed model is validated through extensive experiments performed on a financial dataset of Chinese listed companies. The empirical results demonstrate that the proposed model outperforms existing competing models in the field of FDP. Specifically, our FinMHSPE framework achieves the highest performance, achieving an area under the curve (AUC) value of 0.9574, considerably surpassing all existing methods. A comparative analysis of AUC values further reveals that FinMHSPE outperforms state-of-the-art approaches that rely on financial features as inputs. Furthermore, our investigation identifies several valuable features for enhancing FDP model performance, notably those associated with a company’s information and growth potential.

A novel machine learning model for screening the risk of obstructive sleep apnea using craniofacial photography with questionnaires.

Undiagnosed or untreated moderate to severe obstructive sleep apnea (OSA) increases cardiovascular risks and mortality. Early and efficient detection is critical, given its high prevalence. We aimed to develop a practical and efficient approach for obstructive sleep apnea screening, using simple facial photography and sleep questionnaires. We retrospectively included 748 participants who completed polysomnography, sleep questionnaires (STOP-BANG, SBQ), and facial photographs at a university hospital between 2012 and 2023. Owing to class imbalance, we randomly undersampled the participants, categorized into the moderate/severe or no/mild OSA group, based on an apnea-hypopnea index of 15 events/h. Using a validated convolutional neural network, we extracted the OSA probability scores from photographs, which were used as the input for the questionnaires. Four machine learning models were employed to classify the moderate/severe versus no/mild groups and evaluated in the test dataset. We analyzed 426 participants (213 each in the moderate/severe and no/mild groups). The mean (standard deviation) age was 44.6 (14.7) years; 80.8% were men. Logistic regression achieved the highest performance: the area under the receiver operator curve was 97.2%, and accuracy was 91.9%. Adding OSA probability, retrieved from facial photographs, to the questionnaires improved performance, compared with using questionnaires or photographs alone (area under the receiver operator curve: 97.2%, 64% and 79.1% for threshold SBQ 3 and 4, and 85.7%, respectively). Using simple facial photographs and sleep questionnaires, a two-stage approach (convolutional neural network + machine learning) accurately classified OSA into moderate/severe versus no/mild OSA groups. This method may facilitate optimal OSA treatment and avoid unnecessary costly evaluations.

Prediction of outpatient rehabilitation patient preferences and optimization of graded diagnosis and treatment based on XGBoost machine learning algorithm

BackgroundThe Department of Rehabilitation Medicine is key to improving patients’ quality of life. Driven by chronic diseases and an aging population, there is a need to enhance the efficiency and resource allocation of outpatient facilities. This study aims to analyze the treatment preferences of outpatient rehabilitation patients by using data and a grading tool to establish predictive models. The goal is to improve patient visit efficiency and optimize resource allocation through these predictive models.MethodsData were collected from 38 Chinese institutions, including 4,244 patients visiting outpatient rehabilitation clinics. Data processing was conducted using Python software. The pandas library was used for data cleaning and preprocessing, involving 68 categorical and 12 continuous variables. The steps included handling missing values, data normalization, and encoding conversion. The data were divided into 80% training and 20% test sets using the Scikit-learn library to ensure model independence and prevent overfitting. Performance comparisons among XGBoost, random forest, and logistic regression were conducted using metrics, including accuracy and receiver operating characteristic (ROC) curves. The imbalanced learning library’s SMOTE technique was used to address the sample imbalance during model training. The model was optimized using a confusion matrix and feature importance analysis, and partial dependence plots (PDP) were used to analyze the key influencing factors.ResultsXGBoost achieved the highest overall accuracy of 80.21% with high precision and recall in Category 1. random forest showed a similar overall accuracy. Logistic Regression had a significantly lower accuracy, indicating difficulties with nonlinear data. The key influencing factors identified include distance to medical institutions, arrival time, length of hospital stay, and specific diseases, such as cardiovascular, pulmonary, oncological, and orthopedic conditions. The tiered diagnosis and treatment tool effectively helped doctors assess patients’ conditions and recommend suitable medical institutions based on rehabilitation grading.ConclusionThis study confirmed that ensemble learning methods, particularly XGBoost, outperform single models in classification tasks involving complex datasets. Addressing class imbalance and enhancing feature engineering can further improve model performance. Understanding patient preferences and the factors influencing medical institution selection can guide healthcare policies to optimize resource allocation, improve service quality, and enhance patient satisfaction. Tiered diagnosis and treatment tools play a crucial role in helping doctors evaluate patient conditions and make informed recommendations for appropriate medical care.

Ensemble-based customer churn prediction in banking: a voting classifier approach for improved client retention using demographic and behavioral data

Customer turnover is a crucial issue in banking since maintained profitability depends on keeping clients. This work aims to categorize consumer turnover in banks by using a new ensemble approach combining many machine learning methods, hence enhancing churn prediction models. Using a comprehensive dataset including demographic, financial, and behavioral data—such as credit score, account balance, tenure, and activity levels—the study employs the goal variable revealing if a customer has left the bank. The study starts with univariate, bivariate, and multivariate feature exploration and subsequently uses the Interquartile Range (IQR) approach to identify outliers thereby improving the data quality. Five models—K-Nearest Neighbors, Support Vector Classifier, Random Forest, Decision Tree, and XGBoost—a Voting Classifier ensemble—are used to estimate project churn. Building upon all the strengths of each model, this approach improves the prediction of classification and provides a balanced and highly robust classification system. The applied approaches are K-Nearest Neighbors (KNN), Support Vector Classifier (SVC), Random Forest, Decision Tree, and XGBoost within a Voting Classifier configuration. The performance of the Voting Classifier without SMOTE yields the following results: Accuracy: 0.87, precision: 0.87, recall: 0.80, and F1-Score: 0.87. The proposed model that extend the base model using SMOTE (Synthetic Minority Over-sampling Technique), yields a higher prediction accuracy of 0.90, precision of 0.90, recall of 0.90 and F1-Score of 0.90. This enhancement is proving the efficiency of SMOTE to handle the class imbalance problem in order to render the churn prediction more balanced and reliable system. The proposed approach assures a reliable solution to the strategies to retain the customers in the banking organisations.

Comparison of Resampling Methods and Radiomic Machine Learning Classifiers for Predicting Bone Quality Using Dual-Energy X-Ray Absorptiometry.

Background/Objectives: This study presents a novel approach, based on a combination of radiomic feature extraction, data resampling techniques, and machine learning algorithms, for the detection of degraded bone structures in Dual X-ray Absorptiometry (DXA) images. This comprehensive approach, which addresses the critical aspects of the problem, distinguishes this work from previous studies, improving the performance achieved by the most similar studies. The primary aim is to provide clinicians with an accessible tool for quality bone assessment, which is currently limited. Methods: A dataset of 1531 spine DXA images was automatically segmented and labelled based on Trabecular Bone Score (TBS) values. Radiomic features were extracted using Pyradiomics, and various resampling techniques were employed to address class imbalance. Three machine learning classifiers (Logistic Regression, Support Vector Machine (SVM), and XGBoost) were trained and evaluated using standard performance metrics. Results: The SVM classifier outperformed the other classifiers. The highest F-score of 97.5% was achieved using the Grey Level Dependence Matrix and Grey Level Run Length Matrix feature combination with SMOTEENN resampling, which proved to be the most effective resampling technique, while the undersampling method yielded the lowest performance. Conclusions: This research demonstrates the potential of radiomic texture features, resampling techniques, and machine learning methods for classifying DXA images into healthy or degraded bone structures, which potentially leads to improved clinical diagnosis and treatment.

A Data-Driven Deep Learning Framework for Prediction of Traffic Crashes at Road Intersections

Traffic crash prediction (TCP) is a fundamental problem for intelligent transportation systems in smart cities. Improving the accuracy of traffic crash prediction is important for road safety and effective traffic management. Owing to recent advances in artificial neural networks, several new deep-learning models have been proposed for TCP. However, these works mainly focus on accidents in regions, which are typically pre-determined using a grid map. We argue that TCP for roads, especially for crashes at or near road intersections which account for more than 50% of the fatal or injury crashes based on the Federal Highway Administration, has a significant practical and research value and thus deserves more research. In this paper, we formulate TCP at Road Intersections as a classification problem and propose a three-phase data-driven deep learning model, called Road Intersection Traffic Crash Prediction (RoadInTCP), to predict traffic crashes at intersections by exploiting publicly available heterogeneous big data. In Phase I we extract discriminative latent features called topological-relational features (tr-features), of intersections using a neural network model by exploiting topological information of the road network and various relationships amongst nearby intersections. In Phase II, in addition to tr-features which capture some inherent properties of the road network, we also explore additional thematic information in terms of environmental, traffic, weather, risk, and calendar features associated with intersections. In order to incorporate the potential correlation in nearby intersections, we utilize a Graph Convolution Network (GCN) to aggregate features from neighboring intersections based on a message-passing paradigm for TCP. While Phase II serves well as a TCP model, we further explore the signals embedded in the sequential feature changes over time for TCP in Phase III, by exploring RNN or 1DCNN which have known success on sequential data. Additionally, to address the serious issues of imbalanced classes in TCP and large-scale heterogeneous big data, we propose an effective data sampling approach in data preparation to facilitate model training. We evaluate the proposed RoadInTCP model via extensive experiments on a real-world New York City traffic dataset. The experimental results show that the proposed RoadInTCP robustly outperforms existing methods.