Dealing with class imbalance in classifier chains via random undersampling

CorLabelNet: a comprehensive framework for multi-label chest X-ray image classification with correlation guided discriminant feature learning and oversampling

Recent advancements in deep learning techniques have significantly improved multi-label chest X-ray (CXR) image classification for clinical diagnosis. However, most previous studies neither effectively learn label correlations nor take full advantage of them to improve multi-label classification performance. In addition, different labels of CXR images are usually severely imbalanced, resulting in the model exhibiting a bias towards the majority class. To address these challenges, we introduce a framework that not only learns label correlations but also utilizes them to guide the learning of features and the process of oversampling. In this paper, our approach incorporates self-attention to capture high-order label correlations and considers label correlations from both global and local perspectives. Then, we propose a consistency constraint and a multi-label contrastive loss to enhance feature learning. To alleviate the imbalance issue, we further propose an oversampling approach that exploits the learned label correlation to identify crucial seed samples for oversampling. Our approach repeats 5-fold cross-validation process experiments three times and achieves the best performance on both the CheXpert and ChestX-Ray14 datasets. Learning accurate label correlation is significant for multi-label classification and taking full advantage of label correlations is beneficial for discriminative feature learning and oversampling. A comparative analysis with the state-of-the-art approaches highlights the effectiveness of our proposed methods.

Compound Fault Diagnosis of Wind Turbine Gearbox via Modified Signal Quality Coefficient and Versatile Residual Shrinkage Network

Wind turbine gearbox fault diagnosis is critical to guarantee working efficiency and operational safety. However, the current diagnostic methods face enormous restrictions in handling nonlinear noise signals and intricate compound fault patterns. Herein, a compound fault diagnosis method based on modified signal quality coefficient (MSQC) and versatile residual shrinkage network (VRSN) is proposed to resolve these issues. In detail, the MSQC is designed to remove the noise components irrelevant to wind turbine operation status, and it has the ability to balance the denoised effect and signal fidelity. The VRSN is constructed for compound fault diagnosis, and it consists of two heterogeneous residual shrinkage networks. The former is designed to count the number of faults, and the latter is adopted to identify the single or compound fault pattern. Finally, a self-built wind turbine gearbox compound fault test rig is adopted to verify the proposed method's effectiveness. The results demonstrate that the proposed method is competitive in terms of compound fault diagnosis accuracy.

Prediction of stroke-associated hospital-acquired pneumonia: Machine learning approach

BACKGROUND

Stroke-associated Hospital Acquired Pneumonia (HAP) significantly impacts patient outcomes. This study explores the utility of machine learning models in predicting HAP in stroke patients, leveraging national registry data and SHapley Additive exPlanations (SHAP) analysis to identify key predictive factors.

METHODS

We collected data from a national stroke registry covering January 2014 to July 2022, including 9,840 patients diagnosed with ischemic and hemorrhagic strokes. Five machine learning models were trained and evaluated: XGBoost, Random Forest, Support Vector Machine (SVM), Logistic Regression, and Artificial Neural Network (ANN). Performance was assessed using accuracy, precision, recall, F1-score, AUC, log loss, and Brier score. SHAP analysis was conducted to interpret model outputs.

RESULTS

The ANN model demonstrated superior performance, with an F1-score of 0.86 and an AUC of 0.94. SHAP analysis identified key predictors: stroke severity, admission location, Glasgow Coma score (GCS), systolic and diastolic blood pressure at admission, ethnicity, stroke type, mode of arrival, and age. Patients with higher stroke severity, dysphagia, and those arriving by ambulance were at increased risk for HAP.

CONCLUSION

This study enhances our understanding of early predictive factors for HAP in stroke patients and underlines the potential of machine learning to improve clinical decision-making and personalized care.

Machine learning-based prediction of one-year mortality in ischemic stroke patients

BACKGROUND

Accurate prediction of mortality following an ischemic stroke is essential for tailoring personalized treatment strategies. This study evaluates the effectiveness of machine learning models in predicting one-year mortality after an ischemic stroke.

METHODS

Five machine learning models were trained using data from a national stroke registry, with logistic regression demonstrating the highest performance. The SHapley Additive exPlanations (SHAP) analysis explained the model's outcomes and defined the influential predictive factors.

RESULTS

Analyzing 8183 ischemic stroke patients, logistic regression achieved 83% accuracy, 0.89 AUC, and an F1 score of 0.83. Significant predictors included stroke severity, pre-stroke functional status, age, hospital-acquired pneumonia, ischemic stroke subtype, tobacco use, and co-existing diabetes mellitus (DM).

DISCUSSION

The model highlights the importance of predicting mortality in enhancing personalized stroke care. Apart from pneumonia, all predictors can serve the early prediction of mortality risk which supports the initiation of early preventive measures and in setting realistic expectations of disease outcomes for all stakeholders. The identified tobacco paradox warrants further investigation.

CONCLUSION

This study offers a promising tool for early prediction of stroke mortality and for advancing personalized stroke care. It emphasizes the need for prospective studies to validate these findings in diverse clinical settings.

Prediction of hypertension risk based on multiple feature fusion

OBJECTIVE

In the application of machine learning to the prediction of hypertension, many factors have seriously affected the classification accuracy and generalization performance. We propose a pulse wave classification model based on multi-feature fusion for accuracy prediction of hypertension.

METHODS AND MATERIALS

We propose an ensemble under-sampling model with dynamic weights to decrease the influence of class imbalance on classification, further to automatically classify of hypertension on inquiry diagnosis. We also build a deep learning model based on hybrid attention mechanism, which transforms pulse waves to feature maps for extraction of in-depth features, so as to automatically classify hypertension on pulse diagnosis. We build the multi-feature fusion model based on dynamic Dempster/Shafer (DS) theory combining inquiry diagnosis and pulse diagnosis to enhance fault tolerance of prediction for multiple classifiers. In addition, this study calculates feature importance ranking of scale features on inquiry diagnosis and temporal and frequency-domain features on pulse diagnosis.

RESULTS

The accuracy, sensitivity, specificity, F1-score and G-mean after 5-fold cross-validation were 94.08%, 93.43%, 96.86%, 93.45% and 95.12%, respectively, based on the hypertensive samples of 409 cases from Longhua Hospital affiliated to Shanghai University of Traditional Chinese Medicine and Hospital of Integrated Traditional Chinese and Western Medicine. We find the key factors influencing hypertensive classification accuracy, so as to assist in the prevention and clinical diagnosis of hypertension.

CONCLUSION

Compared with the state-of-the-art models, the multi-feature fusion model effectively utilizes the patients' correlated multimodal features, and has higher classification accuracy and generalization performance.

Machine learning-based prognostication of mortality in stroke patients

OBJECTIVES

Predicting stroke mortality is crucial for personalized care. This study aims to design and evaluate a machine learning model to predict one-year mortality after a stroke.

MATERIALS AND METHODS

Data from the National Multiethnic Stroke Registry was utilized. Eight machine learning (ML) models were trained and evaluated using various metrics. SHapley Additive exPlanations (SHAP) analysis was used to identify the influential predictors.

RESULTS

The final analysis included 9840 patients diagnosed with stroke were included in the study. The XGBoost algorithm exhibited optimal performance with high accuracy (94.5%) and AUC (87.3%). Core predictors encompassed National Institutes of Health Stroke Scale (NIHSS) at admission, age, hospital length of stay, mode of arrival, heart rate, and blood pressure. Increased NIHSS, age, and longer stay correlated with higher mortality. Ambulance arrival and lower diastolic blood pressure and lower body mass index predicted poorer outcomes.

CONCLUSIONS

This model's predictive capacity emphasizes the significance of NIHSS, age, hospital stay, arrival mode, heart rate, blood pressure, and BMI in stroke mortality prediction. Specific findings suggest avenues for data quality enhancement, registry expansion, and real-world validation. The study underscores machine learning's potential for early mortality prediction, improving risk assessment, and personalized care. The potential transformation of care delivery through robust ML predictive tools for Stroke outcomes could revolutionize patient care, allowing for personalized plans and improved preventive strategies for stroke patients. However, it is imperative to conduct prospective validation to evaluate its practical clinical effectiveness and ensure its successful adoption across various healthcare environments.

Label correlation guided discriminative label feature learning for multi-label chest image classification

BACKGROUND AND OBJECTIVE

Multi-label Chest X-ray (CXR) images often contain rich label relationship information, which is beneficial to improve classification performance. However, because of the intricate relationships among labels, most existing works fail to effectively learn and make full use of the label correlations, resulting in limited classification performance. In this study, we propose a multi-label learning framework that learns and leverages the label correlations to improve multi-label CXR image classification.

METHODS

In this paper, we capture the global label correlations through the self-attention mechanism. Meanwhile, to better utilize label correlations for guiding feature learning, we decompose the image-level features into label-level features. Furthermore, we enhance label-level feature learning in an end-to-end manner by a consistency constraint between global and local label correlations, and a label correlation guided multi-label supervised contrastive loss.

RESULTS

To demonstrate the superior performance of our proposed approach, we conduct three times 5-fold cross-validation experiments on the CheXpert dataset. Our approach obtains an average F1 score of 44.6% and an AUC of 76.5%, achieving a 7.7% and 1.3% improvement compared to the state-of-the-art results.

CONCLUSION

More accurate label correlations and full utilization of the learned label correlations help learn more discriminative label-level features. Experimental results demonstrate that our approach achieves exceptionally competitive performance compared to the state-of-the-art algorithms.

Decoding auditory-evoked response in affective states using wearable around-ear EEG system

Objective.In this paper, an around-ear EEG system is investigated as an alternative methodology to conventional scalp-EEG-based systems in classifying human affective states in the arousal-valence domain evoked in response to auditory stimuli.Approach.EEG recorded from around the ears is compared to EEG collected according to the international 10-20 system in terms of efficacy in an affective state classification task. A wearable device with eight dry EEG channels is designed for ear-EEG acquisition in this study. Twenty-one subjects participated in an experiment consisting of six sessions over three days using both ear and scalp-EEG acquisition methods. Experimental tasks consisted of listening to an auditory stimulus and self-reporting the elicited emotion in response to the said stimulus. Various features were used in tandem with asymmetry methods to evaluate binary classification performances of arousal and valence states using ear-EEG signals in comparison to scalp-EEG.Main results.We achieve an average accuracy of 67.09% ± 6.14 for arousal and 66.61% ± 6.14 for valence after training a multi-layer extreme learning machine with ear-EEG signals in a subject-dependent context in comparison to scalp-EEG approach which achieves an average accuracy of 68.59% ± 6.26 for arousal and 67.10% ± 4.99 for valence. In a subject-independent context, the ear-EEG approach achieves 63.74% ± 3.84 for arousal and 64.32% ± 6.38 for valence while the scalp-EEG approach achieves 64.67% ± 6.91 for arousal and 64.86% ± 5.95 for valence. The best results show no significant differences between ear-EEG and scalp-EEG signals for classifications of affective states.Significance.To the best of our knowledge, this paper is the first work to explore the use of around-ear EEG signals in emotion monitoring. Our results demonstrate the potential use of around-ear EEG systems for the development of emotional monitoring setups that are more suitable for use in daily affective life log systems compared to conventional scalp-EEG setups.

Parkinson's disease resting tremor severity classification using machine learning with resampling techniques

In resting tremor, the body part is in complete repose and often dampens or subsides entirely with action. The most frequent cause of resting tremors is known as idiopathic Parkinson's disease (PD). For examination, neurologists of patients with PD include tests such as finger-to-nose tests, walking back and forth in the corridor, and the pull test. This evaluation is focused on Unified Parkinson's disease rating scale (UPDRS), which is subjective as well as based on some daily life motor activities for a limited time frame. In this study, severity analysis is performed on an imbalanced dataset of patients with PD. This is the reason why the classification of various data containing imbalanced class distribution has endured a notable drawback of the performance achievable by various standard classification learning algorithms. In this work, we used resampling techniques including under-sampling, over-sampling, and a hybrid combination. Resampling techniques are incorporated with renowned classifiers, such as XGBoost, decision tree, and K-nearest neighbors. From the results, it is concluded that the Over-sampling method performed much better than under-sampling and hybrid sampling techniques. Among the over-sampling techniques, random sampling has obtained 99% accuracy using XGBoost classifier and 98% accuracy using the decision tree. Besides, it is observed that different resampling methods performed differently with various classifiers.

Machine learning: its challenges and opportunities in plant system biology

Sequencing technologies are evolving at a rapid pace, enabling the generation of massive amounts of data in multiple dimensions (e.g., genomics, epigenomics, transcriptomic, metabolomics, proteomics, and single-cell omics) in plants. To provide comprehensive insights into the complexity of plant biological systems, it is important to integrate different omics datasets. Although recent advances in computational analytical pipelines have enabled efficient and high-quality exploration and exploitation of single omics data, the integration of multidimensional, heterogenous, and large datasets (i.e., multi-omics) remains a challenge. In this regard, machine learning (ML) offers promising approaches to integrate large datasets and to recognize fine-grained patterns and relationships. Nevertheless, they require rigorous optimizations to process multi-omics-derived datasets. In this review, we discuss the main concepts of machine learning as well as the key challenges and solutions related to the big data derived from plant system biology. We also provide in-depth insight into the principles of data integration using ML, as well as challenges and opportunities in different contexts including multi-omics, single-cell omics, protein function, and protein-protein interaction. KEY POINTS: • The key challenges and solutions related to the big data derived from plant system biology have been highlighted. • Different methods of data integration have been discussed. • Challenges and opportunities of the application of machine learning in plant system biology have been highlighted and discussed.

RDPVR: Random Data Partitioning with Voting Rule for Machine Learning from Class-Imbalanced Datasets

Since most classifiers are biased toward the dominant class, class imbalance is a challenging problem in machine learning. The most popular approaches to solving this problem include oversampling minority examples and undersampling majority examples. Oversampling may increase the probability of overfitting, whereas undersampling eliminates examples that may be crucial to the learning process. We present a linear time resampling method based on random data partitioning and a majority voting rule to address both concerns, where an imbalanced dataset is partitioned into a number of small subdatasets, each of which must be class balanced. After that, a specific classifier is trained for each subdataset, and the final classification result is established by applying the majority voting rule to the results of all of the trained models. We compared the performance of the proposed method to some of the most well-known oversampling and undersampling methods, employing a range of classifiers, on 33 benchmark machine learning class-imbalanced datasets. The classification results produced by the classifiers employed on the generated data by the proposed method were comparable to most of the resampling methods tested, with the exception of SMOTEFUNA, which is an oversampling method that increases the probability of overfitting. The proposed method produced results that were comparable to the Easy Ensemble (EE) undersampling method. As a result, for solving the challenge of machine learning from class-imbalanced datasets, we advocate using either EE or our method.

Quantitative Prediction of Microsatellite Instability in Colorectal Cancer With Preoperative PET/CT-Based Radiomics

OBJECTIVES

Microsatellite instability (MSI) status is an important hallmark for prognosis prediction and treatment recommendation of colorectal cancer (CRC). To address issues due to the invasiveness of clinical preoperative evaluation of microsatellite status, we investigated the value of preoperative 18F-FDG PET/CT radiomics with machine learning for predicting the microsatellite status of colorectal cancer patients.

METHODS

A total of 173 patients that underwent 18F-FDG PET/CT scans before operations were retrospectively analyzed in this study. The microsatellite status for each patient was identified as microsatellite instability-high (MSI-H) or microsatellite stable (MSS), according to the test for mismatch repair gene proteins with immunohistochemical staining methods. There were 2,492 radiomic features in total extracted from 18F-FDG PET/CT imaging. Then, radiomic features were selected through multivariate random forest selection and univariate relevancy tests after handling the imbalanced dataset through the random under-sampling method. Based on the selected features, we constructed a BalancedBagging model based on Adaboost classifiers to identify the MSI status in patients with CRC. The model performance was evaluated by the area under the curve (AUC), sensitivity, specificity, and accuracy on the validation dataset.

RESULTS

The ensemble model was constructed based on two radiomic features and achieved an 82.8% AUC for predicting the MSI status of colorectal cancer patients. The sensitivity, specificity, and accuracy were 83.3, 76.3, and 76.8%, respectively. The significant correlation of the selected two radiomic features with multiple effective clinical features was identified (p < 0.05).

CONCLUSION

18F-FDG PET/CT radiomics analysis with the machine learning model provided a quantitative, efficient, and non-invasive mechanism for identifying the microsatellite status of colorectal cancer patients, which optimized the treatment decision support.

An Improved Hybrid Approach for Handling Class Imbalance Problem

Class imbalance issue that presents in many real-world datasets exhibit favouritism toward the majority class and showcases poor performance for the minority class. Such misclassifications may incur dubious outcome in case of disease diagnosis and other critical applications. Hence, it is a hot topic for the researchers to tackle the class imbalance issue. We present a novel hybrid approach for handling such datasets. We utilize simulated annealing algorithm for undersampling and apply support vector machine, decision tree, k-nearest neighbor and discriminant analysis for the classification task. We validate our technique in 51 real-world datasets and compare it with other recent works. Our technique yields better efficacy than the existing techniques and hence it can be applied in imbalance datasets to mitigate the misclassification.

SMOTE-Based Weighted Deep Rotation Forest for the Imbalanced Hyperspectral Data Classification

Conventional classification algorithms have shown great success in balanced hyperspectral data classification. However, the imbalanced class distribution is a fundamental problem of hyperspectral data, and it is regarded as one of the great challenges in classification tasks. To solve this problem, a non-ANN based deep learning, namely SMOTE-Based Weighted Deep Rotation Forest (SMOTE-WDRoF) is proposed in this paper. First, the neighboring pixels of instances are introduced as the spatial information and balanced datasets are created by using the SMOTE algorithm. Second, these datasets are fed into the WDRoF model that consists of the rotation forest and the multi-level cascaded random forests. Specifically, the rotation forest is used to generate rotation feature vectors, which are input into the subsequent cascade forest. Furthermore, the output probability of each level and the original data are stacked as the dataset of the next level. And the sample weights are automatically adjusted according to the dynamic weight function constructed by the classification results of each level. Compared with the traditional deep learning approaches, the proposed method consumes much less training time. The experimental results on four public hyperspectral data demonstrate that the proposed method can get better performance than support vector machine, random forest, rotation forest, SMOTE combined rotation forest, convolutional neural network, and rotation-based deep forest in multiclass imbalance learning.