Information Feature Selection Research Articles

Fusing entropy measures for dynamic feature selection in incomplete approximation spaces

The selection of informative and discriminative features from incomplete dynamic data has emerged as an essential problem because the data preparation in modern applications is a dynamic updating process, and the collected data are often fraught with missing or unobserved values. Using the effectiveness of entropy measures to quantify uncertainty information in incomplete approximation spaces under the framework of tolerance rough sets, a novel incremental feature selection approach is presented in this study from an information-theoretic perspective. Based on the updating patterns of tolerance classification and decision partition induced by conditional and decision features, a novel fused representation of entropy measures in an incomplete approximation space is proposed to accelerate the calculation of feature significance during the heuristic search process. A computationally efficient feature selection algorithm is developed by integrating the incremental fusing strategy of entropy when the data increases dynamically in size. Numerical experiments were performed on several benchmark datasets with data updating for different data sizes and incremental ratios to demonstrate the effectiveness of our method. The superiority of the proposed method is established extensively in terms of computational efficiency, cardinality and accuracy of the selected feature subset.

Pelvic Injury Discriminative Model Based on Data Mining Algorithm.

To reduce the dimension of characteristic information extracted from pelvic CT images by using principal component analysis (PCA) and partial least squares (PLS) methods. To establish a support vector machine (SVM) classification and identification model to identify if there is pelvic injury by the reduced dimension data and evaluate the feasibility of its application. Eighty percent of 146 normal and injured pelvic CT images were randomly selected as training set for model fitting, and the remaining 20% was used as testing set to verify the accuracy of the test, respectively. Through CT image input, preprocessing, feature extraction, feature information dimension reduction, feature selection, parameter selection, model establishment and model comparison, a discriminative model of pelvic injury was established. The PLS dimension reduction method was better than the PCA method and the SVM model was better than the naive Bayesian classifier (NBC) model. The accuracy of the modeling set, leave-one-out cross validation and testing set of the SVM classification model based on 12 PLS factors was 100%, 100% and 93.33%, respectively. In the evaluation of pelvic injury, the pelvic injury data mining model based on CT images reaches high accuracy, which lays a foundation for automatic and rapid identification of pelvic injuries.

Machine learning approach informs biology of cancer drug response

BackgroundThe mechanism of action for most cancer drugs is not clear. Large-scale pharmacogenomic cancer cell line datasets offer a rich resource to obtain this knowledge. Here, we present an analysis strategy for revealing biological pathways that contribute to drug response using publicly available pharmacogenomic cancer cell line datasets.MethodsWe present a custom machine-learning based approach for identifying biological pathways involved in cancer drug response. We test the utility of our approach with a pan-cancer analysis of ML210, an inhibitor of GPX4, and a melanoma-focused analysis of inhibitors of BRAFV600. We apply our approach to reveal determinants of drug resistance to microtubule inhibitors.ResultsOur method implicated lipid metabolism and Rac1/cytoskeleton signaling in the context of ML210 and BRAF inhibitor response, respectively. These findings are consistent with current knowledge of how these drugs work. For microtubule inhibitors, our approach implicated Notch and Akt signaling as pathways that associated with response.ConclusionsOur results demonstrate the utility of combining informed feature selection and machine learning algorithms in understanding cancer drug response.

Prediction of NOx Concentration at SCR Inlet Based on BMIFS-LSTM

As the main energy source for thermal power generation, coal generates a large amount of NOx during its incineration in boilers, and excessive NOx emissions can cause serious pollution to the air environment. Selective catalytic reduction denitrification (SCR) selects the optimal amount of ammonia to be injected for denitrification based on the measurement of NOx concentration by the automatic flue gas monitoring system. Since the automatic flue gas monitoring system has a large delay in measurement, it cannot accurately reflect the real-time changes of NOx concentration at the SCR inlet when the unit load fluctuates, leading to problems such as ammonia escape and NOx emission exceeding the standard. In response to these problems, this paper proposes an SCR inlet NOx concentration prediction algorithm based on BMIFS-LSTM. An improved mutual information feature selection algorithm (BMIFS) is used to filter out the auxiliary variables with maximum correlation and minimum redundancy with NOx concentration, and reduce the coupling and dimensionality among the variables in the data set. The dominant and auxiliary variables are then fed together into a long short-term memory neural network (LSTM) to build a prognostic model. Simulation experiments are conducted using historical operation data of a 300 MW thermal power unit. The experimental results show that the algorithm in this paper reduces the average relative error by 3.45% and the root mean square error by 1.50 compared with the algorithm without auxiliary variable extraction, which can accurately reflect the real-time changes of NOx concentration at the SCR inlet, solve the problem of delay in NOx concentration measurement, and reduce the occurrence of atmospheric pollution caused by excessive NOx emissions.

On the Black-Box Challenge for Fraud Detection Using Machine Learning (I): Linear Models and Informative Feature Selection

Artificial intelligence (AI) is rapidly shaping the global financial market and its services due to the great competence that it has shown for analysis and modeling in many disciplines. What is especially remarkable is the potential that these techniques could offer to the challenging reality of credit fraud detection (CFD); but it is not easy, even for financial institutions, to keep in strict compliance with non-discriminatory and data protection regulations while extracting all the potential that these powerful new tools can provide to them. This reality effectively restricts nearly all possible AI applications to simple and easy to trace neural networks, preventing more advanced and modern techniques from being applied. The aim of this work was to create a reliable, unbiased, and interpretable methodology to automatically evaluate CFD risk. Therefore, we propose a novel methodology to address the mentioned complexity when applying machine learning (ML) to the CFD problem that uses state-of-the-art algorithms capable of quantifying the information of the variables and their relationships. This approach offers a new form of interpretability to cope with this multifaceted situation. Applied first is a recent published feature selection technique, the informative variable identifier (IVI), which is capable of distinguishing among informative, redundant, and noisy variables. Second, a set of innovative recurrent filters defined in this work are applied, which aim to minimize the training-data bias, namely, the recurrent feature filter (RFF) and the maximally-informative feature filter (MIFF). Finally, the output is classified by using compelling ML techniques, such as gradient boosting, support vector machine, linear discriminant analysis, and linear regression. These defined models were applied both to a synthetic database, for better descriptive modeling and fine tuning, and then to a real database. Our results confirm that our proposal yields valuable interpretability by identifying the informative features’ weights that link original variables with final objectives. Informative features were living beyond one’s means, lack or absence of a transaction trail, and unexpected overdrafts, which are consistent with other published works. Furthermore, we obtained 76% accuracy in CFD, which represents an improvement of more than 4% in the real databases compared to other published works. We conclude that with the use of the presented methodology, we do not only reduce dimensionality, but also improve the accuracy, and trace relationships among input and output features, bringing transparency to the ML reasoning process. The results obtained here were used as a starting point for the companion paper which reports on our extending the interpretability to nonlinear ML architectures.

An Unsupervised Mutual Information Feature Selection Method Based on SVM for Main Transformer Condition Diagnosis in Nuclear Power Plants

Dissolved gas in oil (DGA) is a common means of monitoring the condition of an oil-immersed transformer. The concentration of dissolved gas and the ratio of different gases are important indexes to judge the condition of power transformers. Monitoring devices for dissolved gas in oil are widely installed in main transformers, but there are few recorded fault data of main transformers. The special operation and maintenance modes of main transformers leads to the fault modes particularity of main transformers. In order to solve the problem of insufficient samples and the feature uncertainty, this paper puts forward an unsupervised mutual information method to select the feature verified by the optimized support vector machine (SVM) model of particle swarm optimization (PSO) method and tries to find the feature sequence with better performance. The methos is validated by data from nuclear power transformers.

Application of multidomain sensor image fusion and training data augmentation for enhanced CNN image classification

Convolutional neural networks (CNNs) provide the sensing and detection community with a discriminative approach for classifying images. However, one of the largest limitations of deep CNN image classifiers is the need for extensive training datasets containing a variety of image representations. While current methods, such as generative adversarial network data augmentation, additions of noise, rotations, and translations, can allow CNNs to better associate new images and their feature representations to ones of a learned image class, many fail to provide new contexts of ground truth feature information. To expand the association of critical class features within CNN image training datasets, an image pairing and training dataset augmentation paradigm via a multi-sensor domain image data fusion algorithm is proposed. This algorithm uses a mutual information (MI) and merit-based feature selection subroutine to pair highly correlated cross-domain images from multiple sensor domain image datasets. It then re-augments the corresponding cross-domain image pairs into the opposite sensor domain’s feature set via a highest MI, cross sensor domain, and image concatenation function. This augmented image set then acts to retrain the CNN to recognize greater generalizations of image class features via cross domain, mixed representations. Experimental results indicated an increased ability of CNNs to generalize and discriminate between image classes during testing of class images from synthetic aperture radar vehicle, solar cell device reliability screening, and lung cancer detection image datasets.

Fuzzy joint mutual information feature selection based on ideal vector

Feature selection is an effective preprocess that helps the classification task to improve its performance. Information measures, especially fuzzy information measures have been a powerful solution in feature selection. However, existing feature selection frameworks based on fuzzy information measures suffer from one of two limitations: (1) discarding important feature relations as redundancy and complementarity. (2) requiring a high computational cost. To overcome these limitations, a new feature selection framework is proposed in this paper, called fuzzy joint mutual information based on ideal vector (FJMIIV), to extract all possible feature relations with minimum computational cost. The performance of the proposed framework is compared with some of the state-of-the-art frameworks on various benchmark datasets. The comparison results using four well-known classifiers show that the proposed method can effectively improve the classification performance.

Analysis of ultrasound images of chronic diffuse liver diseases using texture parameters of gray gradations

The paper proposes a technology for analyzing ultrasound images based on textural characteristics for the development of a medical decision-making system for determining the presence of pathological changes in diffuse liver diseases. The list of algorithmic solutions known and proposed by the authors for the formation and selection of informative features was applied, a version of the Random Forest classification algorithm was developed, which achieved (compared with known analogues) the highest quality indicators of the classification of normal liver pathology. Ensuring the "explainability" of the computer diagnostic solution for radiologists is achieved by accompanying the diagnostic conclusion with an image indicator, which is a binarized transformation of the primary ultrasound image, where the binary colors are associated with the corresponding norm-pathology class. The transformation is a kind of "amplifier" of the effect of the presence of pathology, which allows radiologists to consciously agree with the conclusion of artificial intelligence or question its diagnosis and recommend additional research. The system was handed over to the co-authors of the work, specialists of the Institute of Nuclear Medicine and Radiodiagnosis of the National Academy of Medical Sciences of Ukraine, for trial operation.

О ФОРМИРОВАНИИ БАЗ ПРЕЦЕДЕНТОВ ДЛЯ РЕШЕНИЯ ЗАДАЧ ИНФОРМАЦИОННОЙ БЕЗОПАСНОСТИ

The issues of the formation of bases of precedents related to the one of “big data” in the field of information security are considered. That manifests itself in the lack of opportunities for using classical methods of the data analysis in order to make informed decisions. As a rule, among the basic requirements for the mathematical and software information security include minimizing the cost of computing resources and high rates of generalizing ability based on the results of machine learning. To record those requirements, it is proposed to use several criteria. The application of criteria is associated with the formation of a latent feature space with a smaller dimension than the original one and the study of the structure of object relations in it. The analysis of the structure diversity of objects relations is carried out in the spaces formed by pairs of classes (type of DDOS attacks, normal traffic). A rule is proposed for including an object in the database of precedents. The effectiveness of using the precedent base is demonstrated (based on a test sample) using the k nearest neighbor recognition algorithms, Random Forest, and SVM. One of the types of DDOS attacks of the test sample is not presented in the training. Descriptions of objects are given without selection and with selection of informative features, according to the results of data normalization and without it.

НЕЙРОФІЗІОЛОГІЧНІ МЕХАНІЗМИ ЗДІЙСНЕННЯ ПРОСТОЇ СЕНСОМОТОРНОЇ РЕАКЦІЇ У ВІЙСЬКОВОСЛУЖБОВЦІВ РІЗНОГО ВІКУ

Introduction.For different types of professional activity, the speed of information processing is one of the most important indicators of performance. It is well known that aging is accompanied by changes in the speed and / or mode of information processing in the brain. Simple sensorimotor reaction (SMR) shows speed and efficiency of information perception processes, its integration and is and reveals how well the basic cognitive mechanisms can be involved in the implementation of target tasks. Purpose.The aim of this work was to investigate the rate of SMR in different age groups and to identify age-related changes in neural networks that promote task performance.Methods.The survey involved 34 right-handed volunteers, aged 18-54, without health complaints - representatives of various military professions (mechanics, shooters, drivers, radar operators, anti-aircraft missile officers, communications, tablets, pilots and conscripts), which were divided into three age groups (1 group - 18-23 years, n = 11; 2 group - 24-34 years, n = 15; 3 group - 35- 54 years, n = 8). SMR was determined using a specialized computer program and while its performance 19 electrode EEG was recorded. Remote synchronization of brain areas was determined using coherent analysis in Neuron-Spectrum-4 / VP complex («NeuroSoft»). Reliable synchronization was determined only for those pairs of leads for which the median was ≥ 0.7.Results.Although the comparative analysis of the PSMR rate in the three groups according to the Kruskal-Wallis test did not reveal significant differences (Kruskal-Wallis test: H (2, N = 54) = 0.889; p = 0.64), coherent analysis of electroencephalograms, which were recorded during task, revealed that there were differences in the neural mechanisms of SSMR between the three age groups. We showed age-related increases in activation of the left secondary visual cortex, while activation of the left frontal lobe decreasedwhich may indicate shift from the strategic system of response to the system of informative features selection and constant attention maintenance. In the third group coherence in the theta range between the prefrontal area and the postcentral gyrus was the strongest in comparison with the first and second groups, which may indicate an increase in the role of visual selective attention with age, as well as greater effort and concentration to perform the task. Also there was an age-related increase in the alpha range coherences and decrease in the delta range, which may indicate shift from the general setting to the input information flow to local highly specific information processing.Originality. Age-related adaptive changes of the neural networks that enable effective performance of the SSMR: different strategies of visual-motor reaction are revealed.72 Серія «Біологічні науки», 202 2Conclusions. Performance of the SSMR test involves different neuronal strategies that were determined for three age groups. The prospect of further research is to determine age-related differences in the chronometric indicators and neuronal mechanisms for SSMR, choice reaction and other psychophysiological tests for different military specialties and according to acquired data create guidelines for conducting surveys to ensure effective performance of work responsibilities with age. Key words: simple sensorimotor reaction; electroencephalogram; coherent analysis; age-related changes; professional selection

Determinant of Covariance Matrix Model Coupled with AdaBoost Classification Algorithm for EEG Seizure Detection.

Experts usually inspect electroencephalogram (EEG) recordings page-by-page in order to identify epileptic seizures, which leads to heavy workloads and is time consuming. However, the efficient extraction and effective selection of informative EEG features is crucial in assisting clinicians to diagnose epilepsy accurately. In this paper, a determinant of covariance matrix (Cov–Det) model is suggested for reducing EEG dimensionality. First, EEG signals are segmented into intervals using a sliding window technique. Then, Cov–Det is applied to each interval. To construct a features vector, a set of statistical features are extracted from each interval. To eliminate redundant features, the Kolmogorov–Smirnov (KST) and Mann–Whitney U (MWUT) tests are integrated, the extracted features ranked based on KST and MWUT metrics, and arithmetic operators are adopted to construe the most pertinent classified features for each pair in the EEG signal group. The selected features are then fed into the proposed AdaBoost Back-Propagation neural network (AB_BP_NN) to effectively classify EEG signals into seizure and free seizure segments. Finally, the AB_BP_NN is compared with several classical machine learning techniques; the results demonstrate that the proposed mode of AB_BP_NN provides insignificant false positive rates, simpler design, and robustness in classifying epileptic signals. Two datasets, the Bern–Barcelona and Bonn datasets, are used for performance evaluation. The proposed technique achieved an average accuracy of 100% and 98.86%, respectively, for the Bern–Barcelona and Bonn datasets, which is considered a noteworthy improvement compared to the current state-of-the-art methods.

Authorship Attribution of Social Media and Literary Russian-Language Texts Using Machine Learning Methods and Feature Selection

Authorship attribution is one of the important fields of natural language processing (NLP). Its popularity is due to the relevance of implementing solutions for information security, as well as copyright protection, various linguistic studies, in particular, researches of social networks. The article is a continuation of the series of studies aimed at the identification of the Russian-language text’s author and reducing the required text volume. The focus of the study was aimed at the attribution of textual data created as a product of human online activity. The effectiveness of the models was evaluated on the two Russian-language datasets: literary texts and short comments from users of social networks. Classical machine learning (ML) algorithms, popular neural networks (NN) architectures, and their hybrids, including convolutional neural network (CNN), networks with long short-term memory (LSTM), Bidirectional Encoder Representations from Transformers (BERT), and fastText, that have not been used in previous studies, were applied to solve the problem. A particular experiment was devoted to the selection of informative features using genetic algorithms (GA) and evaluation of the classifier trained on the optimal feature space. Using fastText or a combination of support vector machine (SVM) with GA reduced the time costs by half in comparison with deep NNs with comparable accuracy. The average accuracy for literary texts was 80.4% using SVM combined with GA, 82.3% using deep NNs, and 82.1% using fastText. For social media comments, results were 66.3%, 73.2%, and 68.1%, respectively.

Computer‐aided diagnosis of COVID‐19 disease from chest X‐ray images integrating deep feature extraction

AbstractThe novel coronavirus (COVID‐19) has an enormous impact on the daily lives and health of people residing in more than 200 nations. This article proposes a deep learning‐based system for the rapid diagnosis of COVID‐19. Chest x‐ray radiograph images were used because recent findings revealed that these images contain salient features about COVID‐19 disease. Transfer learning was performed using different pre‐trained convolutional neural networks models for binary (normal and COVID‐19) and triple (normal, COVID‐19 and viral pneumonia) class problems. Deep features were extracted from a fully connected layer of the ResNET50v2 model and feature dimension was reduced through feature reduction methods. Feature fusion of feature sets reduced through analysis of variance (ANOVA) and mutual information feature selection (MIFS) was fed to Fine K‐nearest neighbour to perform binary classification. Similarly, serial feature fusion of MIFS and chi‐square features were utilized to train Medium Gaussian Support Vector Machines to distinguish normal, COVID‐19 and viral pneumonia cases. The proposed framework yielded accuracies of 99.5% for binary and 95.5% for triple class experiments. The proposed model shows better performance than the existing methods, and this research has the potential to assist medical professionals to enhance the diagnostic ability to detect coronavirus disease.

STUDY NEUR NETWORKS FOR SOFTWARE DEFINED RADIO CONTROL

The scientific article is devoted to the issues of SDR system control. Software Defined Radio is a system designed for software control of information transmission processes in a radio communication channel. Recognition of digital modulation types is used, which automatically classifies the type of digital modulation of the received signal. The following issues are covered in the article: the analysis of existing approaches in the task of automatic recognition of types of digital modulation is carried out; the analysis and classification of informative features in the task of automatic recognition is carried out the following types of digital modulation: 2-PSK, 4-PSK, 8-PSK, 2-FSK, 8-QAM, 16-QAM, 64-QAM and OFDM. This article uses a neural network approach based on cumulative characteristics. To solve these problems, the methods of calculation and selection of informative cumulative features described in the developed model of the system of automatic recognition of digital modulation types by means of a neural network on cumulative features at a known value of carrier frequency are described. A model of the system of automatic recognition of digital modulation types at a known value is constructed carrier frequency. An algorithm for automatic recognition of digital modulation types has been developed using a multilayer neural network. The influence of noise in the communication channel on the probability of recognizing the types of digital modulation at a known value of the carrier frequency is investigated. It was found that regardless of the type of noise in the communication channel, the law of error distribution in IQ data becomes close to normal. This fact is one important argument for the use of cumulative features in the task of automatic recognition of types of digital modulation. Therefore, the task of automatic recognition of digital modulation types is quite relevant. Further research may be aimed at expanding the range of high-order cumulative features used, due to which it is possible to increase the probability of correct recognition of types of digital modulation, and solving the recognition problem at an unknown value of the frequency and initial phase of the carrier signal.

Feature selection and classification over the network with missing node observations.

Jointly analyzing transcriptomic data and the existing biological networks can yield more robust and informative feature selection results, as well as better understanding of the biological mechanisms. Selecting and classifying node features over genome-scale networks has become increasingly important in genomic biology and genomic medicine. Existing methods have some critical drawbacks. The first is they do not allow flexible modeling of different subtypes of selected nodes. The second is they ignore nodes with missing values, very likely to increase bias in estimation. To address these limitations, we propose a general modeling framework for Bayesian node classification (BNC) with missing values. A new prior model is developed for the class indicators incorporating the network structure. For posterior computation, we resort to the Swendsen-Wang algorithm for efficiently updating class indicators. BNC can naturally handle missing values in the Bayesian modeling framework, which improves the node classification accuracy and reduces the bias in estimating gene effects. We demonstrate the advantages of our methods via extensive simulation studies and the analysis of the cutaneous melanoma dataset from The Cancer Genome Atlas.

Survival analysis with semi-supervised predictive clustering trees

Many clinical studies follow patients over time and record the time until the occurrence of an event of interest (e.g., recovery, death, …). When patients drop out of the study or when their event did not happen before the study ended, the collected dataset is said to contain censored observations. Given the rise of personalized medicine, clinicians are often interested in accurate risk prediction models that predict, for unseen patients, a survival profile, including the expected time until the event. Survival analysis methods are used to detect associations or compare subpopulations of patients in this context.In this article, we propose to cast the time-to-event prediction task as a multi-target regression task, with censored observations modeled as partially labeled examples. We then apply semi-supervised learning to the resulting data representation. More specifically, we use semi-supervised predictive clustering trees and ensembles thereof.Empirical results over eleven real-life datasets demonstrate superior or equivalent predictive performance of the proposed approach as compared to three competitor methods. Moreover, smaller models are obtained compared to random survival forests, another tree ensemble method. Finally, we illustrate the informative feature selection mechanism of our method, by interpreting the splits induced by a single tree model when predicting survival for amyotrophic lateral sclerosis patients.

Identification of argumentative sentences in Russian scientific and popular science texts

In this study we analyze the applicability of specific machine learning algorithms to the task of detecting sentences containing argumentation in Russian text. We employ a collection of scientific and popular science texts with manually annotated argumentation to evaluate the quality of identifying argumentative sentences in terms of precision, recall, and F-measure. The experiment involves three algorithms: MNB, SVM, and MLP. The bag of words model is used for representing texts. Lemmas of words in analyzed sentences serve as features for the classification. We perform the automatic selection of informative features in accordance with Variance and χ2 criteria combined with the weight-based filtration of lemmas (via TF*IDF and EMI). The training set includes around 800 sentences, while the test set contains 180. The MNB algorithm demonstrates the highest F-measure and recall scores on almost all feature sets (maximal values reached equal 68.7% and 89% respectively), while the MLP algorithm shows the best precision for about half of feature selection variations (the maximal value is 72.5%).

Manifold Feature Fusion with Dynamical Feature Selection for Cross-Subject Emotion Recognition.

Affective computing systems can decode cortical activities to facilitate emotional human–computer interaction. However, personalities exist in neurophysiological responses among different users of the brain–computer interface leads to a difficulty for designing a generic emotion recognizer that is adaptable to a novel individual. It thus brings an obstacle to achieve cross-subject emotion recognition (ER). To tackle this issue, in this study we propose a novel feature selection method, manifold feature fusion and dynamical feature selection (MF-DFS), under transfer learning principle to determine generalizable features that are stably sensitive to emotional variations. The MF-DFS framework takes the advantages of local geometrical information feature selection, domain adaptation based manifold learning, and dynamical feature selection to enhance the accuracy of the ER system. Based on three public databases, DEAP, MAHNOB-HCI and SEED, the performance of the MF-DFS is validated according to the leave-one-subject-out paradigm under two types of electroencephalography features. By defining three emotional classes of each affective dimension, the accuracy of the MF-DFS-based ER classifier is achieved at 0.50–0.48 (DEAP) and 0.46–0.50 (MAHNOBHCI) for arousal and valence emotional dimensions, respectively. For the SEED database, it achieves 0.40 for the valence dimension. The corresponding accuracy is significantly superior to several classical feature selection methods on multiple machine learning models.

Identification of most important features based on a fuzzy ensemble technique: Evaluation on joint space narrowing progression in knee osteoarthritis patients

ObjectiveFeature selection (FS) is a crucial and at the same time challenging processing step that aims to reduce the dimensionality of complex classification or regression problems. Various techniques have been proposed in the literature to address this challenge with emphasis to medical applications. However, each one of the existing FS algorithms come with its own advantages and disadvantages introducing a certain level of bias. Materials and MethodsTo avoid bias and alleviate the defectiveness of single feature selection results, an ensemble FS methodology is proposed in this paper that aggregates the results of several FS algorithms (filter, wrapper and embedded ones). Fuzzy logic is employed to combine multiple feature importance scores thus leading to a more robust selection of informative features. The proposed fuzzy ensemble FS methodology was applied on the problem of knee osteoarthritis (KOA) prediction with special emphasis on the progression of joint space narrowing (JSN). The proposed FS methodology was integrated into an end-to-end machine learning pipeline and a thorough experimental evaluation was conducted using data from the Osteoarthritis Initiative (OAI) database. Several classifiers were investigated for their suitability in the task of JSN prediction and the best performing model was then post-hoc analyzed by using the SHAP method. ResultsThe results showed that the proposed method presented a better and more stable performance in contrast to other competitive feature selection methods, leading to an average accuracy of 78.14% using XG Boost at 31 selected features. The post-hoc explainability highlighted the important features that contribute to the classification of patients with JSN progression. ConclusionsThe proposed fuzzy feature selection approach improves the performance of the predictive models by selecting a small optimal subset of features compared to popular feature selection methods.