Analysis and multiclass classification of pathological knee joints using vibroarthrographic signals

https://doi.org/10.1016/j.cmpb.2017.10.027Get rights and content

Highlights

  • We extend the VAG signal categorization to a multiclass classification, according to the various PFJ disorders and its stages, diagnosed by MRI, considered as the gold standard for PFJ chondral lesions.

  • Using SimpleLogistic algorithm, we obtained 69% and 90% accuracy and AUC respectively, and sensitivity and specificity over 91% and 69%.

  • Analysis and the classification of the VAG signals give satisfactory results for the screening and constitute a promising tool for classifying signals of various knee joint disorders and their stages.

Abstract

Background and Objective

Vibroarthrography (VAG) is a method developed for sensitive and objective assessment of articular function. Although the VAG method is still in development, it shows high accuracy, sensitivity and specificity when comparing results obtained from controls and the non-specific, knee-related disorder group. However, the multiclass classification remains practically unknown. Therefore the aim of this study was to extend the VAG method classification to 5 classes, according to different disorders of the patellofemoral joint.

Methods

We assessed 121 knees of patients (95 knees with grade I-III chondromalacia patellae, 26 with osteoarthritis) and 66 knees from 33 healthy controls. The vibroarthrographic signals were collected during knee flexion/extension motion using an acceleration sensor. The genetic search algorithm was chosen to select the most relevant features of the VAG signal for classification. Four different algorithms were used for classification of selected features: logistic regression with automatic attribute selection (SimpleLogistic in Weka), multilayer perceptron with sigmoid activation function (MultilayerPerceptron), John Platt's sequential minimal optimization algorithm implementation of support vector classifier (SMO) and random forest tree (RandomForest). The generalization error of classification algorithms was evaluated by stratified 10-fold cross-validation.

Results

We obtained levels of accuracy and AUC metrics over 90%, more than 93% sensitivity and more than 84% specificity for the logistic regression-based method (SimpleLogistic) for a 2-class classification. For the 5-class method, we obtained 69% and 90% accuracy and AUC respectively, and sensitivity and specificity over 91% and 69%.

Conclusions

The results of this study confirm the high usefulness of quantitative analysis of VAG signals based on classification techniques into normal and pathological knees and as a promising tool in classifying signals of various knee joint disorders and their stages.

Introduction

Vibroarthrography (VAG) is an experimental method developed for noninvasive assessment of articular function, especially arthrokinematics. The VAG method is based on the analysis of high frequency vibroacoustic emission, which is a natural phenomenon acquired from the relative motion of articular surfaces of the synovial joint (diarthrosis) [1], [2], [3]. In physiological conditions, articular surfaces covered by hyaline cartilage are smooth and slippery, which determines optimal arthrokinematic motion quality [1], [4]. In contrast, degenerated cartilage results in greater friction during movement, which is reflected in an increase in amplitude and frequency of the VAG signal [2]. Chondral lesions (such as chondromalacia or osteoarthritis) are often observed in a patellofemoral joint (PFJ), a part of the knee joint complex, which can be explained by its specific biomechanical environment and substantial involvement in daily/sports activity. Due to this but also due to having the greatest susceptibility to the VAG test resulting from a superficial position, the knee is the joint most commonly analyzed by VAG.

Although the VAG method is still in development, it shows high accuracy, sensitivity and specificity, when comparing results obtained from controls and a non-specific, knee-related disorder group [5]. Nalband et al. [6] applied the least square support vector machines algorithm based on the time-complexity parameters of the VAG signal and obtained greater than 94% classification accuracy, greater than 98% sensitivity and 86% specificity [6]. Kim et al. [4] presented classification of the neural network with frequency parameters as inputs, which allowed for improvement of the accuracy to more than 95%, sensitivity 92% and specificity of more than 98% [4]. The best results of the normal-abnormal classification signal are found in the work of Rangayyan et al. [7]. The authors used a classifier based on a radial basis function network with statistical parameters in the time domain. Here, the accuracy, sensitivity and specificity reached 100%, with the cross-validation of the leave-one-out method [7].

Nevertheless, it should be noted that the mentioned authors based their analysis only on two classes, normal and abnormal. However, from a clinical point of view, the 2-class classification is insufficient for an appropriate diagnostic process and further adequate treatment, and a more specific categorization of chondral-related changes is necessary [8]. Similarly, the radiological staging of chondral disorders (especially early stages, such as chondromalacia) using X-ray also possess significant limitations, due to the low sensitivity and specificity [9]. On the other hand, availability of modern imaging methods such as magnetic resonance imaging (MRI) is limited due to the high expense [9]. Moreover, current diagnostic methods are entirely observer-dependent and require significant knowledge, expertise, and time. Therefore, due to the constantly increasing incidence of age-related cartilage lesions, there have been calls for the development of noninvasive, observer-independent and financially accessible methods for evaluation of human joints, with sensitivity and specificity comparable with MRI, considered the gold standard for chondral lesion assessment.

Recently it has been demonstrated that the VAG method could be helpful in differentiating particular disorders of the PFJ and its stages, due to the specific, disorder-related character of the VAG signal pattern [2], [9]. However, while the problem of classification of normal and abnormal VAG signals has been studied, extending it to a multiclass classification remains practically unaddressed. Moreover, as previously suggested, further work is needed to determine whether the sensitivity and specificity of the VAG method are sufficient for clinical application [10]. Furthermore, there is a pressing need for description of optimal algorithms for VAG signal multiclass classification, in accordance with the clinical criteria of PFJ chondral lesions [Nalband]. Optimization of diagnostic methods should include the selection of the most relevant and discriminating VAG signal parameters, followed by selection of an optimal predictive model [11], [12], [13], [14]. This will allow us to develop an observer-independent, sensitive, computer-aided diagnostic method, useful for clinicians, in particular for orthopedists and physiotherapists, who are concerned with evaluation of the quality of arthrokinematic motion during physical examination [1].

Thus, the primary goal of our study will be to extend the VAG signal categorization of various PFJ chondral lesions to a 5-class classification (normal and four classes of disorders). Our analyses will be performed with respect to the MRI examination, as a reference method of noninvasive assessment of chondral lesions, which will allow us to evaluate the sensitivity and specificity of the VAG method. For the optimization problem, we applied two algorithms for selecting the best parameters: genetic search and selection based on simple regression functions. Then we compared four classification models representing different approaches to the classification problem: logistic regression with automatic attribute selection based on simple regression functions, multilayer perceptron with sigmoid activation function, sequential minimal optimization algorithm implementation of support vector machine classifier and random forest tree.

The paper is structured as follows: Section 2 describes the analyzed material and the methodology of the study, including the feature extraction techniques, feature selection algorithms and classification methods. Section 3 presents the obtained results, which are discussed in Section 4. Finally, Section 5 presents our conclusions.

Section snippets

Participants

121 knees from 56 patients with chondromalacia patellae (CMP) (38 with bilateral and 20 with unilateral symptoms) and 16 patients with osteoarthritis (OA) (10 with bilateral and 6 with unilateral symptoms) were enrolled in the study. Knees with CMP were classified into 3 grades according to criteria of the International Cartilage Repair Society by MRI imaging: CMP stage I (28 knees; CMPI), CMP stage II (31 knees; CMPII) and CMP stage III (36 knees; CMPIII) [2], [9]. OA patients were selected

Results

In this section we describe the results of VAG signal parameterization and selection of the most relevant discriminate parameters and models using software and algorithms which are described in detail in the previous section.

Discussion

Results presented in this paper confirm the high usefulness of the quantitative analysis of VAG signals based on classification techniques for normal and pathological knees [4], [5], [6], [26]. We obtained over 90% levels of accuracy and AUC metrics, more than 93% sensitivity and more than 84% specificity for the logistic regression-based method (SimpleLogistic). These values are slightly lower than those we can find in the published research works, where the values obtained for accuracy,

Conclusion

In conclusion, we found that the VAG method seems to be sensitive enough for evaluating the biomechanical and morphological changes within the knee joint environment. Analysis and the 5-class classification of the VAG signals give satisfactory results not only for the screening tests but also for classifying signals of various knee joint disorders and their stages, according to the level of chondral lesion [33], [34]. Thus, the VAG method may constitute a helpful tool for clinicians and may be

Acknowledgments

We would like to thank all of the subjects who participated in the study.

The project was approved by the Ethics Committee of Opole Voivodship. Signed informed consent was obtained from all tested persons.

The authors declare no conflict of interest.

The authors report no external funding source for this study.

References (36)

  • Y. Wu

    Knee Joint Vibroarthrographic Signal Processing and Analysis

    Knee Joint Vibroarthrographic Signal Processing and Analysis

    (2015)
  • G. McCoy et al.

    Vibration arthrography as a diagnostic aid in diseases of the knee. A preliminary report

    J. Bone Jt. Surg. Br.

    (1987)
  • D. Bączkowicz et al.

    Joint motion quality in vibroacoustic signal analysis for patients with patellofemoral joint disorders

    BMC Musculoskeletal Disord.

    (2014)
  • E. Petre et al.

    Neural networks-based adaptive control for a class of nonlinear bioprocesses

    Neural Comput. Appl.

    (2009)
  • J. Li et al.

    Design and performance analysis of a 866 MHz low-power optimized CMOS LNA for UHF RFID

    IEEE Trans. Ind. Electron.

    (2013)
  • S.B. Ghosn et al.

    A parallel genetic algorithm for the open-shop scheduling problem using deterministic and random moves

    Int. J. Artif. Intell

    (2016)
  • R.M. Rangayyan et al.

    Analysis of vibroarthrographic signals with features related to signal variability and radial-basis functions

    Ann. Biomed. Eng.

    (2009)
  • R.M. Rangayyan et al.

    Screening of knee-joint vibroarthrographic signals using statistical parameters and radial basis functions

    Med. Biol. Eng. Comput.

    (2008)
  • Cited by (33)

    • A novel computer-assisted diagnosis method of knee osteoarthritis based on multivariate information and deep learning model[Formula presented]

      2023, Digital Signal Processing: A Review Journal
      Citation Excerpt :

      From the point of view of waveform energy, the spectrogram [9], spectral power and frequency features [10], power of spectra density have been taken as the VAG features to automatically perform the knee joint disorders detection. In addition, the approximate entropy, recurrence quantification measurements, and sample entropy [5], fractal dimension [7], multiscale sample entropy, fractal scaling index [11] etc., which parameterize the variability of VAG signals, also have been applied to be the VAG features to distinguish between normal and KOA states. However, such traditional machine learning-based detection method largely relies on artificial designed VAG features, which often has high-cost extraction process and poor adaptability for different subjects.

    • A survey on DE – Duplication schemes in cloud servers for secured data analysis in various applications

      2022, Measurement: Sensors
      Citation Excerpt :

      Usually image data are considered for analysis which while transferring in a network has to be converted into a cipher image and the decryption had to be done as a final stage after transmission of the image [18]. There are number of machine learning and cloud combined approaches which are utilized in healthcare applications so that the data transfer and performance improvement are done simultaneously for each of the clinical areas considered [19–26]. The environment used for deploying the deduplication method is considered as the key focus of the survey.

    • Classification of pathological disorders using optimization enabled deep neuro fuzzy network

      2022, Biomedical Signal Processing and Control
      Citation Excerpt :

      To classify time-series data with DNN the deep learning techniques are used. The imperative aspect of deep networks is removing the requirement of mining handcrafted features and learning the features through training for voice disorder assessment [5] (see Table 1). From previous studies, the voice pathology detection techniques gained huge development, due to machine learning models wherein these techniques are considered as an effective part of medical applications and voice detection models.

    View all citing articles on Scopus
    View full text