Classification of sympathetic skin response based on the morphologic features and Adaptive Neuro Fuzzy Inference system( ANFIS)

doi:10.21203/rs.3.rs-2957332/v1

Download PDF

Research Article

Classification of sympathetic skin response based on the morphologic features and Adaptive Neuro Fuzzy Inference system( ANFIS)

https://doi.org/10.21203/rs.3.rs-2957332/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The prevalence of polyneuropathy (PNP) or peripheral neuropathy (PN) is estimated to be 2%-3% in the general population and may be as high as 8% in people over 55 years of age. It’s the most common type of disorder of the peripheral nervous system in adults and in the elderly. Early detection and accurate classification of PNP can lead to proper diagnosis and treatment of painful symptoms. Our team developed a new method to classify the presence or absence of PNP in a database based on Adaptive Neuro Fuzzy Inference system( ANFIS) using sympathetic skin response (SSR) signal. To realize an efficient detection the output of our classification is divided into four classes such as the severity of PNP: no-PNP, mild, moderate, and severe class. In fact, we propose to extract the morphologic features of SSR signal including Latency, amplitude, rise time, the typical recovery time of 63%, and the typical recovery time of 50% which can be altered by PNP. Thus, the performances of the PNP severity classification system were compared with different machine learning (ML) algorithms such as support vector machine (SVM), K-nearest neighbor (KNN). Hence, The ANFIS model showed better performance in comparison to different ML models. In the classification stage, the best classification performance was achieved as 97.16%, 84.40%, and 87.12%% using ANFIS, KNN, and SVM classifier respectively.

Polyneuropathy (PNP)

Support Vector Machine (SVM)

Adaptive Neuro Fuzzy Inference system( ANFIS) )

K-nearest neighbor (KNN)

sympathetic skin response (SSR)

Polyneuropathy (PNP) or peripheral neuropathy (PN) is one of the most common neurological diseases that damages peripheral nerves and results in a loss in feedback control of the locomotion [1]. PNP etiologies vary by fiber size (small or large fiber), fiber type (somatic motor, somatic sensory, or autonomic), pattern of injury (axonal or demyelinating), distribution (length-dependent and length-independent), and rapidity of onset (acute, subacute, or chronic), Diabetes, B12 deficiency, alcohol, monoclonal gammopathy of undetermined significance (MGUS), and hereditary PNP or Charcot–Marie–Tooth disease (CMT) represent common causes, and most often present in a length-dependent fashion [2–3]. In fact, PNP leads to numerous lower extremity complications, such as numbness, burning, pinprick sensation, and pain [4–5]. Given the severity of this complication, early and accurate detection of PN is inevitable for appropriate treatment and to avoid serious consequences.

In the opening iteration, the researchers employed numerous methods to determine the presence or absence of PNP such as gait, Corneal Confocal Microscopy Image (CCMI), muscle electromyography (EMG), and ground reaction forces (GRF) during gait, Michigan Neuropathy Screening Instrumentation (MNSI), nerve conduction studies (NCS)…etc. In [6], the authors have used robust marker less gait detection system was employed and experiments were conducted in normal clinical conditions on 10 young, 10 old and 10 neuropathy patients. They have extracted four types of gait features, namely temporal features, kinematic joint trajectories in time domain, the Fourier transform of joint angles in frequency domain, and the symmetry indexes. Each type of features is selected using One-way analysis of variance (ANOVA) were served as the input of a two-layer-feed-forward neural network separately. In the classification step, Wang, X et al. have used individual feature set that showed that the kinematic features in time domain reached the highest classification accuracy 94,8%, while the symmetric feature set gave the lowest accuracy 78,5%. However, they have combined 2 features. The combination Temporal and kinematic features set with the highest accuracy of 95,8%, while the combination of Kinematic and symmetric features feature set gave the lowest accuracy 95,3%.

Meng, Y. et al in [7] have employed corneal confocal microscopy (CCM) images of the sub-basal nerve plexus to classify the presence or not of peripheral neuropathy (PN) in participants with diabetes or pre-diabetes. The used database contains 279 participants (type 1 diabetes without PN [n = 49], type 1 diabetes with PN [n = 39], type 2 diabetes without PN [n = 74], type 2 diabetes with PN [n = 67], pre-diabetes without PN− [n = 26], and pre-diabetes with PN [n = 24]) utilizing one image per participant. For the detection of PN case, the authors have developed an artificial intelligence (AI) based deep learning algorithm (DLA) for the binary classification of patients that achieved a sensitivity of 91%, a specificity of 93%, and an area under the curve (AUC) of 0.95.

In the last 50 years, the researchers are trying to observe the biomechanical changes due to PNP by studying Electrodiagnostic (Electrodermal activity (EDA), Electromyography (EMG)… etc.) testing that can not only identify whether or not a PNP may be present, but can help give the clinician information to determine the etiology, the severity, and the prognosis of the disorder [3].

In [8], the researchers used changes in muscle activity during gait from three different lower extremity muscles (vastus lateralis (VL), tibialis anterior (TA), and gastrocnemius medialis (GM)) electromyography (EMG) of Diabetic sensorimotor polyneuropathy (DSPN) patients. They have used a database that contains 147 volunteers’ adults that were divided into five groups; 30 subjects with no previous history of DM are considered as a control group, and 4 DSPN severity groups were classified: 43 subjects taken from diabetic group and non-neuropathic (in this project classified as absent DSPN), 30 diabetic individuals from mild DSPN, 16 moderate DSPN subjects, and 28 from the severe DSPN subjects. The dataset has been classified into four classes (Absent, Mild, Moderate, and Severe) using machine learning. The authors have extracted nineteen-time domain features from each lower limb muscles (GM, VL, and TA), and dimensional GRF (GRFx, GRFy, and GRFz). The model accuracy was validated with the results from different machine learning algorithms. In fact, the authors have used the combination of Top 13 features from GRFx set the highest accuracy 90,11%, while the combination from GRFy set gave the lowest accuracy 86.44%. In addition, they have used the combination of two channels, the machine learning model achieved 94.78% accuracy by Top 15 features combined GRFx and GRFz signals, 92.56% accuracy for Top 16 features combined from GRFx and GRFy signals, 93.78% accuracy by Top 13 features combined from GRFy and GRFz signals. From the three-channel analysis, the machine learning model achieved 94.67% accuracy by Top 15 features combination from GRFx, GRFy and GRFz.

Autonomic nerves are made up of small myelinated and unmyelinated fibers that are present in the skin (somatic and sudomotor fibers), peripheral nerves and organs. When these fibers are damaged, it is called small fiber neuropathy (SFN) [9–10].

Unfortunately, the EMG which is the usual examination for diagnosing SFN, only explores the large fibers and does not provide information on the small ones [11]. Therefore, the sympathetic skin response (SSR) provides information about a part of the peripheral nervous system small unmyelinated C fibers that cannot be assessed by currently performed techniques used in clinical EMG laboratories or by physical examination [12–13]. Thus, researchers have diagnosed the different types of polyneuropathies (diabetic polyneuropathy [13], Diabetic Cardiac Autonomic Neuropathy [15], transthyretin familial amyloid polyneuropathy [16] …etc.) using SSR.

After an extensive literature study, we found a state-of-the-art solution for PNP detection. This solution surpasses all the literature mentioned above because of its advantages.

This study can help the health professionals with making an accurate decision as for diagnosis SFN. In this method, we use the SSR signals. These signals are enriched with information, which help researchers and doctors to detect SFN. The proposed method is based on two main steps: the first one is the feature extraction and the second one is the anomaly classification.

Therefore, as a first step we compute the morphologic features (amplitude, latency time, rise time, the typical recovery time…etc.) of the SSR signal. After that, we classify the extracted features by using Adaptive Neuro Fuzzy Inference system (ANFIS).

The rest of the study is divided into following sections: Section II presents the methodology of this study which describes following processes: data collection, the morphologic features, methods of classification, we report the experimental results of the proposed method defined in section Ⅲ, and in section Ⅳ our discussion is the design, in the last section, we concluded.

In the following section, we describe the dataset that has been used in this paper. Then, we describe the methodology adopted in our work and we highlight the investigated tools in detail.

Furthermore, in this work, we propose an automated approach for the classification of PNP diseases. Therefore, our method involves two major steps including feature extraction and classification. The principle of the proposed approach is presented in Fig. 1 and is mainly divided into two tasks:

The first task represents the morphologic feature extraction providing the useful information for the second task.
The second task is the classification process. This system is used to detect and classify PNP; in this task the classification model is simulated using Adaptive Neuro Fuzzy Inference system.

A. Database

EDA recordings of hand and foot skin were taken from 36 normal subjects (17 F,19 M, aged 19–68) recordings of hand and foot skin. Table 1 illustrates the physical characteristics of these groups of subjects. Controls had no clinical peripheral or autonomic nerve involvements and used no drug which could influence sympathetic activity.

105 pathological subjects also studied (aged between 26–86) who were otherwise neurologically (Table 1). Clinical work-up including EMG, motor and sensory nerve conduction measurements were all normal.

Table 1

physical characteristics of these groups of subjects (F: Female, M: Male, SD: Standard Deviation).
	Gender		Age		Height(cm)		Mass (Kg)
	M	F	Mean	SD	Mean	SD	Mean	SD
Healthy subjects	24	12	45	14.3	174.52	8.15	71.4	7.2
Pathologic subjects	56	50	66.72	12.77	170.58	10.12	76.8	7.56

The examination room was at a constant temperature between 22 and 24 ° C. The subjects were comfortably seated on the examination table in a semi-sitting position. Except for control subjects, they first benefited from the ENMG conducted in order to characterize PNP. All subjects had at least, in addition to a detection examination, a study of the motor conduction of the deep fibular and posterior tibial nerves and a study of the sensory conduction of the superficial and deep fibular nerves on the side where the symptomatology was maximal. For SSR measurement, after skin cleaning and electrode placement, they were asked to relax and breathe regularly. The stimulations consisted of electrical pulses of 0.2 ms duration, applied to the skin opposite the median nerve on the left wrist The intensity was 20 mA. They were delivered when response monitoring was stable, with a random interval greater than 20 seconds. Their number was the one that allowed to obtain a response with a reliable measurement of latency and amplitude, without exceeding five. The recording was done by 1 cm diameter adhesive electrodes (Nutab Diagnostic Electrodes, Kendall) put at hand and foot contralateral to the stimulation. Cathodes were on the sole of the foot and on the palm of the hand, and anodes on the dorsum of the foot and the hand. Amplification and visualization of the signal were provided by an electromyographic device (Neuro-Mep4, Neurosoft). The input dynamics was 6 mV. The sensitivity was adjusted during the measurement to obtain a clear take off from the response. The bandwidth was between 0.2 and 20 Hz.

In fact, to improve our database for each acquisition, we are remade 5 times. Each measurement began with a rest period of 2 minutes, to ensure that SSR signals were not overlapping. In fact, Fig. 2 illustrates an example of typical examples of SSR data recorded from healthy and pathologic subjects, the upper graph Fig. 2a) shows the SSR data recorded from the hand of subjects haven’t PNP. b) recorded from the foot of subjects haven’t PNP. In the bottom graph Fig. 2c) the EDA data were recorded from the hand subjects have PNP. d) recorded from the foot of subject with PNP.

Measurements of the response parameters were made by the sliders on the device and were the onset of the first negative wave, for its latency, and the peak-to-peak interval between the negative wave and the positive wave for the amplitude. In the absence of a negative wave, the latency was the onset of the positive wave, and the amplitude was the difference between the trough of this wave and the starting point. Of the five trials performed, with horizontal baseline and clear response takeoff, we selected the one with the shortest latency, which most often corresponded to the one with the largest amplitude, and which was usually the first trial. If the responses were not of sufficient quality, the same technique was applied on the other side. In the case where the SSR was not obtained, either in the hand or in the foot, it was tried to record it with stimulations of intensity higher than 30 mA, and after a deep inspiration, to differentiate abolitions by central deficit or deficit of sensory afferences. The length between the foot and the hand corresponded to the distance between the palm and the ground when the subject was standing (see justification in Fig. 2). Hand and foot skin temperature was measured in all subjects at the end of the examination where the cathodes had been placed (Neurosoft thermal probe).

B. features extraction

The SSR is most often utilized in peripheral neuropathies to identify functional impairment of non-myelinated postganglionic sudomotor sympathetic fibers. In the literature, several terms are used depending on the method of measurement such as electrodermal activity (EDA), electrodermal response (EDR), psychogalvanic reflex (PGR), galvanic skin response (GSP), peripheral autonomic surface potential (PASP) and the most frequently used SSR [17–18].

Thus, SSR is characterized by the morphologic features such as the amplitude, latency time, rise time, the typical recovery time of 63% of SSR, and the typical recovery time of 50% of SSR. These features have been previously used in [19–24] and they have showed a very important potential for classification. Figure 3 presents how to extract these features via an SSR signal.

The latency of SSR (second) is the time interval between the stimulation artifact and the start of the SSR [17].
the amplitude of SSR (mV) is measured from the peak of the first deflection to the peak of the next one (peak to peak) [17].
The rise time (seconds) Time interval between the start (the measured latency) of the signal response and the second peak [22].
Half-recovery time (seconds) the time interval between the second peak and the point at which the signal increases by half the second peak.
recovery time of 63% (seconds) the time interval between the second peak and the point at which the signal increases by 63% of the second peak.

C. Adaptive Neuro Fuzzy Inference System (ANFIS)

ANFIS belongs to a family of hybrid system, was introduced by Jyh-Shing Roger Jang [24–26] in 1993 which is the combination of both advantages of fuzzy logic principals and artificial neural network. ANFIS is a fuzzy mapping algorithm based on Takagi-Sugeno fuzzy inference system. Its inference system corresponds to a set of fuzzy IF–THEN rules that have learning capability to generate the input/output (I/O) pairs [25], can be written using the following Equations (2) and (3), where the inputs are x and y, the fuzzy sets are B_i and A_i. p_i, q_i and r_i are the parameters of design that are denoted in training and f_i are the outputs described by the fuzzy rule.

Rule 1: if x is A₁ and y is B₁ then \(\mathop f\nolimits_{1} =\mathop p\nolimits_{1} x+{q_1}y+{r_1}\) 1

Rule2: if x is A₂ and y is B₂ then \(\mathop f\nolimits_{2} =\mathop \text{p}\nolimits_{2} \text{x}+{\text{q}_2}\text{y}+{\text{r}_2}\) 2

ANFIS structure is comprised of five layers (i.e., layer 1 to layer 5) with directional links and nodes. In fact, ANFIS training algorithm is a hybrid Learning Algorithm, which consists of two passes:1) the ordinary least squares algorithm to update coefficients of the output function. 2)the error back propagation algorithm is used to update fundamental factors of the system [27]. Figure 4 shows two input ANFIS architecture. The function of each layer of the ANFIS architecture is described below[24–28].

Layer 1

The first layer is called as the fuzzification layer. The fuzzification layer takes the input value and determines the membership function and belonging to them. Every node i in this layer is a square node with a node function the output of each node is calculated using Equations (3) (4).

\(O_{\text{i}}^{1}={\mu _{{\text{A}_\text{i}}}}(\text{x}), \text{i}=1,2\) 3

\(O_{\text{i}}^{1}={\mu _{{B_{\text{i}-2}}}}(\text{x}), \text{i}=3,4\) 4

where, x is the input to node i, and Ai, is the linguistic label (small, large, etc.) associated with this node function. Parameters in this layer are referred to as premise parameters. Node output is the membership value of the input.

Layer 2

The second layer is called rule layer [28]. It’s responsible for generating the firing strengths for the rules. Every node in this layer is a circle node labeled π, which multiplies the incoming signals and sends the product out as in the Eq. (4).

\(O_{\text{i}}^{2}={\text{w}_i}={\mu _{{\text{A}_\text{i}}}}(\text{x})*{\mu _{{B_\text{i}}}}(x), \text{i}=1,2,...\) 5

Layer 3

this layer is called as Normalization layer. His role is to normalize the computed firing strengths by diving each value with the total firing strength. Every node in this layer is a circle node labeled N.

\(\text{O}_{\text{i}}^{3}={\bar {\text{w}}_\text{i}}=\frac{{{\text{w}_\text{i}}}}{{{\text{w}_1}+{\text{w}_2}}}, \text{i}=1,2,...\) 6

Layer 4

This layer is called as defuzzification layer. Weighted values of rules are calculated in each node of this layer as given in Eq. (6).

\(O_{i}^{4}={\bar {w}_i}{f_i}={\bar {w}_i}({p_i}{x_i}+{q_i}y+{r_i})\) 7

Node outputs are the evaluation of right-hand side polynomials of Eq. 6 where (\(\stackrel{-}{w}\)_i) is the output of layer 3.and {p_i,q_i,r_i} is the parameter set. These are called the consequence parameters. The number of consequence parameters of each rule is one more than the number of inputs.

Layer 5

It is called as the summation layer. The actual output of ANFIS is obtained by summing the outputs obtained for each rule in the layer 4.

\(O_{i}^{5}=\sum\limits_{i} {{{\bar {w}}_i}{f_i}=\frac{{\sum\limits_{i} {{w_i}{f_i}} }}{{\sum\limits_{i} {{w_i}} }}}\) 8

In the order to detect the presence or absence of PNP that attack the small myelinated and unmyelinated fibers. The current section, we propose to extract the morphologic features(amplitude, latency, rise time, typical recovery time…etc.) that extracted from SSR using ANFIS.

In our dataset, we have considered 720 samples per no-PNP, 700 samples per severe neuropathy, 700 samples per moderate neuropathy and 700 mild neuropathies, a total of 2820 samples that recorded from hands and feet. We have selected these five parameters as inputs. So, the final dataset was consisting of five inputs: latency, amplitude, rise time, half-recovery time, 63% recovery time and one output: PNP severity level (grade 0: absent or non-PNP, grade 1: mild PNP, grade 2: moderate PNP, grade 3: severe PNP). For each database,30% of the data set was reserved for testing set and the remaining 70% was used for training set running with Matlab® software. Figure 5 shows the number of samples in each class among the original, train, and test sets.

The SSR signals classification results are presented in Table.2. Figure 5 shows the confusion matrix for PNP severity classification using the ANFIS model. 11 absent patients, 26 mild patients, 4 moderate patients and 33 severe patients were misclassified by the ANFIS model.

To enhance the evaluation study, the performance of the proposed approach was evaluated using and the performance metrics such as accuracy (Acc), Sensitivity (Sen), Specificity (Spe), Precision (Pre) and F1-score was calculated using the following formulas. There are four classes of PNP severity: absence (no-PNP), mild, moderate and severe PNP. where :

\(\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}(\% )=\frac{{\text{T}\text{P}+\text{T}\text{N}}}{{\text{T}\text{P}+\text{T}\text{N}+\text{F}\text{P}+\text{F}\text{N}}}*100\) 9

\(\text{S}\text{e}\text{n}\text{s}\text{i}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}(\% )=\frac{{\text{T}\text{P}}}{{\text{T}\text{P}+\text{F}\text{N}}}*100\) 10

\(\text{S}\text{p}\text{e}\text{c}\text{i}\text{f}\text{i}\text{c}\text{i}\text{t}\text{y}(\% )=\frac{{\text{T}\text{N}}}{{\text{T}\text{N}+\text{F}\text{P}}}*100\) 11

\(\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}(\% )=\frac{{\text{T}\text{P}}}{{\text{T}\text{P}+\text{F}\text{P}}}*100\) 12

\(\text{F}1\_\text{s}\text{c}\text{o}\text{r}\text{e} (\% )=\frac{{2\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}*\text{S}\text{e}\text{n}\text{s}\text{i}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}}}{{\text{P}\text{r}\text{e}\text{c}\text{i}\text{s}\text{i}\text{o}\text{n}+\text{S}\text{e}\text{n}\text{s}\text{i}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}}}*100\) 13

TP of healthy is all the healthy data that are correctly classified.

TN of healthy is all the non-healthy data that are not classified as healthy.

FP of healthy is all the non- healthy data that are classified as healthy.

FN of healthy is all the healthy data that are not classified as healthy.

Same of the remaining classes PNP severity: mild, moderate and severe.

The performances of the PNP severity classification system were compared with different conventional machine learning (ML) algorithms such as support vector machine (SVM), K-nearest neighbor (KNN). Table 2 shows the comparative performance analysis of ANFIS with different ML algorithms for PNP severity classification. Accuracy, sensitivity, precision, and specificity of the ANFIS are 97.16%, 100%, 97.14% and 94.60%% respectively. Figure 7 shows the ROC curve were an area under the curve (AUC) of 0.989.

Table 2

Performance evaluation of the proposed ANFIS model using morphologic features
Classifier	Performance
Classifier	Spe (%)	Sen (%)	Pre (%)	Acc (%)	F1-score(%)
ANFIS	94.60%	100%	97.14%	97.16%	98.54%
KNN	90;15%	67.59%	87.24%	84.40%	76.16%
SVM	86.66%	88.42%	87.22%	87.12%	87.81%

Recently, various researches had proposed the automatic PNP severity classification using among dataset such as gait [6], Corneal Confocal Microscopy Image (CCMI) [7], muscle electromyography (EMG) and ground reaction forces (GRF) during gait [8][29], Michigan Neuropathy Screening Instrumentation (MNSI)[30][32], nerve conduction studies (NCS) [31]…etc. In this work, we describe a novel method to diagnostic the absence or the presence of PNP diseases using SSR data. Tables 2 prove that the accuracy values of the classification were greater than 97.16%. This result outperforms the results obtained by other research who used different data base illustrated in the Table 3.

In Table 3, we represent a comparative study between the proposed approach and the conventional method presented in the literature

Table 3

comparison between our proposed method and other methods
References	Classifier	Type of data base	Used Features	Accuracy
Wang, X.et al [6]	two-layer-feed-forward neural network	markerless gait	Temporal features Kinematic features Frequency features Symmetry features Temporal & Kinematic Temporal & Frequency Temporal &Symmetry Kinematic & Frequency Frequency & Symmetry Kinematic & Symmetry Temporal & Kinematic & frequency Temporal & Kinematic & symmetry Frequency & symmetry & temporal Frequency & symmetry & kinematic All features Very significant features using ANOVA Very significant & significant	83.3% 94.8% 93.8% 78.8% 95.8% 95.6% 93.0% 96.7% 95.3% 94% 96% 95.3% 96.7% 95.4% 96.0% 95.9% 97.0%
Haque, F.et al [8]	Machine learning(ML)	muscle electromyography (EMG), and ground reaction forces (GRF)	GM (Top10 features) TA(Top10 features) VL (Top10 features) GM&VL(Top13 features) GM&TA(Top13 features) VL&TA(Top14 features) GM&VL&TA(Top 14 features) GRFx (Top13 features) GRFy(Top10 features) GRFz(Top10 features) GRFx & GRFy(Top15 features) GRFx& GRFz(Top16 features) GRFy & GRFz(Top13 features) GRFx & GRFy& GRFz(Top15 features)	86.56% 86.89% 86.89% 91.67% 91% 88.11% 92.89% 90.11% 89.11% 86.44% 94.78% 92.56% 93.78% 94.67%
Haque, F et al [30]	RF ANN SVM EA KNN DAC NB LR	Michigan Neuropathy Screening Instrumentation (MNSI).	Questionnaire, Vibration Perception, and Tactile Sensitivity	91.8% 90% 89.69% 89.98% 87.81% 87.21% 87.81% 87.21%
Haque, F et al [32]	ANFIS EC SVM KNN NB DAC ANFIS	MNSI EMG&GRF	Questionnaire, Vibration Perception, and Tactile Sensitivity peak magnitude peak magnitude and time to peak	93.06% 90.07% 89.07% 88.09% 87.93% 87.24% 96.13% 80.54%
Proposed method	ANFIS KNN SVM	SSR	Using Morphologic features (Latency, amplitude, rise time…etc.)	97.16% 84.40% 87.12%

In [30], Haque, F et al. diagnosed diabetic peripheral neuropathy (DSPN) using composite scoring techniques such as the Michigan Neuropathy Screening Instrumentation (MNSI). MNSI data were collected (1053 samples) from the Epidemiology of Diabetes Interventions and Complications (EDIC) clinical trials. To classify the severity of DSPN, the authors extracted six characteristics: questionnaire, foot appearance (FA), ulceration (Ulc), ankle reflection (AR), vibration perception (VP), and tactile sensitivity (TS). In fact, they used the three most important indices of all MNSI variables that are extracted using the XGBoost model for the classification of the severity of DSPN using the 8 machine learning models: ensemble classifier (EC), random forest (RF), K-nearest neighbour (KNN), decision trees (DT), support vector machine (SVM), Naive Bayes (NB), logistic regression (LR), artificial neural network (ANN). Thus, RF has better accuracy (91.87 ± 1.42), sensitivity (91.8 ± 1.66), specificity (97.23 ± 0.55) compared with other ML algorithms, afterward, ANN, and SVM showed second-best performance( see Table 3).

On the other hand, Haque, F et al. in [31], have investigated the performance of eight different conventional ML algorithms such as EC, RF, KNN, DT, SVM, NB, LR, and DAC for severity classification of DSPN using NCS. Thereby, ten nerve attributes have been considered for DSPN severity grading in this study were classified into four severity classes as absent, mild, moderate, and severe based on the severity. The authors have used three feature selection algorithms: minimum redundant maximum relevant (MRMR), Feature selection using neighbourhood component analysis (fscnca), and Relief algorithms. In fact, EC has better accuracy (93.25 ± 0.95), sensitivity (91.69 ± 1.03), and specificity (98.44 ± 0.61) compared with other ML algorithms, afterward, RF, and DT showed second-best performance using MRMR feature selection technique for NCS. Using relief feature selection technique, EC has better accuracy (93.40 ± 0.97), sensitivity (91.77 ± 1.15), and specificity (98.44 ± 0.73) compared with other ML algorithms. However, EC has better accuracy (93.26 ± 0.91), sensitivity (91.95 ± 1.03), and specificity (98.95 ± 0.62) compared with other ML algorithms using fscnca feature selection.

The proposed method accuracy is large than the others method where they use different database ( see Table 3). According to the Table 3, our method is ranked among the best methods compared to the published works.

In this work, the main novelty to extract the morphologic features of SSR data for the implementation of ANFIS. It results an automated classification of SSR signals into healthy, mild, moderate and severe classes.

In conclusion, the experimental results prove that the proposed approach represents a powerful method with high performances in detecting and classifying PNP diseases.

Peripheral neuropathy (PNP) is a type of nerve disease that can cause various complications in the lower limbs, including numbness, burning, pinprick sensation, and pain. In severe cases, the PNP can lead to ulceration and amputation. It is estimated that 40 to 60 million people with diabetes are affected by lower limb complications due to PNP, and every 30 seconds, one lower limb is amputated due to PNP. Accurate detection of PNP is essential for appropriate treatment and avoidance of severe consequences. Many research and different diagnostic methods have been introduced, yet no specific data is available and conflicting diagnosing criteria are available in a different part of the world.

Therefore, in this study we proposed a now automated methodology for the detection of PNP diseases using morphologic features of SSR signals. The obtained features of each dataset were fed into an Adaptive Neuro Fuzzy Inference system. The proposed approach not only avoided the complex feature extraction process but also used a simple ANFIS structure. The ANFIS model showed better performance in comparison to different ML models. It achieved Accuracy, sensitivity, precision, and specificity 97.16%, 100%, 97.14% and 94.60% respectively, in four classes problems (Healthy, mild, moderate, and severe).

In the future, this proposed technique can be applied for the analysis and diagnosis of other cerebral pathologies. In addition, the classification accuracy performance can be increased by using different decomposition methods to SSR signals. or by using different features.

Figure 1

Framework of the proposed method

Figure 4:

ANFIS architecture that has two input and one output.

Kramer, M., Grimm, A., Winter, N., Dörner, M., Grundmann-Hauser, K., Stahl, J. H., ... & Willikens, S. (2021). Nerve Ultrasound as Helpful Tool in Polyneuropathies. Diagnostics 2021, 11, 211.
Feldman, E. L., Russell, J. W., Löscher, W. N., Grisold, W., Meng, S., Feldman, E. L., ... & Meng, S. (2021). Polyneuropathies. Atlas of Neuromuscular Diseases: A Practical Guideline, 215-261.DOI: 10.1007/978-3-030-63449-0_12.
Novello, B. J., & Pobre, T. (2020). Electrodiagnostic evaluation of peripheral neuropathy.
Burgess, J., Frank, B., Marshall, A., Khalil, R. S., Ponirakis, G., Petropoulos, I. N., ... & Alam, U. (2021). Early detection of diabetic peripheral neuropathy: a focus on small nerve fibres. Diagnostics, 11(2), 165.
Xiong, Q., Lu, B., Ye, H., Wu, X., Zhang, T., & Li, Y. (2015). The diagnostic value of neuropathy symptom and change score, neuropathy impairment score and Michigan neuropathy screening instrument for diabetic peripheral neuropathy. European Neurology, 74(5-6), 323-327.
Wang, X., Kuzmicheva, O, Spranger, M, Graser, A.(2015)“Gait feature analysis of polyneuropathy patients”. IEEE International Symposium on Medical Measurements and Applications (MeMeA) Proceedings. doi:10.1109/memea.2015.7145172.
Meng, Y., Preston, F. G., Ferdousi, M., Azmi, S., Petropoulos, I. N., Kaye, S., ... & Zheng, Y. (2023). Artificial Intelligence Based Analysis of Corneal Confocal Microscopy Images for Diagnosing Peripheral Neuropathy: A Binary Classification Model. Journal of Clinical Medicine, 12(4), 1284.doi: 10.3390/jcm12041284.
Haque, F., Reaz, M. B. I., Chowdhury, M. E. H., Kiranyaz, S., Abdelmoniem, M., Hussein, E., .shaat, M., Ali, S. H. M., Bakar, A.A., Srivastava, G., Bhuiyan, M.A.S., Mokhtar, M. H. H., & Kurniawan, E. (2022). “Evaluating Performance of Machine Learning Models for Diabetic Sensorimotor Polyneuropathy Severity Classification using Biomechanical Signals during Gait”. arXiv preprint arXiv:2205.10581.doi: 10.48550/arXiv.2205.10581.
Lacomis, D. (2002). Small‐fiber neuropathy. Muscle & Nerve: Official Journal of the American Association of Electrodiagnostic Medicine, 26(2), 173-188.
Gerawarapong, C. (2015). Association of peripheral autonomic neuropathy and sympathetic skin response in the patients with diabetic polyneuropathy: a pilot study in Thailand. J Med Assoc Thai, 98(12), 1222-30.
Shahani, B. T., Halperin, J. J., Boulu, P. H., & Cohen, J. (1984). Sympathetic skin response--a method of assessing unmyelinated axon dysfunction in peripheral neuropathies. Journal of Neurology, Neurosurgery & Psychiatry, 47(5), 536-542.
Messina, I., Palmieri, A., Sambin, M., Kleinbub, J. R., Voci, A., & Calvo, V. (2013). Somatic underpinnings of perceived empathy: The importance of psychotherapy training. Psychotherapy Research, 23(2), 169-177.
Nazlıel, B., Arıkan, Z., Irkeç, C., & Karakılıç, H. (2007). SSR abnormalities in chronic alcoholics. Addictive Behaviors, 32(6), 1290–1294. doi:10.1016/j.addbeh.2006.08.002 .
Oishi, M., Mochizuki, Y., Suzuki, Y., Ogawa, K., Naganuma, T., Nishijo, Y., & Mizutani, T. (2002). Current perception threshold and sympathetic skin response in diabetic and alcoholic polyneuropathies. Internal medicine, 41(10), 819-822.
Lin, X., Chen, C., Liu, Y., Peng, Y., Chen, Z., Huang, H., & Xu, L. (2021). Peripheral Nerve Conduction And Sympathetic Skin Response Are Reliable Methods to Detect Diabetic Cardiac Autonomic Neuropathy. Frontiers in Endocrinology, 12, 709114.
Castro, J., Miranda, B., Castro, I., de Carvalho, M., & Conceição, I. (2016). The diagnostic accuracy of Sudoscan in transthyretin familial amyloid polyneuropathy. Clinical Neurophysiology, 127(5), 2222–2227. doi:10.1016/j.clinph.2016.02.013
Kucera, P., Goldenberg, Z., & Kurca, E. (2004). Sympathetic skin response: review of the method and its clinical use. BRATISLAVSKE LEKARSKE LISTY., 105(3), 108-116.
Greco, A., Valenza, G., Lanata, A., Scilingo E, & Citi, L.(2016) "cvxEDA: a Convex Optimization Approach to Electrodermal Activity Processing". IEEE Transactions on Biomedical Engineering. doi:10.1109/tbme.2015.2474131.
Kamath, C. (2013). “Teager energy based filter-bank cepstra in EEG classification for seizure detection using radial basis function neural network”. International Scholarly Research Notices.doi:10.1155/2013/498754.
Benedek, M., & Kaernbach, C. (2010). “A continuous measure of phasic electrodermal activity”. Journal of neuroscience methods, 190(1), 80-91.
Li, M., & Narayanan, S. (2010, August). “Robust ECG biometrics by fusing temporal and cepstral information”. In 2010 20th International Conference on Pattern Recognition,pp. 1326-1329 IEEE.doi:10.1109/icpr.2010.330.
Posada-Quintero, H.F.; Chon, K.H. Innovations in Electrodermal Activity Data Collection and Signal Processing: A Systematic Review. Sensors 2020, 20, 479. https://doi.org/10.3390/s20020479.
Shukla, J., Barreda-Angeles, M., Oliver, J., Nandi, G. C., & Puig, D. (2019). Feature extraction and selection for emotion recognition from electrodermal activity. IEEE Transactions on Affective Computing, 12(4), 857-869. doi:10.1109/taffc.2019.2901673.
Jang, J. S. (1993). ANFIS: adaptive-network-based fuzzy inference system. IEEE transactions on systems, man, and cybernetics, 23(3), 665-685.
Yadollahpour, A., Nourozi, J., Mirbagheri, S. A., Simancas-Acevedo, E., & Trejo-Macotela, F. R. (2018). Designing and implementing an ANFIS based medical decision support system to predict chronic kidney disease progression. Frontiers in physiology, 9, 1753.
Wang, J., Gao, X. Z., Tanskanen, J. M., & Guo, P. (2012, November). Epileptic EEG signal classification with ANFIS based on harmony search method. In 2012 Eighth International Conference on Computational Intelligence and Security (pp. 690-694). IEEE.
Mathur, N., Glesk, I., & Buis, A. (2016). Comparison of adaptive neuro-fuzzy inference system (ANFIS) and Gaussian processes for machine learning (GPML) algorithms for the prediction of skin temperature in lower limb prostheses. Medical engineering & physics, 38(10), 1083-1089.
Karaboga, D., & Kaya, E. (2019). Adaptive network based fuzzy inference system (ANFIS) training approaches: a comprehensive survey. Artificial Intelligence Review, 52, 2263-2293.
Haque, F., Reaz, M. B. I., Chowdhury, M. E. H., Ezeddin, M., Kiranyaz, S., Alhatou, M., ... & Srivastava, G. (2022). Machine Learning-Based Diabetic Neuropathy and Previous Foot Ulceration Patients Detection Using Electromyography and Ground Reaction Forces during Gait. Sensors, 22(9), 3507.
Haque, F., Bin Ibne Reaz, M., Chowdhury, M. E. H., Srivastava, G., Hamid Md Ali, S., Bakar, A. A. A., & Bhuiyan, M. A. S. (2021). Performance analysis of conventional machine learning algorithms for diabetic sensorimotor polyneuropathy severity classification. Diagnostics, 11(5), 801.
Haque, F., Reaz, M. B., Chowdhury, M. E., Kiranyaz, S., Ali, S. H., Alhatou, M., ... & Srivastava, G. (2022). Performance Analysis of Conventional Machine Learning Algorithms for Diabetic Sensorimotor Polyneuropathy Severity Classification Using Nerve Conduction Studies. Computational Intelligence and Neuroscience, 2022.
Haque, F., Reaz, M. B., Chowdhury, M. E., Hashim, F. H., Arsad, N., & Ali, S. H. (2021). Diabetic sensorimotor polyneuropathy severity classification using adaptive neuro fuzzy inference system. IEEE Access, 9, 7618-7631.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Classification of sympathetic skin response based on the morphologic features and Adaptive Neuro Fuzzy Inference system( ANFIS)

Status:

Version 1

Abstract

Figures

I. INTRODUCTION

II. METHODOLOGY

III. Result

IV. Discussion

V. Conclusion

Declarations

Figure 1

Figure 4:

References

Additional Declarations

Status:

Version 1