Assessment and Evaluation of Diabetic Foot using Biothesiometry and Artificial Neural Networks
Correspondence Address :
Dr. Mahesh Veezhinathan,
Associate Professor, Department of Biomedical Engineering, SSN College of Engineering, Chennai, Tamil Nadu, India.
E-mail: maheshv@ssn.edu.in
Introduction: Diabetes is a common disorder that is prevalent in the general population. As it advances, it causes a multitude of consequences, some of which are fatal. Diabetic Neuropathy (DN) is one such illness that causes nerve damage, and it is quickly recognised and diagnosed using a technique known as biothesiometry.
Aim: To create an assessment tool based on an Artificial Neural Network (ANN) that evaluates diabetic foot based on Vibration Perception Threshold (VPT) values.
Materials and Methods: This experimental and predictive study was conducted at Sri Sivasubramaniya Nadar College of Engineering, Chennai, Tamil Nadu, India, from January 2022 to March 2022 and was done using VPT values for 696 controlled and diabetic groups selected by purposive sampling. The VPT was measured by a biothesiometre. A metal probe was placed under the foot of the person and the voltage was increased gradually from zero and the transition from no vibration to the point of vibration is marked as VPT. Average of three measurements were taken to calculate the VPT value of the given patient. This involves the VPT value which helps in the assessment of severity of the condition. The recorded data was fed as an input to the ANN model which predicted the average VPT value of the left and right foot. Furthermore, the ANN model was assessed my means of statistical measures and parameters.
Results: The results of the study, confirmed the correlation between the values of VPT acquired at different points of the foot and the coefficient for the left and right foot was found to be 0.99549 and 1.0000 respectively. Furthermore, the efficiency of the proposed ANN model was assessed using statistical measures like Mean Square Error (MSE), Mean Absolute Error (MAE), Square Sum Error (SSE) and Coefficient of determination (R2). The predicted values were very close to the experimental VPT results, and the correlation coefficient R are 0.99549 and 0.99975 for the left and right feet, respectively, which shows the best settlement.
Conclusion: The study concluded that VPT acquired from the foot of diabetic patients is useful in categorising the level of severity of DN. Furthermore, the results of ANN model proved that, there exists a strong correlation between the average VPT values of left and right foot and those that are acquired from different points of the foot such as big toe, first metatarsal, third metatarsal, fifth metatarsal, In step and heel, which concluded the study to be effective in the assessment and diagnosis of DN.
Diabetes, Vibrometre, Vibration perception threshold
Diabetes is common in the general population, particularly among people aged 40-60 years. According to Saeedi P et al., the incidence of diabetes would rise to 537 million by 2030 (1). In such people with diabetes, damage to nerves in foot and legs can occur which can range from severe to fatal. The condition is called Diabetic Neuropathy (DN), this condition has numerous complications associated with it such as foot ulcerations, leg amputation and pain and it is mostly identified in the later stage (2). The early diagnosis of such conditions can be done using the procedure known as biothesiometry. A biothesiometre is an electromechanical device that is used to measure the VPT values that serves as a simple technique to diagnose peripheral neuropathy in people (3). In most cases DN is detected by using VPT values, wherein it gives a quantitative measure of the vibration that is being perceived by a patient. Aruna BM and Haragopal R, conducted a study on 60 people, 30 of whom were diabetic and 30 of whom were not; the findings were reached by calculating the mean VPT between the two groups, and there was a statistical significance between the groups (4). In a related study based on VPT, Dash S and Thakur AK; Kaur J and Batra AP found no significant difference between individuals with clinical neuropathy and those without clinical neuropathy, concluding that the severity of the condition cannot be detected by biothesiometry (5),(6). Furthermore, in a study conducted by Ashok BH et al., it was discovered that obtaining thermal images in addition to VPT data proved to be effective in producing an accurate diagnosis report for plantar lesions in the foot (7).
The presence of neuropathy is a crucial factor in evaluating if a foot is particularly prone to painless damage, which can progress to a persistent and spreading ulcer. This has major ramifications and might result in massive amputations below or above the ankle and knee. Neuropathy can be qualitatively assessed by means of using a tuning fork, hot and cold sensation and pricks. But quantitative assessments of the protective sensation in the foot should be done in order to report any incremental risks and damage. Many procedures, such as nerve conduction investigations, are available but unfeasible due to their high cost (8). Hence, simple techniques like using vibration perception prove very useful and effective. The device that produces vibration at specific voltage levels is known as a Vibrometre/Biothesiometre.
In the present study, VPT values were acquired from various point of the foot of diabetic and non diabetic individuals with the help of a biothesiometre and an ANN model based on the Levenberg Marquardt algorithm was constructed with the acquired VPT values as input and the average VPT value as output. ANN is increasingly being used as a method to diagnose DN, but many algorithms remain untested. Hence, the present study, proposes the adoption of a new ANN model and the proposed ANN was tested for its performance based on parameters such as MSE, MAE, SSE and coefficient of determination R2.
The present experimental and predictive survey was conducted among 696 volunteers at Sri Sivasubramaniya Nadar College of Engineering, Chennai, Tamil Nadu, India, from January 2022 to March 2022. The subjects were selected by purposive sampling technique. The subjects who attended the HCL health camp for VPT measurement were both diabetic and non diabetic. Out of the 696 volunteers 78 were normal subjects where the maximum VPT value for both the foot was found to be less than 15. The minimum age of the participant involved in the study was 35 years and the majority consisted of participants above the age of 50 years. The measurements were made from the subjects as shown in (Table/Fig 1)a,b.
Vibration Perception using Biothesiometre
The vibrometre model that was used in the study is the Vibrometre VPT by Diabetik Foot Care India Pvt. Ltd. The technical specifications of the device are shown in (Table/Fig 2) (9). The overall framework of the work involved in the present study is illustrated in (Table/Fig 3).
Data generation: In the present study, a health camp was organised in a hospital and the data was collected from 696 volunteers using the aforementioned model of vibrometre. All the volunteers were informed about the procedure and an informed consent was obtained from the all the participants involved in the study. The subjects were then asked to sit in a relaxed position. The vibrometre probe was initially placed on the hands in order to provide a sense and impression of vibration. Then the probe was applied at six points on the foot as illustrated in (Table/Fig 4) and the voltage level was simultaneously increased. The points shown in the figure represent the parts of foot namely the (i) big toe, (ii) 1st metatarsal, (iii) 3rd metatarsal, (iv) 5th metatarsal, (v) In step and (vi) heel, respectively. The level of severity was classified and assessed by using the voltage perception range as shown in (Table/Fig 5) (5),(10). which led to categorising the subjects into right mild loss, left mild loss, right moderate loss, left moderate loss, right severe loss and left severe loss as shown in (Table/Fig 6). The procedure was stopped immediately, if the volunteer felt any slight discomfort. Several trials were conducted, and the data generated was transmitted via the RS23C serial interface to the PC.
Artificial Neural Network (ANN) Approach
ANN is a data-driven modelling tool that is used to imitate the structure and functional abilities of a biological neural network, as shown in (Table/Fig 7)a (11),(12). ANN is a parallel distributed processor made up of basic and highly linked processing components known as neurons that process data. The brain network learns and stores that knowledge in the interneuron connection strengths known as synaptic weights (13). The neural model has a set of connecting channels, an adder, and an activation function. The synaptic weights w1, w2,...,wm are multiplied by the input signals x1, x2,...,xm. In order to determine the neuron’s net input, the weighted inputs are summed and multiplied by bias (b) (net input). The net input is then passed via an activation function (F(·)), which generates the output (0) of the neuron as shown in (Table/Fig 7)b. The bias term is used to improve network performance by adjusting the net input of the activation function. The activation function limits the amplitude range of a neuron’s output to a limited value (14).
The net input (netinput) and output (0) of the neuron are expressed as
The most widely used activation functions are the sigmoid and linear functions, and they are described in equations (3) and (4), respectively.
The multilayer feed forward neural network with the back propagation training technique is the most often used ANN. A Multilayer Perceptron (MLP) is made up of an input layer, one or more hidden layers of linear threshold units, and an output layer. The Back Propagation (BP) technique computes the difference between expected and actual outcomes and propagates the mistake back to the network. The BP algorithm modifies the network’s weights and biases at each iteration to decrease error, and the process is repeated until a particular level of accuracy is attained.
Artificial Neural Networks (ANN) application in diabetic foot: In this work, authors have imported an ANN model onto a database of diabetic foot occurrences in order to show the reliability of neural networks in the biomedical sector. Hence, an ANN model with one hidden layer and 25 neurons was built utilising the VPT values of diabetic and non diabetic measured using a biothesiometre. The chosen network design includes eight neurons in the input layer that represent the age, big toe, first metatarsal, third metatarsal, fifth metatarsal, instep, and heel, as well as one neuron in the output layer. The number of hidden neurons that was selected was 23 and 15 in one hidden layer for the left and right foots which is demonstrated in (Table/Fig 8),(Table/Fig 9) respectively.
STATISTICAL ANALYSIS
The coefficient of determination (R2) for the left and right foots were 0.990963 and 0.974916, respectively. This selection process was made upon experimenting all the probability of growing hidden neurons from 1 to 25, R2 from 0.1 to 0.9 for each. Levenberg–Marquardt BP was used as a training function and sigmoid was used as a transfer function. The coefficient of determination was estimated based on network performance, with different modifications to connection weights made throughout training. The input data set was distributed as 70% for training, 15% for testing, and 15% for validation, respectively. The correlation coefficient (R) captured for training, testing, validating, and overall data for left foot (LF) and right foot (RF) ANN models are shown in (Table/Fig 10).
Based on an analysis of (Table/Fig 8),(Table/Fig 9), it can be observed that the proposed Levenberg-Marquardt BP neural network, which has one hidden layer with 23 hidden neurons for the left foot and one hidden layer with 15 neurons in the hidden layer for the right foot, shows superiority in terms of reduced standard statistical errors and decreased overall error rates (Table/Fig 11). As a result, the Levenberg-Marquardt BP neural network has been proposed as the best structural framework for the proposed ANN model. It has eight inputs, one hidden layer, 23 and 15 hidden neurons in the hidden layer for the left and right feet, respectively, and one output layer with one output neuron. The convergence of the VPT value of a diabetic using a biothesiometre and ANN model in terms of MSE, states of convergence, Error Histograms (EH) and the correlation coefficient (R) are depicted in (Table/Fig 12),(Table/Fig 13),(Table/Fig 14),(Table/Fig 15),(Table/Fig 16),(Table/Fig 17),(Table/Fig 18),(Table/Fig 19). From (Table/Fig 12),(Table/Fig 13), it can be observed that the best validation performance of left and right foots was accomplished at 37 and 52 epochs with the value of MSE around 0.032724 and 1.2021e-10 respectively. The corresponding gradients are [0.12792 and 9.5743e-0.8] while the step size Mu for left and right foots are [1e-06 and 1e-14] and at epochs 43 and 52, respectively as displayed in (Table/Fig 14),(Table/Fig 15). Furthermore, the bins near to apt zero line having EH values for the aforementioned cases are 0.03727 and 1.14e-07, respectively are illustrated in (Table/Fig 16),(Table/Fig 17). To scrutinise the capability of the developed ANN model, the regression analysis between the numerical and predicted results were plotted for the stated cases that can be seen in (Table/Fig 18),(Table/Fig 19). The values of R, is always closer to 1, for training, validation, testing, and overall dataset, which confirms that a good agreement lies between the actual and the targeted numerical value. Hence, the proposed Levenberg-Marquardt BP neural network modelling approach is reliable, effective, accurate, fast, and alternative tool particularly in healthcare sector.
Furthermore, to estimate the performance of this ANN configuration, the following standard statistical measures such as MSE, MAE, SSE and coefficient of determination (R2) were adopted.
where Êi measured output using biothesiometre, i predicted value by ANN model, = average measure and Nd number of input data set, respectively.
The error calculations such as MSE, MAE, SSE and the coefficient of determination (R2) of developed ANN model of one hidden layer with 25 neurons are presented in (Table/Fig 11). It should be noted that the suggested ANN model produced better results, and the predicted levels of diabetes severity were in close agreement with data obtained using a biothesiometre, with correlation coefficients of 0.99096 (LF) and 1.0000 (RF) for each.
The ANN was chosen, since it offers an effective and credible solution and has been widely employed in a variety of engineering and science disciplines (15),(16),(17),(18). In the health industry, ANN is a suitable tool for prediction of illness such as malaria and cancer, widely employed in image processing in the medical profession, and finds its application in Decision Support Systems (DSS) (19),(20),(21),(22). Several studies have also demonstrated that ANN is an appropriate method for predicting early diabetes (23),(24),(25). A research work on ANN introduced a network trained with BP to study the inadequate accuracy in diagnosis of diabetes mellitus (26). Another work was conducted, using ANN to identify appropriate factors that disturb the health conditions and to also assess their impact on diabetes (27). A few investigators did the timely prediction of diabetes using a hybrid neural network and logistic regression model (28). Many studies have recently investigated the application of machine learning and deep learning algorithms for diabetes prediction in the early stages, such as the artificial BP scaled conjugate gradient neural network (ABP-SCGNN) algorithm, deep learning model, and ANN algorithms (26),(27),(28).
Limitation(s)
To efficiently handle big data is one limitation of the proposed Levenberg–Marquardt BP neural network model. Thus, the authors plan to implement the newly developed prediction models like Gaussian process regression, and Gradient boosted regression trees via Bayesian inference in future work.
In the present study, it can be concluded that biothesiometry is an effective technique and tool that can be used to assess diabetic foot and evaluate the same by means of VPT values. The subjects were categorised into groups based on the severity of DN and the range of VPT which helped in further analysis and provided information for the construction of the ANN model. The ANN based on the Levenberg-Marquardt algorithm, was utilised to predict the average VPT values of the left and right foot and the value of (R2) determined for both the foot were found, to be highly correlated. The performance of the ANN was tested using the performance validation plot, state of convergence and error histograms, which also showed appealing results, thereby proving the reliability and efficiency of the proposed ANN model.
The authors would like to express their gratitude to Sri Sivasubramaniya Nadar College of Engineering for giving the facility to carry the research work and Dr. Palaniappan for his valuable insights and guidance.
DOI: 10.7860/JCDR/2022/56348.17168
Date of Submission: Mar 15, 2022
Date of Peer Review: Jun 30, 2022
Date of Acceptance: Sep 17, 2022
Date of Publishing: Nov 01, 2022
AUTHOR DECLARATION:
• Financial or Other Competing Interests: None
• Was Ethics Committee Approval obtained for this study? No
• Was informed consent obtained from the subjects involved in the study? Yes
• For any images presented appropriate consent has been obtained from the subjects. Yes
PLAGIARISM CHECKING METHODS:
• Plagiarism X-checker: Mar 28, 2022
• Manual Googling: Jun 30, 2022
• iThenticate Software: Sep 16, 2022 (9%)
ETYMOLOGY: Author Origin
- Emerging Sources Citation Index (Web of Science, thomsonreuters)
- Index Copernicus ICV 2017: 134.54
- Academic Search Complete Database
- Directory of Open Access Journals (DOAJ)
- Embase
- EBSCOhost
- Google Scholar
- HINARI Access to Research in Health Programme
- Indian Science Abstracts (ISA)
- Journal seek Database
- Popline (reproductive health literature)
- www.omnimedicalsearch.com