Blood Glucose Level Prediction Using Least-Square Support Vector Machine and Fuzzy Clustering

Essay, Pages 7 (1535 words)

Views

Hybridization involves combining two or more different approaches, either at the preprocessing, feature extraction, or learning stage when looking for improved performance. The majority of the BG prediction models involve the hybridization of physiological (compartmental) models along with different machine learning techniques. Regarding support vector regression, for example, Plisetal combined support vector regression along with a physiological model, where the latter generates informative input features to be used to train the SVR model.

Furthermore, Georga etal combined support vector regression with compartmental models, which are used to quantify the absorption of subcutaneously administered insulin, glucose from the gut following a meal, and the effects of exercise on plasma glucose and insulin dynamics.

Don't use plagiarized sources. Get your custom paper on

“ Blood Glucose Level Prediction Using Least-Square Support Vector Machine and Fuzzy Clustering ”

Get high-quality paper

NEW! smart matching with writer

Regarding the hybridization of an artificial neural network with other approaches, some researchers have reported success in this direction. For example, Mougiakakouetal. [50] combined an artificial neural network with a compartmental model, where the latter is used to estimate the effect of food on BG levels and the influence of injected insulin on plasma insulin concentration; this output along with the previous BG measurements were used to train the ANN model.

Mougiakakouetal further investigated the combination of a recurrent neural network along with three compartmental models, which estimated the effect of short-acting (SA) insulin intake on blood insulin concentration, intermediate-acting (IA) insulin intake on blood insulin concentration, and, carbohydrate intake on BG absorption from the gut. Zecchinetal combined an artificial neural network and a physiological model to exploit meal information to be used along with the CGM data.

Moreover, Zecchinetal further explored the applicability of a jump neural network, which is feed by a meal physiological model and 30 CGM data, and compared their result with a previously proposed artificial neural network Briegeletal explored a nonlinear state space model for modeling an individual BG dynamic using a compartmental model and an artificial neural network.

Furthermore, Otto developed a hybrid model combining an artificial neural network and fuzzy logic, where the fuzzy logic was used to approximate food, insulin, and the level of exercise.

Several researchers have attempted to hybridize genetic programming along with physiological models. For example, Contreras etal. developed a hybrid model using a genetic programming-based algorithm known as grammatical evolution and a physiological model. Self-organizing maps (SOMs) have been used to develop a hybrid model along with a physiological model.

For example, Zarkogiannietal used the physiological model to simulate the subcutaneous insulin kinetics and glucose absorption from the gut into the blood, which are in turn fed into the SOM. Jankovicetal developed a two-layer (prediction and correction layer) online adaptive personalized BG prediction model. The prediction layer consisted of an autoregressive model with external input (ARX) and an artificial neural network, which made the first estimates and then the output was further optimized in the second (correction) layer through an extreme learning machine (ELM). Figure 2 shows the simulation for the blood glucose level prediction using the proposed algorithm in the presented paper.

Simulation Results

By considering the mean squared error (MSE), the performance of the proposed algorithm could be evaluated. Fig. 3 shows the predicted glucose levels versus actual levels. As shown, LSSVM shows good performance in predicting blood glucose levels using FCM clustering. The MSE in this experiment is equal to 1.1849. This test was also done without FCM clustering. In this case, the MSE is 1.4019, which is slightly more than when FCM clustering is applied to the data. The γ and σ parameters of LSSVM were considered 100 and 50, respectively.

For a better comparison of the results of the SVM performance with the FCM algorithm and SVM without the FCM algorithm, 10 random samples from each of the parameters γ and σ are produced that are used in common for both cases. The results are presented in Table 2. MSE is calculated as the average for both modes. According to the results of table 2 and the implementation of the LSSVM regression algorithm with and without FCM clustering, it could be concluded that fuzzy clustering led to the reduction in prediction error of the blood glucose level. In fact, fuzzy clustering data resolve uncertainties in the model made by LSSVM.

Therefore, the importance of some data is considered more than other data and the effect of noise and inadequate data is reduced in the prediction and estimation of output. As a result, irrelevant and inadequate data would have less effect on prediction. On the other hand, LS SVM is much faster than SVM in calculating the model by data as well as optimization of the cost function. LSSVM training is simpler than neural networks in terms of regression and, does not use local optima to build models unlike the neural network.

Table 2. MSE Values For Fuzzy And Non-Fuzzy Support Vector Machine With Different Values of And

MES	LSSVM
19.40	57.11
2.65	3.28
70.36	82.47
1.17	1.59
98.28	67.72
1.13	1.77
80.85	99.94
0.99	1.25
70.65	96.20
1.06	1.35
49.01	6.82
7.87	39.00
12.34	36.67
4.39	5.62
66.82	55.30
1.44	2.37
37.17	26.91
2.62	5.34
14.86	60.13
3.06	3.42

Average

MES 1.52 2.10

The results of this method were compared to other common methodologies in order to better evaluate the performance of the proposed algorithm to predict blood glucose levels. Table 3 shows a comparison between these methods. Therefore, the good performance of the algorithm proposed in the paper is proved according to Table 3 and results obtained from the respective papers. However, it is difficult to justify different algorithms of prediction and regression without testing algorithms with the same data of experimental data. In fact, the urgent need is felt to establish a set of standard data to predict blood glucose levels so that various algorithms can be properly assessed. In this case, it provides more reliable results.

Optimization of the Hybrid Svm and Fcm Methodologies (Proposed Method)

Optimization of σand γ coefficients using a genetics algorithm is performed in this section. A standard genetics optimization algorithm with appropriate crossover and mutation has been applied to optimize the quantity of σand γ. An evolutionary algorithm (EA) is a biologically inspired approach to problem solving [96]. The two most used variants of EA in BG prediction and modeling approaches are genetic programming (GP) and genetic algorithms (GA).

Hidalgoetal used a genetic programming-based symbolic regression known as grammatical evolution to develop an individualized model of BG dynamics. Moreover, Contreras used the grammatical evolution approach to develop a standalone BG prediction model. Furthermore, Hidalgoetal. [69] assessed the performance of different predictors, genetic programming, random forests, k-nearest neighbors, and grammatical evolution along with a new enhanced modeling algorithm, a variant of grammatical evolution that uses optimized grammar, and a variant of tree-based genetic programming that uses a three-compartment model for carbohydrate and insulin dynamics.

In this study, the number of considered population for each generation is 50 and the number of created generations to achieve the optimal result is 25. The considered cost function to achieve the optimal point will be the root mean square error (RMSE) in predicting the level of the blood's glucose; therefore the chromosomes with less RMSE will have more fitness value as well as more chance to be selected. In addition, the elite operators are selected 4% for the reproduction of the population in each generation.

Considering the selection of the 50-chromosome population in each generation, two chromosomes will directly migrate from the previous generation to the next one. Single-point crossover operator with 80% selection probability is chosen to apply the crossover process in the middle point of chromosomes. Thus, each chromosome with an 80% probability to be selected and two selected chromosomes cut from the middle and their genomes were replaced with each other. The mutation operator is applied with the selection probability of 0.02 or 20 percent. Therefore, we consider the following specifications for genetic algorithms:

Population-size=50

N-generation=25

Elite-count=0.02*populationsize

Let's optimize the support vector machine parameters by genetics algorithm with the above particulars which utilized for estimation and prediction of glucose level. Figure 4 illustrates the reduction of the applied cost function in each generation achieving optimal points.

According to the above image, the achieved optimized amounts are: Figure 5 illustrates the estimation of glucose level based on the achieved optimized amounts.

With the optimization of the genetic algorithm . In addition, Figure 6 illustrates the amounts of cost function per chromosomes of the last generation.

Let's optimize parameters of support vector machine considered in the which has been used to predict and estimate the glucose level by genetics algorithm. Figure 7 illustrates the reduction of the cost function applied in each generation until achieving the optimized point in this method.

According to the above image, the achieved amount was . Figure 8 illustrates the estimation of glucose level based on the achieved optimized amounts.

Finally, we could obtain optimized amounts of σand γ using a genetics algorithm and a more accurate prediction for blood glucose of diabetic patients while achieving the minimum error.

Conclusion

This paper presents a new algorithm to predict blood glucose levels of diabetic patients using a least-squares support vector machine regression based on Fuzzy C-Means Clustering. Fuzzy clustering was used to weigh data and reduce inappropriate and noise data for regression. Testing on the real data showed the proper functioning of the proposed algorithm compared to other common methods.

In fact, eliminating unrelated and additional characteristics of data, a selective characteristic algorithm can help to improve learning by decreasing unpleasant dimensional effect, increasing overall capabilities and promoting the interpretability of the model. The use of signal processing algorithms to receive information within data and selection of appropriate characteristics from among them for regression and better forecast can be considered as a ground for future studies.