An effective integrated machine learning approach for detecting diabetic retinopathy

1 Introduction

Several classification techniques of machine learning (ML) algorithms were discussed, and these techniques greatly helped the stakeholders of the medical field for predicting heart disease. A model of the artificial neural network (ANN) was proposed by Dangare and Apte was outperformed with 100% accuracy [1]. To detect anomalies in hyperglycemia classification, techniques such as feedforward ANN, deep belief network, genetic algorithm (GA), support vector machine (SVM), and Bayesian neural network were proposed and implemented [2]. Some of the ML techniques were rarely implemented or not implemented at all. Besides, the accuracy of some ML techniques was lower than the accuracy obtained by DL techniques. Hence, the combined models of ML and deep learning (DL) techniques were discussed to enhance accuracy for diabetes prediction [3]. The principal component analysis (PCA) was discussed to deal with large datasets. The dimensions of these large datasets could be reduced by using PCA to observe the correlation between the attributes and for better interpretability [4]. Diabetic retinopathy would affect eyes. The current trends of disease, mechanisms, and approaches to treat diabetic retinopathy were discussed [5]. A system with ML algorithms namely, k -nearest neighbor (KNN), variants of SVM, and NB was discussed to detect exudates in retinal images automatically. The proposed system detected the exudates with an accuracy of 98.58% that was greater than other existing techniques [6]. The classification techniques such as C4.5, Naïve Bayes (NB), and clustering technique k -means clustering were used to detect the risk factors of diabetes disease complications. The proposed system achieved an average accuracy of 68% [7]. Moth-flame optimization algorithm was discussed, which would improve the accuracy of classification [8]. To detect and classify characteristics such as micro-aneurysms (MA) and hemorrhages in retinal images, a model with convolutional neural networks (CNN) was proposed. The proposed model achieved 95% accuracy for the two-class classification of the dataset size 30,000 images and 85% for the five-class classification of the dataset size 3,000 images [9]. Diseases related to heart, breast cancer, and diabetes were analyzed using ML techniques. This study revealed the significance of predicting the risk factors of diseases [10]. The aforementioned discussions exhibit the role of ML in predicting the symptoms and risk factors of different kinds of chronic diseases. At this moment, the statement “prevention is better than cure” is to be reminded. If the symptoms of the disease are identified before the occurrence, then it will help the people to take necessary precautions. Hence, the ML algorithms have great participation and impact on medical diagnosis.

2 Proposed methodology and implementation

To implement the proposed methodology, the dataset diabetic retinopathy is retrieved from the UCI ML repository. The proposed techniques such as normalization, PCA, and SVM are briefly described as follows. The dimensions of a dataset may consist of different levels of data. If the dataset is directly taken for the computation process, the higher-ordered dimensions may dominate the other lower-ordered dimensions. The obtained result is no way useful for decision-making. So, it is necessary to scale the data to make all the dimensions to be at the same level. The normalization technique scales the data and removes anomalies existing in the data. This is a preprocessing technique generally applied to the dataset before proceeding with the analysis process. The normalized data usually lies between 0 and 1. The main functionality of the PCA is to reduce the dimensions of a dataset. This is also referred to as dimensionality reduction. When a dataset contains more dimensions, several dimensions might be highly correlated. This problem is referred to as multicollinearity. The existence of multicollinearity affects the quality of data analysis. The PCA technique is best to deal with the multicollinearity problem. The main elements in PCA are principal components (PCs). The generation of the number of PCs is based on the number of dimensions given as input. The same number of PCs will be generated for the given number of dimensions, and they are ordered by their variance. PCs with high variance come first and then next higher level and so on. SVM is one of the most widely used supervised ML techniques. It is largely used for classification in high-dimensional data. It can be applied to both linearly and non-linearly separable problems. The key elements of the SVM are hyperplanes. In SVM, the classification of data will be done by identifying the hyperplanes. Support vectors are the vectors that describe the hyperplanes. In Section 2.1, the description of the dataset is given. In Section 2.2, the algorithm and objective function for the moth-flame optimization technique is described. The flow of activities in the proposed model is shown in Figure 1.

Figure 1

Flowchart for the proposed integrated ML approach.

As an initial step, the collected diabetic retinopathy dataset is inputted to the proposed model. The details of the retinal images presented in this dataset are used to classify the images and to decide whether there are any symptoms of the diabetic retinopathy existing. Before proceeding to implement ML algorithms, it is important to normalize the data. This normalization can be done by using a standard scaler. After normalization, the ML algorithms such as DT, NB, RF, and SVM will be individually applied to the dataset. Next, the PCA technique is implemented for reducing the features in the dataset. This feature reduction will enable all the features to be at the same level. It means that the domination of higher-order features will be avoided. Now, the reduced features are inputted into aforesaid ML algorithms. The performance of ML algorithms before and after the implementation of PCA is observed. If high performance is observed, then we can continue with the recently applied technique. Otherwise, to improve the performance of ML algorithms, the resultants of previously applied techniques are feeded into the proposed integrated approach, i.e., SVM + PCA + moth-flame optimization technique. Then, the performance of ML algorithms was evaluated and compared concerning the performance measures discussed in Section 1.1. By looking into the comparative analysis, it can be understood that the ML approach that outperformed among all the others represents the correct classification of class labels.

3 Results and discussion

The implementation of ML algorithms was performed on a diabetic retinopathy dataset retrieved from the UCI ML repository. The performance of the proposed integrated ML model is evaluated using measures such as precision, F 1-score, specificity, accuracy, recall, and sensitivity (12)–(16). Precision describes the correctness and is given in equation (12). Recall/sensitivity represents the wholeness and is expressed in equation (13). F 1-score is described in equation (14). Accuracy characterizes the rightness and is given in equation (15). Specificity is described in equation (16).

where PT = positive (true value), PF = positive (false value), NT = negative (true value), and NF = negative (false value). The simulation results of ML algorithms and corresponding performance measures are described as follows: First, the dataset has experimented with four popular ML algorithms, namely, DT, NB, RF, and SVM. Figure 2 depicts the performance of these algorithms.

Figure 2

Performance of ML algorithms on the original dataset.

From Figure 2, it can be observed that the DT classifier has achieved 57% of precision, recall, and F 1-score, and 57.1, 54.5, and 59.2% of accuracy, sensitivity, and specificity, respectively. The NB classifier has achieved 64% of precision, 63% of recall and F 1-score, and 63.2, 64.2, and 62.4% of accuracy, sensitivity, and specificity, respectively. The RF classifier has achieved 70% of precision, recall, and 69% of F 1-score and 68.8, 76.2, and 63% of accuracy, sensitivity, and specificity, respectively. The algorithm SVM has achieved 79% of precision, 76% of recall and F 1-score, and 76.1, 89.4, and 65.3% of accuracy, sensitivity, and specificity, respectively. When these performances are compared with each other, it is observed that the SVM is outperformed.

The dataset is then fed in to the PCA for dimensionality reduction. The PCA has reduced the dataset from 20 dimensions to 12 dimensions. These reduced features of the dataset are given as input to the aforementioned classifiers. Figure 3 depicts the results obtained after applying ML algorithms on reduced dimensions.

Figure 3

Performance of ML algorithms after reducing dimensions using PCA.

From Figure 3, it can be observed that the DT classifier has achieved 67% of precision, recall, and F 1-score, and 67.09, 66.3, and 67.6% of accuracy, sensitivity, and specificity, respectively. The NB classifier has achieved 61% of precision, 60% of recall and F 1-score, and 59.7% of accuracy, sensitivity, and specificity. The RF classifier has achieved 69% of precision, recall, and F 1-score, and 69.2, 67.8, and 70.4% of accuracy, sensitivity, and specificity, respectively. The SVM has achieved 71% of precision, 66% of recall and F 1-score, and 65.8, 81.3, and 55.7% of accuracy, sensitivity, and specificity, respectively. Figure 3 shows that the performance of the classifiers NB, RF, and SVM has been reduced when PCA is applied. But, the performance of DT has been enhanced with PCA. The main reason for the degradation of performance is dimensionality reduction. It is understood that these classifiers work better with more samples and dimensions.

To achieve better performance of PCA-based classifiers, the moth-flame optimization technique is applied. This technique eliminated the attributes that are affecting the performance of the classification model negatively. It chooses the optimal features/attributes that have a positive impact on performance. By implementing the moth-flame technique, the 12 dimensions were further reduced to 9. Based on this reduction, it is understood three dimensions are affecting the performance of the model. These three dimensions are eliminated by the moth-flame algorithm. Figure 4 depicts the result of the model with PCA and moth-flame optimization techniques.

Figure 4

Performance of ML algorithms after implementation of moth-flame optimization technique.

Figure 4 shows that the DT classifier has achieved 68% of precision, recall, and F 1-score, and 68.3, 67.4, and 69.2% of accuracy, sensitivity, and specificity, respectively. The NB classifier has achieved 62% of precision, 63% of recall and F 1-score, and 61.8% of accuracy, sensitivity, and specificity is 62.1%. The RF classifier has achieved 80% of precision, 79% of recall, and F 1-score, and 78.2, 85.8, 75.1% of accuracy, sensitivity, specificity, respectively. The SVM has achieved 86% of precision, recall, and F 1-score, and 86.3, 94.2, and 75.2% of accuracy, sensitivity, and specificity, respectively. From Figure 4, it is evident that the performance of all the classifiers except NB has improved dramatically with the addition of the moth-flame optimization technique. This technique eliminated the dimensions that impact the performance of the classifiers negatively. It is observed from the proposed model that the integration of SVM, PCA, moth-flame outperformed than all other ML algorithms. The performance of all ML algorithms is summarized in Table 2. In the table, PC, MF, Prscn, Rcl, F 1, Acc, Sens, Spec refer to priniciple component analysis, moth-flame, precision, recall, F 1-score, accuracy, sensitivity and specificity respectively. It can be noticed that the proposed integrated model of SVM, PCA, moth-flame has achieved high performance than the other ML algorithms. The high performance of this model represents that the classification of class labels has been achieved correctly.

4 Conclusion

To identify the diabetic retinopathy, this article proposed an integrated approach of ML algorithms and achieved high performance. The dataset that is retrieved from the UCI ML repository is used for the proposed approach. The key observations of the proposed integrated approach are as follows:

1. From Figure 2, it is observed that when the ML algorithms are implemented individually, SVM is outperformed than other ML algorithms.

2. From Figure 3, it is evident that the reduction of the dimensions using the PCA technique has negatively influenced the performance of majority ML algorithms.

3. From Figure 4, it is understood that the integration of SVM, PCA, and moth-flame optimization techniques improved the performance of classification and identified the class labels correctly.

4. From Table 2, it is easy to interpret and compare the performances achieved by the implemented ML algorithms.