Exploring Machine Learning Algorithms for the Prediction of Dengue: A Comprehensive Review

In the algorithm 1, the required libraries for SVM are imported and then the dataset from a CSV file are loaded using the read_csv function from Pandas. The function train_test_split from Scikit-Learn is then used to split the dataset into training set and testing set. Next, an SVM model using the SVC class from Scikit-learn are created, specifying a linear kernel. The model object's fit method is used to train the SVM model on the given training set. In this algorithm, the trained model is utilized to forecast the target variable for the testing set. The function accuracy_score from Scikit-learn is used to relate the predicted values with the actual values in the testing set in order to evaluate the model's accuracy.

To optimize the model's performance, the dataset must be checked whether it is appropriately prepared for the SVM algorithm, including handling missing data, scaling the predictor variables, and dealing with any imbalanced classes in the target variable. Dourjoy et al. have proposed a comparative analysis of dengue fever prediction using machine learning algorithms which aimed to develop an SVM-based system for initial prediction of dengue fever by using symptoms reported by the patients [22]. The study used a dataset of patients with suspected dengue fever. The dataset included information on various symptoms such as fever, headache, joint pain, and nausea. The authors used an SVM algorithm to predict whether a patient had dengue fever based on the presence of specific symptoms. It also sought to incorporate other parameters associated to the disease that are required to improve the model's ability to accurately predict outcomes when used in regions with similar climatic conditions and when weather variables like temperature, humidity and total rainfall are not significantly different. Their model described that the SVM-based system was able to detect dengue fever with an accuracy of 94.4%. The authors identified several symptoms that were strongly associated with dengue fever such as fever, headache and rash which were included in the SVM model.

2.4 Decision tree

A Supervised learning technique known as Decision Tree (DT) makes predictions and decisions by learning from the previous history of data. It splits the dataset as two major groups based on the attribute value or random value and then it compares both the groups to find the best split point. Few approaches like Classification and regression trees (CART), chi-squared automatic interaction detection (CHAID) and Iterative dichotomiser (ID) are used to construct a DT. DTs are employed in decision-making, object classification, result prediction, and data analysis. DTs are used in medicine to find new drugs and diagnose diseases [23-25].

Several researchers have used decision tree algorithms for early prediction of dengue fever from symptoms for patients on the base of geographical locations. Kalansuriya et al. [26] aimed to build a model for location-based dengue prediction of dengue fever based on symptoms reported by patients. A methodology has been put out by Pravin et al. to identify dengue patients at an early stage. They have retrieved patient data from suspected dengue patients using IoT sensors in the framework. This framework has classified people records on the basis of symptoms [27].

Rao et al developed a prediction model in real time which is used to identify dengue at the initial stage. It also minimizes the number of false negative and false positive results of the prediction model. It is then trained and tested with the datasets. The implementation of decision tree on prediction model has achieved highest accuracy [28]. The propose model can be used as a universal tool for dengue diagnosis.

2.5 Naive bayes

A probabilistic classifier namely Naive Bayes (NB) is modified with some simplifications that is built on the Bayes theorem. The features in the NB classifier are presumably conditionally independent based on the label. NB has been used for weather forecasting, spam filtering, and medical diagnosis [29, 30].

Naive Bayes classifiers are built on feature selection algorithms, which determine the conditional likelihood of features belonging to a class, which are frequently used for text classification in machine learning [31]. After determining features using a recognised feature selection approach, an auxiliary feature technique is utilised to choose an auxiliary feature that can reclassify the text space targeted at the supplied features. The appropriate conditional probability is then altered to improve classification accuracy.

Harumy et al. [32] discussed about the different factors responsible for spreading dengue in the domain of his country by using Artificial Neural Network (ANN), Naïve Bayes, Backpropagation, Regression. The proposed model has three layers. In the first layer, data collection and identification of variables are performed. While second phase is of comparison of prediction of dengue cases. Third phase identifies factors that most influences. The outcome reveals that the proposed prediction model has an accuracy rate of 87.16%.

Peng et al. [33] proposed a prediction model using time series which was active to dengue and chikungunya. The following methods namely Gaussian processes, Kernel Ridge Regression are used in the findings. The study concludes that the Kernel Ridge Regression method has the maximum performance for predicting dengue. It is also capable of modelling linear and nonlinear relationships between predictor variables and outcomes.

During the period 2013 to 2017 a survey was conducted in five areas in and around   Malaysia which saw the greatest incidence of dengue sickness. The survey evaluated the most reliable machine learning model for predicting dengue outbreaks [34]. The factors that affect the climate like temperature, humidity, wind speed and rainfall were included in every model. According to the data, the SVM showed the best capacity to predict with an accuracy of 70%, sensitivity of 14%, specificity of 95% and precision of 56%. However, with reference to the imbalanced data (original data), the testing sample's SVM (linear) sensitivity increased to 63.54% from 14.4%. The SVM model's most significant predictor was the week of the year. This study serves as an example of the potential machine learning for the forecasting of dengue outbreaks.

The datasets include symptoms such as blood pressure, fever, heart rate, temperature and platelets. It also includes symptoms that both the patient and the doctor notice.  The use of the Naive Bayes classification algorithm for predicting dengue sickness is described in the consultation. Real-time data is accepted and saved on the server through the analysis of the patient's earlier medical information [35]. Additionally, the dataset was compared to the acceptable real-time data, which provides the likelihood that dengue will occur.

A machine learning-based classification algorithm that uses a patient's likelihood of being admitted to the paediatric intensive care unit as a surrogate for the severity of dengue. A huge number of machine learning techniques were well trained and validated with the help of Stratified 5-Fold cross-validation. The finest model was chosen and then it was evaluated on a separate test set. The findings of the Cross-validation reveal that SVM with a Gaussian Kernel beat the other existing models and with Area under the ROC curve (AUC ROC) achieve a score of 0.81. Later results across the test set displayed an average AUC ROC scores a value of 0.75. The findings of validation and testing are encouraging and also encourage additional study and advancement [36].

2.6 Neural networks

Neural networks (NN) are a collection of layers of neurons and are modelled after the organisation of the human brain. Input, hidden, and output layers are how these neurons are organised. Input x_i, weights w_i, and bias (b) are collected as a numerical value. The input to the neuron Sum is determined as defined in Eq. (1).

$\operatorname{Sum}=\sum_{i=1}^n x_i w_i+b$              (1)

The neurons in the buried layers of the brain receive this Sum value as input. The threshold function t uses the estimated value to activate a neuron. Some known standard neural networks are Artificial Neural Network (ANN), Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN). A number of neurons are coupled by back-propagation and feed-forward forms in ANN [37, 38]. CNN employs a multi-layer perceptron variant [39]. ANN models rely heavily on hardware but the CNN model requires a large amount of training data to function effectively and does not record the object's location or orientation. Voice, handwriting, and picture recognition are all visual patterns that these networks can be trained to recognise [40, 41]. Modern NNs are hence much more effective than earlier versions.

Gambhir et al. [42] has developed a model for an early detection of dengue disease taking a dataset of 110 patients in Delhi from different hospitals during 2015-16. The dataset is further classified in to two different classes namely dengue positive and dengue negative. In this study, the dataset is composed of 85 patients as dengue positive and rest as negative. The author proposed model has been divided into three phases. The first phase of propose prediction model has performed data collection and pre-processing work. In the second phase, the K-cross fold validation method has performed for reducing unwanted data. Further machine learning technologies specifically Decision Tree (DT), Artificial Neural Network (ANN) and Naïve Bayes (NB) are used in the proposed model. This K-Cross Fold validation method validates results produced by three machine learning algorithms. Researcher has found that ANN based diagnostic model has given better performance in the study than the Naïve Bayes (NB) and Decision Tree (DT). In third phase, the proposed model is capable of taking decision whether a patient is found dengue positive or dengue negative. The results shows that ANN -based prediction model has achieved 79 % accuracy, 55.55 % sensitivity, 88.5 % Specificity. On other hand, Naïve bayes -based prediction model has achieved 76.36 % accuracy, 47.66 % sensitivity, 84.88 % Specificity while Decision Tree -based prediction model has achieved 73.63 % accuracy, 41.66 % sensitivity, 82.55 % Specificity. The study's conclusion is that an ANN-based prediction model is ideal for identifying dengue patients at an early stage.

The public health sector's absence of a robust medical infrastructure has a particularly negative impact on rural communities. As members of the lower middle class or those who fall below the poverty line, a significant portion of the population cannot afford private hospitals. As a result, the burden on the public health sector has multiplied. In the event of an epidemic, the mortality rate is significant since the majority of the populace lacks access to healthcare and prompt treatment. A proper medical infrastructure must be set up in case of an outbreak because a substantial portion of the people in our nation lives in rural areas. As one of the most frequent epidemics in our nation and the cause of a high fatality rate, dengue needs to be predicted well in advance so that medical resources may be set up on time and the fatality rate is lowered. Although dengue does not require a particular course of therapy or a lengthy course of treatment, it is crucial to identify and address the condition quickly.

The technology seeks to forecast the outbreak of an epidemic (dengue), allowing the health sector to prepare the necessary resources in advance. Recurrent neural networks will be used by the system to produce predictions. Climate, pollution, and the numbers of people who have been diagnosed with dengue in prior years are only a few of the variables included in the forecast process [43]. The information was gathered from a number of official sources as well as medical facilities. In the future, if comparable environmental circumstances exist, the model will learn from this data and forecast the likelihood of an outbreak. The outcome of the potential outbreak will be displayed by using Google heatmaps to indicate the potential dengue-endemic areas.

In order to implement more effective vector management techniques, it is critically necessary to develop mechanisms that can more accurately predict dengue cases. A unique hybrid architecture that combines the advantages of recurrent neural network and convolutional neural networks is used to identify the weekly dengue incidence [44]. In dengue predicting method, this hybrid architecture beats the other frequently used deep learning models. The suggested models are also compared to cutting-edge studies in the literature, illuminating how significantly dengue forecasting can benefit from our hybrid models' use of recurrent networks and convolutional layers.

2.7 Ensemble techniques

In machine learning techniques, an ensemble of basic models is combined to provide an ideal prediction model and/or learner that enhances the outcome. In Bagging the training data is divided into multiple copies, and after that, each copy is used to train a different classifier. By merging the predictions made by various classifiers, the final prediction is formed. Homogenous weak classifiers are constructed in series to correct the classifier errors [45]. The first model was developed using the whole dataset, while the second model used the predictions from the first model. All of the weak classifiers are combined to create the final classifier. Staggering process builds heterogeneous weak classifiers and the stacking ensemble approach works using a meta model. This meta model develops the ability to mix base model predictions. The level 0 classifier output is used as training data for the level 1 classifier, which will then try to approach the same goal function, according to the stacking combining technique. An ensemble learning technique called a random forest consists of numerous decision trees, or different classifiers [46]. Both classification and regression issues can be resolved using it. The following steps are used to make a random forest: Collect the required input data, then for each tree select the randomised subsets of the data as candidate inputs, now the candidate predictors are used to build a tree, final step is to identify the prediction from each tree in the forest and calculate the average of all predictions. Both numerical and categorical data can be handled by a random forest, but the method performs best with continuous response variables. It is broadly used with logistic regression, support vector machines and linear regression.

A research study in south China was conducted based on statistically approach to predict dengue outbreak in real time [47]. The main idea behind this research was to developed an ensemble forecast model for predicting dengue outbreak in real time. The prediction model has used dengue data, climate data, search query data and social media data as an input data in the prediction model. This model undergoes four phases namely Raw data, Process data, Ensemble data and Aggregated data. The outcome of this survey paper is that EPRA prediction model shows best performance for 1-2 week ahead. The model has achieved 94% accuracy. This prediction model can be further implemented for central region to other regions of China.

Iqbal et al. [48] concentrated on building a model that can effectively recognise the extent of the dengue outbreak in a large number of specimens at once. Seven well-known machine learning systems were evaluated for their capacity to predict dengue outbreaks. Eight different performance parameters are used to assess these approaches' effectiveness. The best classification accuracy was reported to be 92% using the LogitBoost ensemble model, with specificity and sensitivity of 94% and 90%, respectively.

The study that shows how environmental elements including temperature, humidity, precipitation, and others affect dengue transmission. Transmission of the dengue virus is more likely in regions with higher vapour pressure and precipitation rates. It made use of classification techniques to identify the key traits that caused dengue to spread. Ensemble learning technique was employed in hybrid integrations to identify the characteristics that spread the illness caused because of dengue [49].

The reliability and accuracy of a number of automated Aedes larvae detection techniques that have been previously proposed were insufficient [50]. Therefore, an automated approach was proposed which are very effective and accurate in distinguishing the Aedes larvae by providing an accuracy of about 99%. When focus on accuracy, this methodology is far better when compared to any of the earlier techniques.

The dataset for dengue prediction highly relies on labelled data and hence the supervised Machine algorithms are chosen for the study. Before deciding which algorithm or model is ideal for a certain task, it is often suggested to test a few and evaluate their performance. Prior to using any machine learning or deep learning method, it is crucial to correctly pre-process and normalise the data. The Table 1 lists few pros and cons for machine learning techniques depending on the specific dataset and problem being solved depending on which the model selection can be done. The Table 2 describes the common features of dengue defined as binary data type.

Survey articles were taken on the basis of the above analysation in predicting dengue for diagnostic models for infectious diseases. Out of the 41 survey articles, it was discovered that 20% of research publications used SVM, 7% used ANN, 7% used KNN and 5% used LR techniques. The remaining 15% DT, 20% NB, "boosting" ensemble approach 12% and 5% CNN approaches. Other machine learning methods include deep learning, J48, adaptive network-based fuzzy inference systems (ANFIS), and random forest. (VQ). Figure 3 shows the usage of various ML techniques that was surveyed.

3.png

Figure 3. Frequency of ML algorithms used in the study

The machine learning algorithms discussed in the study are evaluated to identify the performance of the model and also to test the generalization capabilities using Cross validation and K-Fold validation. The Cross-validation method is a statistical approach for comparing and evaluating algorithms by splitting the data into two sets where, one set of data is used for training and the other set is used for testing. The K-Fold validation on the other hand, splits the data into k number of folds which is used to evaluate the model with new data set. Precision and F1 score are the performance metrics used to evaluate the algorithm’s effectiveness. The Table 3 shows the evaluation of ML techniques for Cross validation, K-Fold Validation, Precision and F1 score. Table 2 parameters are input to the algorithm. The values are given in binary. The output of each algorithm is shown in Table 3.

The Figure 4 shows the comparative analysis of dengue disease prediction models of various ML algorithms like logistic regression, K-nearest neighbour, Support Vector machine, Decision tree, Naïve Bayes and Artificial Neural Network.

Table 1. Comparisons of ML algorithms

4.png

Figure 4. Comparative analysis of various ML algorithms for the dengue prediction model

The study is carried out with the same binary text input dataset for all algorithms mentioned in Table 3. Among which Naive Bayes can outperform other algorithms especially when dealing with text data. As a result, Naive bayes algorithm provides output with 97.3% through Cross Validation and 97.2% through K-Fold Validation. The Precision value is 99.1% and F1 score is 99.2% from the above analysis. SVM also provides overall performance with cross validation of 98.5%, K-Fold validation of 97.5%, Precision of 98.2% and F1 Score of 98.0%.