A hybrid deep learning model for predicting and targeting the less immunized area to improve childrens vaccination rate

3.1 Data set

The authors [15] reported 337 publications using data from the National Family Health Survey (NFHS); on the other hand, only 48 publications used data from the District Level Health Survey (DLHS) and 3 publications used the Annual Health Survey (AHS). The research papers were published in the AHS data sets between January 2011 and March 2015. Unfortunately, the NFHS and DLHS round four surveys were not used. DLHS-4 round four survey data set is used in this research work; more precisely, data is taken from the Salem District Health Level Survey in the state of Tamil Nadu. This work focuses on native district, and in the future it will be extended to neighboring districts and states.

In 2013, DLHS-4 was collected by the International Institute for Population Sciences, Mumbai, on behalf of the Ministry of Health and Family Welfare, Government of India. All the data sets were updated in 2016 as per the information provided on the official website of the Ministry of Health and Family Welfare. The updated version of the DLHS-4 data can be accessed by sending formal request to the Director of Indian Institute of Population Sciences.

Since none of the parents have been identified or contacted directly, no ethical approval is required for researchers to undertake this study. The DLHS-4 survey provides detailed information at the district level and at the village level to group the less immunized areas. It helps to identify and improve the level of immunization in particular region in the state level rather than at the national level. National level surveys have only been performed annually and provide data on national-level. This research work concentrates more on state level information. Hence, DLHS-4 survey data set is used in the proposed work.

There are approximately 5,057 children aged between 12 to 23 months in the DLHS-4 data set. Out of these, 2,050 (40.5%) were able to show their vaccination card (recorded by Indian health workers), 1,751 (34.5%) claimed to have a card but were unable to produce it, and the remaining 1,256 (25%) did not have a card. The survey did not include mother’s self-reports on the timing of their children’s vaccinations, so only 2,050 children’s vaccination cards could help to determine age-appropriateness and also find out were measles, DPT-3 and other vaccines were used.

There are 28 predictors in the DLHS data set, all of which are numerical values and with three binary variables as shown in Table 1. A maximum of 5000 specimens are included in the data set. A 100-fold cross-validation approach was used to eliminate bias. Training and testing data sets were chosen randomly along with cross-validation at a ratio of 2:3 or 1:3. This data set was collected without any senseless, noisy information, but it had missing values. These missing values were mostly obtained due to misinterpretation of the questions asked during the survey process, Some of the questionnaires were not answered accidentally. Therefore, these Non-Respondent data were called missing values in the data set. There is no question about the authenticity of the data, because if someone wished to have the data, they were to provide the proper identification proof that the government would recognize it.

The collection of documents obtained from the location-based social net-works in the Salem area was divided into training and testing on the basis of 2:3 and 1:3 ratios. This collection contains 2000 sample files. It contains information on outbreaks of disease and the doctor’s comments on the disease and vaccines. To maintain the validity, this data set was cross-verified with Anganwadi workers. Table 2, shows the information about the data set such as target variables and predictor variables as given below.

From the given DLHS data set, the attributes are identified for the model creation and pre-processed those data columns which as not answered’ record that is similar to missing values. The “Ministry of Human Resource Development” survey department allocated median value for the documents with “NA” or “Not Answered” in the data set. With the help of statistical domain knowledge, It is found that, during the analysis process, this median value affects the prediction process and produces over-adapted results.

So, these not answered’ or Missing value are imputed with the predicted value from multivariate imputation via chained equations package in R tool’. Instead of creating single imputation value this R package creates multiple imputations to handle uncertainty in missing value. In this package, predictive mean matching is applied to numeric variables to predict the missing values from the given data set.

The data set was imported into python Spyder IDE, from CSV data format into the data frame. The data frame is split into X and Y list variables. Again the split X and Y list into X-train, X-test, Y-train, and Y-test sets. Data set is divided into two halves such as X and Y. X part contains socio-demographic data such as: Highest education qualification of women in the house, occupation of husband and wife, whether the woman is living with her husband, husband ever attended school, highest standard of men, women’s employment, sex of the child, received antenatal care, treatment for house-hold persons during the time of illness, enrolled in health schemes, frequency of antenatal check up, places of antenatal care, place of first check-up for a child, vaccination card and; Y part contains child immunization related data sets such as: vaccination received but not recorded, ever had a vaccination, vaccinations like BCG, polio, polio vaccination received or not, DPT, number of DPT doses received, measles, hepatitis b, Vitamin A dose, pulse polio dose, the main reason for not taking the vaccination.

Sci-kit learns machine learning library is imported into python to per- form pre-processing. The data set is having attributes with varying scales, so re-scaling is done on those attributes to make it useful for many deep learning algorithms. This method is useful for the gradient descent method to get optimized results. It also successfully supports regression and neural network algorithms. Data set has been re-scaled using a sci-kit package with a min-max scalar class.

Similarly, the data set that is obtained from the set of relevant documents of Location-based social networks available within Salem region is gathered and pre-processed before giving it as an input into the DSS-RSM model. The text-based data set is converted into tokens by using word hashing method such as Tokenizer. It is a default package from Keras and TensorFlow. The tokens generated are further converted into vector matrix with each text in the document being converted into vectors. This vector matrix is ready to get embedded inside the DSS-RSM model.

4.1 Rank similarity learning based multi layer perceptron neural network

Multi-Layer Perceptron Neural Networks has brought many breakthrough results in speech recognition, computer vision, and text processing. The MLP recommendation system is used together to recommend some information to the user, based on the given input data. For example, MLP is applied in the YouTube recommendation system to recommend videos to the users based on the user information and kind of the videos they watched over the internet [16]. MLP is also applied in prediction kind of activities; it helps in making drought forecasting over time-series data [17]. In these areas, MLP can generate predictive results for many kinds of processes and the results are satisfactory [18].

In this work, a novel deep learning technique R-MLP proposed and implemented as shown in the Fig. 1. This includes ranking similarity learning at the end of the model to rank the output. The proposed work uses 2 hidden layers in the R-MLP neural network with a maximum of more than 20 neurons in each layer to represent the attributes or features from the data set. A sampled output vector for each vaccine is obtained from the proposed model at the output layer. A ranking similarity function uses Pareto dominated sets at the end of the model to classify the cross matrix values obtained at the output layer.

Based on the sampled output vector and target values of the vaccine, the Pareto dominated sets are created for each vaccination coverage in the panchayat and town regions [19]. Based on this method, Pareto sets are compared with a binary quality indicator to find the sum of 0 s and 1 s. The sum of different output is compared with the final results. This serves as a basis to rank those output. All these results will be finally grouped as a list.

Each of the neurons in the input layer contains the data for each variable and presents them as an input vector.

χ p = ( χ 1 p , χ 2 p , … , χ N p ) T(1)

OPEN IN VIEWER

This input vector is propagated to the intermediate layer using the following propagation rule:

s k p = ∑ i = 1 N w j ⁢ i ⁢ χ j p + 𝑟𝑎𝑛𝑑𝑜𝑚 ⁢ ( b ) i(2)

Where N is the number of input neurons of the network w j ⁢ i is the value of the weight of the connection between neuron j of the input layer and neuron i of the middle layer, and b i is the value of random bias associated with neuron i .

Considering the activation value of neuron i is a function of the input vector.

Therefore, its output vector is given as:

y k p = F i ⁢ ( S i p )(3)

Similarly, for a neuron k of the output layer, the equations that determine its activation state is:

s k p = ∑ i = 1 L w i ⁢ k ⁢ y i p + 𝑟𝑎𝑛𝑑𝑜𝑚 ⁢ ( b ) k(4) y k p = F k ⁢ ( S k p )(5)

Here ‘ L ’ is the number of neurons in the intermediate layer, ‘ w i ⁢ k ’ is the weight value of the connection between neuron ‘ i ’ of the intermediate layer and neuron ‘ k ’ of the output layer, and b k is the value of the random bias associated with the neuron ‘ k ’. Here, the error term for the output neuron is calculated by the Eq. (6)

F k ⁢ ( S k p ) = 1 1 + e - s k p(6)

The activation equation used in this ANN is the sigmoidal or logistic function.

From the obtained output, the two sampled output vector has been created and it will be taken for ranking similarity learning. Before the output is getting ranked, the Pareto dominated set is generated with the comparison between the output from two different sampled output vector. Taken two sampled solutions S ⁢ ( y i ) and S ⁢ ( y j ) ∈ Ω f , y i will dominate y j or vice-versa if and only if, for all i ∈ { 1 , 2 , . , m } and for all j ∈ { 1 , 2 , . . , m } .

This Pareto dominated set is associated with the vaccination indicator such as 1 for partially immunized and 0 for non-immunized again and compared with the dominated set for ranking the sampled output vector.

If the vaccinated indicator is taken as X k ∈ { 1 , 0 } , from this indicator it is compared with the sampled solutions as follows to find the binary similarity indicator:

(7)

Then these results will be summed up based on the binary similarity indicator I i o ( + 1 , - 1 ) .

If the number of positive value similarity is greater than negative value similarity it is ranked as high. If the number of positive value similarity is lesser than negative value similarity it is ranked as low. If the number of positive value similarity is approximately equal to negative value similarity it is ranked as medium. Finally, the ranked results from the two different regions are grouped to form a ranked vaccinated coverage list.

DSSM is a deep neural network (DNN) modeling technique for representing text strings in a continuous semantic space and modeling semantic similarity between two text strings. DSSM has wide applications including information retrieval and web search ranking [20, 21], advertisement selection/relevance, contextual entity search and interesting tasks, question answering [22], knowledge inference [23], machine translation [24], etc.

On the other hand, one recent semantic matching method called Deep Structured Semantic Model (DSSM) was proposed by [25] in a form of supervised learning, which utilizes the relevance labels in the model training. With the labels incorporated into its objective function, the DSSM maps query, and a relevant document into a common semantic space. In semantic space, the similarity measure is calculated as positive similarity score, even though there is no shared term between query and documents.

OPEN IN VIEWER

Furthermore, it has been reported in [26, 27] that applying the DSSM to natural language processing yields superior single feature performance. The DSSM similarity score contains the label information so that not only high relevance without term match receives high similarity score, but also the low relevance with term match also has a low similarity score as a result of penalization. In Fig. 2, the new proposed work shows that the DSS-RSM is used to find the disease outbreak regions from the set of documents on location-based social networks available within Salem region and a query is given as an input to find most occurrence or most spoken topic in the diseases outbreak regions. The query will represent the name of the disease that will be given as input and obtains the percentage of semantic similarity from a word document. The hashed input object from the word document and query are taken as an input into the ranking similarity learning algorithm to rank the document which matches mostly with the query. In this task, a group of documents is used for ranking the similarity process.

Finally, the hashed input object from the word document and query are taken as an input into soft cosine similarity measure function. Soft cosine similarity between document and query vectors considers similarities between a pair of features, which is completely different from the traditional cosine similarity, Since it measures the similarity of features in the vector space model. This generalizes the concept of cosine as well as soft similarity. In the other end, when Ranking SVM is applied to document retrieval, a feature vector is created from query-document pair. The ranking support vector machine is used to rank the documents. It applies the pairwise classification problem and it uses support vector classifier to solve the problem. Each feature is defined as a function of queries and documents. The input is taken as a word sequence into the word hashing method by formatting the text from the query and answer word document into the following sequence:

x t = w 1 , w 2 , … , w T Q ⁢ y t = w 1 , w 2 , … , w T D(8)

The word hashing takes the input query and word sequence and convert into vectors as letter-trigram vectors for each word [28].

l t = [ f t - d T , … , f t T , … , f t + d T ] T , t = 1 , … , T(9)

Here, f t is the letter-trigram representation of the i th word.

The input given to the convolutional layer is the letter trigram as vector, the feature vector is calculated with activation function as follows:

h t = tanh ⁢ ( W c ⋅ l t ) , t = 1 , ⋯ , T(10)

The output of the convolutional layer is given as an input into the max-pooling layer to retain only the most useful feature vectors produced by the convolutional layers.

v ⁢ ( i ) = max t = 1 , … , T ⁡ { h t ⁢ ( i ) } , i = 1 , … , K(11)

The sentence-level feature is given as an input into the semantic layer to generate high-level semantic representation by y .

y = tanh ⁢ ( W s ⋅ v )(12)

Here, ‘ v ’ is the global feature vector after max-pooling ‘ W s ’ is the semantic projection matrix, and ‘ y ’ is the vector representation of the input query in the latent semantic space. Finally, the semantic similarity score is computed between query ‘ Q ’ and document ‘ D ’ by using cosine similarity.

(13)

Learning to rank is a typical method for ranking the documents based on the query. So in this method a ranking SVM algorithm is included to rank the documents based searching query. It uses the pairwise classification technique to classify the document that mostly matches the search query. This method will create classes based on the output from the pairwise classifier as Class 1 and Class 2. Assuming there is an input vector space for a set of documents as X d o R n , and query as X q .

There will be output space vector for ranking different categories such as Y = c 1 , c 2 . . , c m , here c presents the classes as follows:

$f(X_d)>f(X_q)\text{or}(w_i,f(X_d)>f(X_q))$(14) $iM(w_i)={\displaystyle \frac{1}{2}}||w_i||\widehat{2}+C{\displaystyle \sum _i^m}(w_i,f(X_d)>f(X_q))$(15) (16)

This equation compares the relationship between two instance values from document and query with weighted feature vector and creates a new vector with a new label for ranking as c1, c2. etc. Equation (8) gives minimum regularization parameter which is used for cross-validation and counts all positive and negative score. Based on this score the output is classified to the respective classes. If more documents match with a query it comes under Class 1 or it goes to Class 2. So, Class 1 is denoted as high ranked and Class 2 is denoted as low ranked.

OPEN IN VIEWER

The proposed hybrid model R-MLP and DSS-RSM is one novel model which differs from other models [29]. The output result that has been obtained from the R-MLP is a predicted result and it has to be cross verified with low dimensional semantic results from the DSS-RSM model. The above mentioned ensemble model has been shown in Fig. 3. Further, the output from R-MLP and DSS-RSM models are combined using an ensemble method to produce one optimal predictive meta-model and at it produces predicted output. Here, the bootstrap aggregation ensemble method is used to group predicted results to cross-validate and identify partially immunized area. Based on this, counter-measures have to been taken to reduce the percentage of partially immunized and increase the percentage of fully immunized.

5.1 Evaluation measures

The content of the confusion matrix is used to calculate the precision, recall, accuracy and F1-measure. Here, F1-measure [31] is used to evaluate our models prediction performance. F1-measure value is calculated by averaging the F1-measures for each category.

F ⁢ 1 = 2 ⁢ ( Precision * Recall / Precision + Recall )(17)

OPEN IN VIEWER

where, if TP denotes the size of the true positive set, FP denotes the size of the false-positive set, and FN denotes the size of the false-negative set, precision and recall are defined by

Precision = TP / TP + FP and(18) Recall = TP / TP + FN .(19)

As the geometric mean of precision and recall, the F1-measure penalizes heavily both false positive and false negatives. Accuracy is the proportion of the total number of predictions that were correct and is calculated by the following equation:

Accuracy = TP + TN / TP + FP + TN + FN(20)

OPEN IN VIEWER

After executing the R-MLP and other discussed algorithms with a given data set, the rate of children’s partially immunization rate is predicted in spite of each vaccination that has been offered to the children in the panchayat and town regions of Salem.

The regions in and around the district of Salem are grouped into panchayat and town regions. At first, the comparison has been done between socio-demographic attributes against each immunization attributes separately to find out the predicted result for each vaccination in one specific grouped region as a sample from panchayat and town regions.

The predicted result of the R-MLP with test set has been compared with other discussed algorithms with the same test set, as a bar graph, which shows whether there is reasonable improvement in the prediction while applying R-MLP. For the above-obtained results, the precision, recall, accuracy, and F1-measure are shown in Table 3. Results from the above algorithm are classified based on output vector, partially (or) not-immunized and vaccination coverage area into predicted results at the panchayat and town regions. These classified results are taken as an input into the ranking similarity-based learning algorithm with Pareto dominated sets for ranking the predicted results. This ranking has been done separately in the panchayat and town regions. The following Table 4 shows the ranking of each vaccination coverage in the panchayat and town regions with the percentage of partially immunized vaccinated rate.

OPEN IN VIEWER

OPEN IN VIEWER

After compiling and fitting the model with given the data set, evidence for occurrence of the disease is predicted around Salem region. Specifically in the villages of Salem by evaluating cosine similarity between queries and documents in low dimensional space. At first, the prediction has been made on both input queries and the set of documents on location-based social net-works available within the Salem region. With each predicted input queries and resulting word from a predicted document, the query/document semantic matching has been done to find the percentage of occurrence of particular predicted results in low dimensional space by mapping two predicted feature vectors in a latent semantic space through a deep neural network.

Using the ranking SVM, the pairwise classification processes are used to classify and rank the classes. The documents which match repeatedly with the query are labeled as highly ranked documents. A similar kind of process is carried out for all the documents. List of documents are denoted as classified and ranked in the form of Class 1 and Class 2.

For every possible query, the F1-measure has been calculated and recorded in the following Table 5 using soft cosine similarity measure function.

Cross verifying the proposed Composite model performance by using ensemble method. Finally, find the cross-relationship between R-MLP and DSS- RSM results by using the ensemble method. Table 5 shows how the DSS-RSM results influence the predicted results from R-MLP in the panchayat and town regions in case of partially immunized and non immunized vaccinated coverage area.

Cross-validation usually finds comparison between two different predicted aggregated results. The results of the two models are aggregated separately using the bootstrap aggregation method. The aggregate equation given blow finds the aggregates on each output.

f ⁢ ( x ) = 1 M ⁢ ∑ m = 1 M ( f m ⁢ ( x ) )(21)

The value of the one-hot distribution, either true or false is represented as 0.5 and 0.0. The value 0.5 represents the serious predicted immunization problem in the respective region. The value 0.0 represents the region with fewer immunity problems. This distribution is compared with the combined results to find the probability of loss function as follows [32]:

H ⁢ ( p , f ) = - ∑ x p ⁢ ( x ) ⁢ log ⁡ f ⁢ ( x )(22)

In the above equation, the probability of loss function is given. were, ‘ p ⁢ ( x ) ’ is the needed probability and ‘ f ⁢ ( x ) ’ is the actual probability. In the following Table 6, the loss of functions is calculated in two different regions.

The cross-validation result from the above table shows that there are some minimum and maximum deviations in the loss function. The area which shows minimum loss function will have greater possibilities to get affected by the disease and it may go too serious by affecting nearby villages. So those kinds of villages must be identified from the above results and proper care should be given by conducting needed and useful medical awareness camps in the villages that are vulnerable to the disease outbreak in future.