Classification weight of new born baby at Sayang Ibu maternity Hospital with probabilistic neural network (PNN)

The purpose of this research is to classify baby weight before baby was born by considering maternal factors such as maternal weight, maternal height, maternal hemoglobin (Hb), maternal blood pressure, maternal age, total parity and fetal factors such as twin status using PNN (Probabilistic Neural Network). PNN is one of the models in Artificial Neural Networks (ANN) that exists for classification. PNN can be used in classification because it can classify with optimal results, higher accuracy and faster than other neural network models. PNN is structured by four layers, namely the input layer, pattern layer, summation layer and output layer. From classification with PNN on weight of new born baby using 7 factors, we obtained that the best proportion of accuracy classification on weight of new baby born is 90:10 with σ value is 0,2. The accuracy of classification within training data is 100% whereas on testing data is 92%. Based on classification on testing data we obtain eleven data are classified into the same as original class and one data os predicted to the other class.


Introduction
Probabilistic Neural Network (PNN) is one of the models of Artificial Neural Network (ANN) which is for classification. PNN is often used in classification because it can map each input pattern to a number of classifications optimally, faster, and more accurate than other neural network models [1]. PNN is an algorithm that use probability functions, it does not require large data in the process, and has the advantages to overcome the problems that previously existed in back propagation method such as overcome the long training time, getting stuck in a global minimum, and the difficulty of network architecture settings. PNN is supervised learning so it does not require a repetitive training process (iteration) to improve the parameters that will later be used in recognizing the class of a data, namely parameter smoothing σ [2].
Infant mortality rate is one of the most important aspects to describe public health. The main cause of neonatal death is Low Birth Weight (LBW). LBW is one of the 3 main causes of infant mortality [3]. Surasmi, et al (2003) [4], stated that factors whivh caused may premature birth were divided into maternal factors, fetal factors, placental factors and unknown factors. The maternal factors that may caused LBW are age, nutrition, total parity, distance parity, hemoglobin (Hb), height, and weight. As  According to Department of Healthy Republic Indonesia [3], the determination of LBW status is only based on the weight of the baby born without involving another factor. In this research we will identify and predict the class of birth weight based on 7 factors, namely maternal hemoglobin, maternal age, maternal blood pleasure, maternal height, maternal weight, total parity and twin status using historical data in 2017 with PNN classification analysis.

Normalization data
The normalization technique is used in this research to overcome the discrepancy between the values of variables. The normalization value is calculated by the average and variance as in the following formula:

Data training and data testing
Training data is used by classification algorithms to form a classifier model. Data testing is used to measure the extent to which the classifier successfully classifies correctly. The classification model is then built based on training data and then its performance is measured based on testing data. Distribution of training data and testing data uses the following proportions: Training data= proportion of training data× N (4) Testing data=N-training data (5) Where N is total data (training data + testing data).

Probabilistic Neural Network (PNN)
PNN is an Artificial Neural Network (ANN) method using principles from statistical theory namely Bayesian Classification to replace the heuristic principles on back propagation algorithms [2]. The process of PNN is faster compared to back propagation ANN. In addition, PNN does not need a large data set in the learning phase, and has the advantage of being able to overcome the problems that exist in back propagation ANNs, such as overcome the long training time, getting stuck in the global minimum, and the difficulty of planning network architecture. However, PNN has a problem in determining smoothing parameters which are usually determined by trial and error or user defined [5]. Let Vnp is the matrix of training data, where n is the number of training data and p is the number of variable. The algorithm for PNN is: 1. Let X(qxp) is input matrix where q is the number of input data and p is the number of input variables. Input data can be in the form of training data and testing data.
Define bias weight input layer as follows: Where below bias weight input layer and σ is smoothing parameter with 0 <  ≤ 1

4.
Each point is conducted for k = 1 to q for all input data (training data and testing data): a. Calculate the distance of matrix X to matrix V. The calculation of distance is the Euclidean distance between the input vector k x and the weight vector where is Euclid distance between input data X and matrix V. The th row input data on matrix X and the th training data on matrix V. b. Search for a1 ki activation value. The calculation of the activation value using the radbas function is performed on each distance matrix index as follows: ih W is data ꀘ data on ꀘ ꀘ class on matrix W d. Find z so that a2kz = max ( a2 kh │ꀘ=1,2,...,r) e. Assign z as class of classification.

Accuraty of classification
A system that performs classification is expected to be able to classify all data sets correctly, but it is undeniable that the performance of a system cannot be 100% correct so that a classification system must also measure its performance [4]. To calculate the percentage of accuracy used the equation:

Low Birth Weight (LBW)
According to the health profile of Samarinda City in 2015, the birth weight of a baby is the weight of the baby weighed within the first hour after birth. Birth weight is divided into two: low birth weight (LBW) and normal birth weight. According to the Indonesian Ministry of Health (1996), LBW is a baby born weighing 2500 grams or less regardless of gestational age. Surasmi, et al (2003) [6] stated that the factors causing LBW birth were divided into maternal factors, fetal factors, placental factors and unknown factors. Maternal factors such as hemoglobin, age, weight, height, blood pressure, amount of parity and fetal factors such as twin status.

Data sources and research variables
The data in this research is obtained from Sayang Ibu Maternal Hospital. Total data as much as 117 data with 102 normal weight of new baby born and 15 data on low birth weight. Factors used in this study were maternal factors such as hemoglobin (hb), age, weight, height, blood pleasure, total parity and fetal factor such as status of twins.

Data analysis
The following are the stages of data analysis in this study using the Octave GUI is: 1. Data Normalization The purpose of data normalization is making data on the same range and make some variable that have a high value are not dominating variable in the classification. Normalization data using equation number (1), (2) and (3) .

Define Smoothing Parameter
Determine the value of σ with interval 0 < σ ≤ 1. This research using value of σ as much as 0.2; 0.4; 0.6; 0.8; and 1. Then, the σ value is used to determine bias weight of input layer using Equation number (6).

Proportion Training Data and Testing Data
Proportion training data and testing data is 50:50, 60:40, 70:30, 80:20, and 90:10. Distribution of training data and testing data based on proportion is assisted by using Octave Software. Proportion of training data and testing data is calculated by equation (4) and (5) Figure 4.1 shows weight of baby new born at Sayang Ibu Maternal Hospital. From 117 baby, has been seen 87% as much as 102 babies has normal weight and 13% or as much as 15 babies has low weight or notified LBW. From these data it can be seen that more of baby new born has a normal weight compared with LBW at Sayang Ibu Maternal Hospital.

Classification of PNN
The result from classification weight of new baby born using algorithm PNN with combination of 5 proportion training data and testing data and 5 smoothing parameter as follows From Table 1. We know that higher σ make lower accuration training data. The best accuracy results is on proportion 90:10 with a value of σ as much as 0,2. The accuracy of the training data as much as 100% and the accuracy of the testing data as much as 92%.

Conclusion
The result from classification with PNN on weight of new born baby using 7 factors, namely maternal hemoglobin, maternal height, maternal weight, maternal age, maternal blood pressure, number of parity and twin status of baby we obtained that the best proportion of accuracy classification on weight of new baby born is 90:10 with σ value is 0,2. The accuracy of classification within training data is