Effectiveness of Artificial Neural Networks in forecasting failure risk for pre-medical students

This research paper evaluates the ability of Artificial Neural Networks (ANN) to predict the performance of the applicants students to Medical Sciences, in order to predict their failure/ risk in their premedical year. Educational institutions in general, consider a variety of factors when making admission decisions. Traditionally, academic researchers have developed several statistical models to predict an applicant's success in the academic programs. An ANN model is designed based on existing academic acceptance criteria for medical college. The Cascade Correlation Learning structure of ANN is used to predict students' final grades in their premedical year. The result of this research shows that the neural network model can predict students' performance even better when the similar characteristics for input data have been maintained.


Introduction
The registration department in universities usually regulates a set of rules in order to accept new students. These rules are set to select applicants who have the abilities and skills to pursue and succeed in their academic career in a particular field of studies. Acceptance to the College of Medicine and Medical Sciences at the Arabian Gulf University is no exception. The admission body of the college firstly makes sure that qualifications and other particulars of the individual students fulfil the requirements of the college admission according to its rules and conditions. Secondly, they arrange acceptance tests that result in accepting, or rejecting applicants. However, it has been noticed throughout the years that the decisions based solely on the results of these acceptance tests, including an additional personal interview, are not sufficient. A number of students still fail in the first year of education (premedical stage). This creates many problems both to the University and to the students; whether administrative, financial and/or psychological. The Artificial Neural Networks techniques can play a role in this process. However, it should use the same set of entrance rules but with different weightings. This research aims at modelling an effective and intelligent tool that can help the decision makers to have better judgment on possible ability of each individual student to succeed in the premedical stage and hence in his/her subsequent studies in the medical school.

Artificial Neural Networks approach
Artificial Neural Networks are sophisticated modeling techniques; capable of modeling extremely complex functions. Cascade Correlation Network (CCN) architecture of Artificial Neural Networks is used to deal with this research problem. It is a supervised learning algorithm developed by Fahlman and Lebiere [1]. It is an attempt to overcome certain drawbacks and limitations of popular Back Propagation Learning Algorithm developed by Rumelhart et al. [2]. It is trained using the Quick Propagation Algorithm (QPA) which is the enhanced version of Back Propagation (BP). It is a convenient approach to solve problems since it is an ontogenic neural network that generates its own topology during training. It overcomes the limitation of BP which is slow at learning from examples. A BP network may www.intechopen.com Artificial Neural Networks -Methodological Advances and Biomedical Applications 356 require many thousands of epochs to learn the desired behaviour from examples. An epoch is defined as one pass through the entire set of training examples. Previous studies involve some researches close to this research. Predicting MBA student performance by evaluating the ability of three different models, namely, logistic regression, probability analysis, and neural networks; is reported by Fahlam and Lebiere [3]. The result was that the neural network model performs better than statistical models. Back-Propagation Neural Network was used in selecting surgical residents via the National Residency Matching Program applied to medical students interested in surgery in their fourth year of study. It is used to facilitate the surgical residents' selection process using 36 variables. The study showed that neural networks are capable of producing significantly better results than traditional statistical approaches [4]. Many other studies (Entwistle, 1988;Ardila, 2001;Busato et al., 1999, Furnham et al., 1999 were undertaken in order to try to explain the academic performance or to predict the success or the failure of students; they highlighted a series of explanatory factors associated to the student [5]-[8].

Implementation of ANN in forecasting the failure risk of applicants to Medical College
Premedical students' data including their grades at the end of first academic years were collected from both registration department and Medical College Acceptance Committee. Data for the academic years: 2003/2004, 2004/2005, and 2005/2006

Neural inputs and output
The main variables affecting the process of selecting medical students were identified. The main input variables used and involved in the acceptance process are shown in Table 1. The notations column is used for abbreviation. The following input variables were added in order to test if they have any significant effect on the forecasted results. The added variables are shown in Table 2, and the Output variable is shown in Table 3.  Table 2.

Forecasting model
Neural network software tool used in this research is "Forecaster XL' (v. 2.3)". It uses the Cascade Correlation supervised learning algorithm -trained using Quick-Propagation. It divides the rows of inputs x i and their related output y i into training and validation subsets. 83% of rows are used for training, while 17% of rows are used for validating the model. Network training is designed to adjust network weights for maximizing predictive ability and minimizing forecasting error. Validation subset is that part of data used to tune the network topology or network parameters other than weights. For example, it is used to define the number of hidden units to detect the moment when predictive ability of neural network started to deteriorate. Test subset is a part of data set used to test how well the trained neural network forecasts using new data. Test subset is used after the network is trained, and hence ready to forecast. This subset is not used during training and thus consists of a new data to the model. The regular Sigmoidal activation function is used. It is expressed in relation to the input variable "x" as follows: 1 / (1 + exp(-x)) -0.5 , (Where x is the input).

Results and discussions
The Artificial Neural Networks model is built based on students' historical data. The input variables are identified together with their related output variable (Input / Output variables pair). A typical example of the prepared data Inputs for a particular student are: Gender (F), Age (18) Data reprocessing is done by analyzing the input and output variables. Data was cleaned by removing any column of data that was identified as unsuitable for neural network (e.g. that contains text or too many missing values or repeated values). Numeric columns are marked with numeric mark. Categorical columns are marked as categorical. Table 4 shows input and output variables used in this research work as recognized by the used software.  Table 5 shows an analysis of the data for training process. The data has been classified into training set and validation set. Some columns (input variables) were excluded because it proved that they have no effect on the training process. Specifically, namely; marital and Batch input variables. 'Marital' was excluded since all the students were 'single' and hence it has no contribution to the network output. 'Batch' variable was also excluded since all students enrolled are fresh graduates, and hence the variable is not an informative one.   Figures 2 and 3. Figure 2 shows the training MSE using data of the academic year 2003/2004 with 9 input variables, the lowest MSE reached at the early stages of training. Figure 3 shows the training MSE using data of academic year 2003/2004 with 5 input variables. The network restores the structure that resulted in the lowest error (this is at iteration number 1956). The resulted network structure is presented in Table 7 for each academic year. Since input data are numerical and categorical; numerical inputs would be recognized as one input in the training set, while categorical inputs would be recognized as two inputs. The table shows the number of hidden units in each network structure. The categorical inputs such as (male/ female) is represented by {-0.4, 0.4}, while the forecasted output (target) is the GPA. Hidden units vary when using a different academic year as training set in order to achieve the lowest error.  In academic year 2003/2004 with 9 inputs, the input variables Marital and Batch were excluded as discussed before. Therefore 7 inputs were left. And since Gender is categorical, it had been recognized as 2 inputs and may take 'Male' or 'Female' values. Because of that, the final number of input variables would be 8. In academic year 2004/2005 with 9 inputs, the input variable Batch was excluded, hence, the final number of input variables was 10. In Table 7, the number of hidden units varies among the different networks to achieve the lowest error for each network. Inputs importance (Sensitivity Analysis) is shown in Table 8. Table 9 shows the obtained results in each trained network.

Conclusion
The target of this research work is to study the effectiveness of Artificial Neural Networks in forecasting failure risk for pre-medical students at the Arabian Gulf University. The model structure for forecasting uses the Cascade Correlation Networks technique, and trained using the Quick Propagation algorithm. Different models with different input variables were built. The data of the academic years 2003/2004, 2004/2005, and 2005