Neural network used in predicting cardiovascular risks

JOURNAL OF SCIENCE OF HNUE FIT., 2011, Vol. 56, pp. 40-47 NEURAL NETWORK USED IN PREDICTING CARDIOVASCULAR RISKS Nguyen Thi Thu Thuy Informatics Department, The Commercial University of Viet nam, HoTungMau Rd, CauGiay, Hanoi Email: nguyentthuthuy@gmail.com 1. Introduction No gold standard exists for assessing the risk of individual patients. Current techniques use a generic technique applied to the patient’s cardiovascular record. This data itself is inconsistent, over a history of patients at any one clinical site, and not always immediately useable. The research is applying data mining methods to make the clinical data more useable, meaningful and open to the use of neural and other classifier techniques. Risk assessment systems were designed and implemented to help the clinicians in their decision for the patients particular cardiovascular uses. These systems sup- port the diagnosis based on medical data and knowledge domain. The quality of medical decision making will be improved by the support from these systems and clinical experiences. This research focuses on the popular system, which is using broadly in Britain medical decision support system, The Physiological Operative Severity Score for enUmeration of Mortality and morbidity (POSSUM). The research focuses on the using of both supervised learning, and unsuper- vised techniques in the medical domain, in particular to the cardiovascular domain. These techniques are Multi-Layer Perceptron (MLP), Radial Basic Function (RBF), and Support Vector Machine (SVM) for supervised learning, and Self Organizing Maps (SOM) for unsupervised learning. The techniques of supervised ones are ap- plied to the data domain in order to have a comparison between the evaluated system of POSSUM and the advantage of Neural network. The comparisons are based on the rate of mortality and morbidity of patients. The outcome set of unsupervised learning techniques is compared to the results of supervised ones. 2. Data The given data is collected from Hull site and Dundee site in the hospitals of Britain. They are allocated in many excel files. The original data included the infor- mation about cardiovascular patients. There are errors in row data such as duplicate 40 Neural network used in predicting cardiovascular risks information, missing values, or are inconsistent. Original databases from Hull site has 98 attributes and 497 patients. This data is scored from the risk assessment system of POSSUM, PPOSSUM. Data from the Dundee site is taken from patients’ information at Dundee hospital. It includes 35 attributes and 341 patients. A part of the data is used from previous researcher [1]. This data was trans- formed to be appropriate values for using of neural networking of a number of soft- ware such as SNNS (Stuttgard Neural Network Simulator), and NeuroDimension. The research had a comparison of results between former research and the present research. This is mentioned in [2]. This research focuses on the row data from Hull and Dundee site. Other files might be additional information for the research in the choosing of the input, or output set for training processes. The data mining methods of preparing data for the process of neural networks are data cleaning with the duplicated ID information, missing value. The duplicated ID is treated as alternative person. The missing value is replaced by the default value of “Null”. For the patient who has numerous missing values in its content this patient could potentially be ignored. Row data is transformed to an appropriate value for each neural network technique. The normalization of data is scaling data into the range of [0,1]. In the method of linear normalization the new values can be calculated by the formula: New value = (original value-minimum value)/(maximum value- minimum value). For example, the original value of IPSI% in Hull site data domain is 80, the minimization, and maximization value of this attribute is 30, and 99 respectively. So, the new value for IPSI% attribute will be (80-30)/(99-30) =0.725. Alternatively other methods such as using mean and standard deviation, or decimal scaling of each attribute are used in order to scale the values in specific ranges of [0,1]. The data transformation is represented in attribute construction. The at- tributes with Boolean values can be transformed to the value of “0” or “1”. Moreover, the attributes with “symbol” values might be transformed to the following method: Dividing the original attribute to sub-attributes, which are equal the number of its values. For example “Indication” attribute in Hull site has 4 values as A-F, ASx, CVA, TIA, so we divide by Indication_A-F, Indicattion_ASx, Indication_CVA, and Indication_TIA . Therefore, the sub-atrributes continue to be transformed to smaller groups as possible. The last transformation usually is Boolean values of “0” or “1”. An example about the transformation of data can be seen in Table 1 below. Table 1. Example of one new sub-group value IDC (Indication) IDC_A-F IDC_ASx IDC_CVA IDC_TIA A-F 1 0 0 0 ASx 0 1 0 0 CVA 0 0 1 0 TIA 0 0 0 1 41 Nguyen Thi Thu Thuy 3. Risk assessment systems The medical decision support system are populary reviewed in such as MYCIN; INTERNIST-QMR; the GLAsgow system for the diagnosis of DYSpepsia; Acute Physiology and Chronic Health Evaluation [3]; or POSSUM and PPOSSUM [4]. The research focuses on the popular risk assessment systems in Britain (POSSUM and PPOSSUM). The Physiological and Operative Severity Score for enUmeration of Mortality and morbidity (POSSUM) is a system, which provide a tool for predicting risk adjustment and comparative audit. POSSUM (and PPOSSUM) rely on an a-priori scoring of physiological and operative severity parameters. The calculation of the physiological and operative severity parameters is based on multivariate discriminant analysis of factors measured in a broad group of general surgical patients. The logistic regression analysis in this model tries to produce statistically significant equations for both mortality and morbidity based on a 12 factors/4 grades physiological score and 6 factors operative severity score [4]. There is another model which based on POSSUM, it is Portsmouth POSSUM (P-POSSUM). This equation was derived from a heterogeneous general surgical pop- ulation and has been used as an audit tool to provide risk-adjusted operative mor- tality rates. The equations correlated to mortality and morbidity were introduced in [4]. Experiment: The data in Hull site were already scored in POSSUM and PPOSSUM system. However, the evaluation of mortality and morbidity rate need to be calculated in order to have a comparison between scoring system prediction and neural network prediction. The data in the Dundee site did not include scores from risk assessment system of POSSUM and PPOSSUM. This data needs to be highlighted for further investigation. There is an example of experiment of calculation mortality and morbidity rates. The data used in this experiment is from the Hull site, and already scored for the physiological and operative severity attributes. The equations of POSSUM to predict the morbidity and mortality for each patient are used. The performance of the system is measured by the ratio of death between the prediction of the system and the observed death. Table 2 below shows the mean predicted risk of mortality in 7 groups of patients, and the comparisons between predicted death and observed death of the POSSUM system. The performances of the mortality rate for PPOS- SUM can be seen in Table 3 below. 42 Neural network used in predicting cardiovascular risks Table 2. Comparison of observed and predicted death from POSSUM logistic equations Range of predicted death rate Mean predicted risk of Mortality (%) No of operations Predicted deaths Reported deaths The ratio (O/E) 0-10 % 7 130 9 9 0.99 10-20 % 15 81 12 19 1.57 20-30 % 25 31 8 2 0.26 30-40 % 36 9 3 0 0 40-50 % 43 15 6 5 0.78 >50 % 62 5 3 3 0.97 0-100 % 15 265 41 38 0.93 Table 3. Comparison of observed and predicted death from PPOSSUM logistic equations Range of predicted death rate Mean predicted risk of Mortality (%) No of operations Predicted deaths Reported deaths The ratio (O/E) 0-10 % 3 222 8 30 3.75 10-20 % 14 24 3 2 0.67 20-30 % 23 12 3 2 0.67 30-40 % 33 4 1 3 3.00 40-50 % 44 2 1 2 2.00 >50 % 57 1 1 0 0.00 0-100 6 265 17 38 2.24 The ratio between observed and expected number of adverse outcomes indi- cates the prediction performance. A ratio of 1 indicates that there is an average performance; greater than 1 means the performance is worse than expected; and less than 1 means the performance is better than expected predictions. For example from table 3, the ratio (O/E) for the range of predicted death rate of 20-30% is 0.26. This means the performance of operation is better than the predicting operation. However, the ratio for the range of 10-20% is 1.57. This means the performance of operations is worse than predicting operation. Overall POSSUM gives close to accurate risk estimation, with a O/E ratio of 0.93. However its performance varies across the different risk categories, and is particularly poor for low risk operations (10-20% bands). Overall PPOSSUM 43 Nguyen Thi Thu Thuy underestimates the risk (O/E = 2.24), and for no one group does it give an accurate risk estimation. The need for better estimators is therefore obvious. 4. Neural network techniques The Neural Network (NN) approach adopted is that of an information pro- cessing system that consists of a graph representing the processing system as well as various algorithms which access that graph [5]. Neural Network techniques can be divided into two methodologies: supervised learning and unsupervised learning. For supervised learning, the data is trained via networks with expected (a-priori defined) outputs. The supervised techniques used are MultiLayer Perceptron (MLP), Radial Basic Function (RBF), and Support Vector Machine (SVM). Conversely, with the unsupervised method, no a-priori classifications are used. The comparison of previous research results in neural network and current research is mentioned in [2]. The results can be seen in Table 4 below. The mis- classification is based on Mean Square Error (MSE) by using the same data domain, and structure of neural network techniques such as MLP, and RBF. Table 4. Comparison of current research (*) and previous research (**). MLP-0H MLP-1H MLP-2H RBF-0H RBF-1H RBF-2H SVM SOM MinMSE (CV) 0.072 0.059 0.06 0.06 0.06 0.06 0.122 0.06 Epochs 903 96 107 53 899 194 84 52 MSE (*) 0.061 0.048 0.046 0.052 0.052 0.052 0.057 0.052 MSE (**) 0.12 0.15 0.13 0.25 0.07 0.09 — — The current results (except SVM) seem to be appropriate more than from the initial clinical study, despite being hampered by reduced training volume. It is suggested that a more rigorous data preparation stage is responsible. One way will be to investigate the confusion matrix for each classifier to see the type of errors being made. There is an example of experiments for using neural network with the evaluation of mis-classification by confusion matrix. Using the neural network techniques are supervised learning. The data in the experiment is taken from the Hull site. It includes 265 patterns with 86 attributes, which are used for risk assessment in section 3 above. However the data needs to be prepared in order to be appropriate for use with the different networks. The methods to transform data are mentioned in section 2 above. By eliminating irrelevant attributes, the transformed data set has 83 attributes with 265 patterns. In the first experiments using neural network techniques, they are compared with POSSUM as a means for predicting mortality rates. The cleaned 44 Neural network used in predicting cardiovascular risks data has a mortality rate of 14.34% (38 from 265 patterns with status= “dead”). The accuracy results are obtained through the generation and analysis of the confusion matrix. The results are compared to the predictions given in Tables 2 and 3. Overall, the predicted mortality rate for each neural network technique was lower than the observed one (see detail in Table 5 below). The percentage of mis-classification in each model is obtained by dividing the sum the mis-classification of “dead” or “alive” patients by the total number of patterns. The results show that although POSSUM gives a better result for the ratio of observed and expected deaths, its mis-classification is the highest. Table 5. The comparison of results of experiments with supervised neural network techniques, POSSUM and PPOSSUM Models No Predicted Misclassification The ratiooperations deaths Dead Alive % (O/E) POSSUM 265 41 32 29 23 0.9 PPOSSUM 265 17 11 32 16 2.25 MLP 265 15 23 12 13 2.53 RBF 265 0 38 0 14 N/A SVM 265 11 28 13 15 3.45 To ensure the provision of the highest quality of care a comparative audit of the data for different outcomes can be investigated. Patient parameters such as strokes, myocardial relapse within 30 days of operations (30Day_MR), and cardiovascular arrest within 30 days (30Day_CVA) may be used as indicators for outcome risk for individual patients. Subsequently a new summary output attribute (risk) is built based on the value for the two main post-operative outputs. This attribute takes three values (High, Medium, Low) based on the heuristic rules:∑ (Status, 30Day_MR) = 0 → Risk =Low∑ (Status, 30Day_MR) = 1 → Risk =Medium∑ (Status, 30Day_MR) = 2 → Risk=High WEKA software is used to develop the different neural classifiers to be applied. The data set is split in two ways. A test set is taken by using 50% of the overall pattern set or using a 10 fold cross validation partition. With the latter technique, the data set is divided into 10 partitions. One partition is used as a test set whilst the rest is for training; the procedure is repeated 10 times, so that each partition acts as a separate test set. The results can be seen in Table 6. It is clear from Table 6, the results with cross validation test data set are better than other type of test sets. The reason may be that because of having many times for training. From table 6, the MLP model has provides the best predictions of patient risk with a Mean Square Error (MSE), and a mis-classification (0.02, 3.7% 45 Nguyen Thi Thu Thuy with type 1, 0.01, 1.9% with type 2 respectively). Table 6. The comparisons of alternative neural network techniques Models Test set Misclassification MSEL M H % MLP 50% split 0 5 0 3.7 0.02Cross validation 0 2 3 1.9 0.01 RBF 50% split 0 7 3 7.5 0.05Cross validation 0 4 6 3.8 0.03 SVM 50% split 0 2 0 1.5 0.08Cross validation 0 2 3 1.9 0.07 5. Discussions and futher works POSSUM and PPOSSUM are generic clinical tools that allow a metric factor to be used in assessing the severity of illness. The risk assessments are compared to reported mortality across a group of patients. The ratio between the predictions of POSSUM, PPOSSUM and the observed mortality shows the performances of the system. However, each individual patient has an assessment of risk, which is based on clinical judgement. POSSUM and PPOSSUM seem to over predict mortality for the data. Furthermore, the outcomes of these models are just mortality and morbidity whilst for cardio-vascular disease the combining of other outcomes such as 30 day MR for strokes or dead may be appropriate outcome. The results of using neural network techniques seem to be appropriate for data domain. This is represented by the comparison in Table 5 and 6. The risk of “Low”; “Medium”; and “High” for cardio-vascular patients is a more appropriate assessment than indication of status of “Dead” or “Alive”. By using the confusion matrix, the mis-classification of each model is eval- uated. From Table 5 and Table 6, it seems that using different models of neural network produces smaller mis-classification errors than with POSSUM, and PPOS- SUM. More interestingly, the models using new outcome of risk (High, Medium, Low) had smallest percentages of mis-classification compared to the other risk pred- ication models (i.e. mortality or morbidity). The results of neural network, particular to the models with the new outcome of risk need to be subjected to further investigation. Moreover, a comparison of supervised versus unsupervised classifiers may help in determining more appropriate patient classifications. These results can then be applied in determining what of the original data should be used to generate a better set of classifiers and indicators of use in predicting cardio-vascular risk. The selection of input attributes for patient classification is an issue for this and further work. The reduction of attribute set which is relevant in the data domain 46 Neural network used in predicting cardiovascular risks might improve the classifiers. This is based on the theory of mutual information If the domain derived techniques are not to be trusted or are to be independently vali- dated, then alternative means of clustering patients (according to risk) are required. Unsupervised neural techniques are used to achieve this. REFERENCES [1] Kuhan, DN Davis*, IC Chetter, CN McCollum+, PT McCollum, 2003. The use of Artificial Neural Networks for risk prediction following Carotid endarterectomy. Internal Report. [2] TNguyen TT and Davis ND, 2005. Predicting CardioVascular Risk Using Neural Net Techniques. Poster in The Computer Science Conference. [3] Lisboa P.J.G, 2002. A review of evidence of health benefit from artificial neural networks in medical intervention. Elsevier: Neural Network 15, pp. 11-39. [4] Copeland G P, Jones D, Walters M, 1991. POSSUM: a scoring system for surgical audit. Br J Surg 78, pp. 355-360. [5] Dunham, M. H, 2002. Data mining introductory and Advance Topics, Upper Saddle River, NJ : Prentice Hall/Pearson Education. ABSTRACT Neural network using in predicting cardiovascular risks Neural Networks are broadly applied in a number of fields such as cognitive science, diagnosis, and forecasting. Medical decision support is one area of increasing research interest. This research looks at the application of data mining techniques to the assessment of risk in the clinical domain of cardiovascular medicine. The primary technique is the modelling of medical data using a number of alternative neural net- work techniques. The research focuses on classification for risk assessment in given medical data using both supervised and unsupervised learning. Specific techniques investigated are Multi-Layer Perceptron, Radial Basic Function, and Support Vector Machine for supervised classification, and Self Organizing Map, Cluster algorithm for unsupervised ones. The popular risk scoring systems used in British Surgery are described. The Physiological and Operative Severity Score for enumeration of Mortality and morbidity (POSSUM), and Portsmouth POSSUM (PPOSSUM) are systems, which provide a tool for predicting risk adjustment and comparative au- dit. The POSSUM and PPOSSUM models are built assuming a linear relationship between the outcome and other variables. The risk assessment of popular medicine scoring system such as POSSUM and PPOSSUM is compared to neural network predictions in order to increase the advantages of using NN. All results of predicting risks are analysed in order to evaluate the agreements between NN techniques, and the decision will be further investigated. 47