MEDICAL DECISION SUPPORT SYSTEM USING DATA MINING TECHNIQUES

TJPRC Publication

2013, TJPRC

visibility

…

description

16 pages

link

1 file

Abstract

The healthcare industry collects a huge amount of data which is not properly mined and not put to the optimum use. Discovery of these hidden patterns and relationships often goes unexploited. Advanced data mining modeling techniques can help overcome this situation. The health-care knowledge management especially in heart disease can be improved through the integration of data mining and decision support. This paper presents a prototype heart disease decision support system that uses two data mining classification modeling techniques, namely, Naïve Bayes and Decision Trees. It extracts hidden knowledge from a database containing information about patients with two important heart diseases in Egypt, namely, AMI (Coronary artery), and HTN (High blood pressure) disease. The models are trained and validated against a test dataset. Lift Chart and Classification Matrix methods are used to evaluate the effectiveness of the models. The results showed that the two models are able to extract patterns in response to the predictable state. Five mining goals are defined based on exploration of the two heart diseases dataset and the objectives of this research. The goals are evaluated against the trained models. The two models could answer complex queries, each with its own strength with respect to ease of model interpretation, access to detailed information and accuracy.

Figures (14)

Figure 1: Medical Data Mining Decision Support System M odel for Analyzing Medical Data Support System (MDMDSS) for analyzing medical data. In the preparation stage, we use the data integration techniques such as data cleaning to remove noise and preparation stage and the production stage. Figure 1 shows the structure of the proposed medical data mining Decision

Table 1: Description of AMI Disease Attributes

Table 2: Description of HTN Disease Attributes Moheb R. Girgis, Tarek M Mahmoud & Entesar H. Eliwa

To determine if there was sufficient information to learn patterns in response to the predictable attribute, columns To determine if there was sufficient information to learn patterns in response to the predictable attribute, columns in the trained model were mapped to columns in the test dataset. The model, predictable column to chart against, and the state of the column to predict patients with one of heart disease were also selected. Figures 2(a) and (b) show the Lift Chart output results for random-guess and ideal models.

The probability that this AMI disease is a target, as determined by the Naive Bayes algorithm, is at least 2.85% yercent as shown in the Mining Legend. This number is the probability threshold required to achieve the performance a he indicated location of the lift chart. The score associated with a mining model expresses the performance of th ‘espective model as a fraction of the performance of the ideal model. Therefore, the score is always less than or equal to ‘ and the performance is better as it gets closer to 1. As we can see in Figure 2(a) the performance of Naive Bayes is bette han Decision Tree. and the performance is better as it gets closer to 1. As we can see in Figure 2(a) the performance of Naive Bayes is better Lift Chart with No Prediction Value. It tells how well each model fared at predicting the correct number of the

Moheb R. Girgis, Tarek M Mahmoud & Entesar H. Eliwa number of correct predictions (58.74%) followed by Naive Bayes (52.32%) as can be seen in Figure 4(b).

line for Decision Trees appears above the lift line of Naive Bayes. Classification Matrix: The second method to measure the effectiveness of the model is classification matrix, Classification Matrix: The second method to measure the effectiveness of the model is classification matrix, which sorts all cases from the model into categories, by determining whether the predicted value matched the actual value. All the cases in each category are then counted, and the totals are displayed in the matrix. The classification matrix is a standard tool for evaluation of statistical models and is sometimes referred to as a confusion matrix. Tree appears to be better as they give high percentage of correct predictions than Naive Bayes. This is so because the Lift standard tool for evaluation of statistical models and is sometimes referred to as a confusion matrix.

Table 3: Model Results for AMI Disease Table 4: Model Results for HTN Disease

DATA MINING GOALS We present here five mining goals that were defined based on exploration of the two heart diseases dataset and the

Figure 8: Screen Shot for AMI & HTN Heart Disease Decision Support System Identify the Effect of the Inputs and their Relations with Each one of the Two Heart Diseases we are Studying in this Research This goal can be satisfied by allowing the user to enter the values of medical attributes for a diagnosing patien vith heart disease and predict whether the patient is likely to be diagnosed with one of the two heart diseases. The Decisio: ‘rees and Naive Bayes models could answer this question using “what if’ scenarios. For example, in the case of the AM eart disease, entering the values A ge = 64, FBS = 83, LDL = 188, into the models, would produce the output shown i igure 8. The two models showed that the person has AMI heart disease. Decision Tree gives the highest probabilit 99.96%) with 41 supporting cases, followed by Naive Bayes (42.25%). As the values are high, doctors can recommen 1at the patient undergoes further heart examination. Performing “what if’ scenarios can thus help prevent a potential hear For example, the Dependency Viewer of the Naive Bayes and Decision Trees models shows a quick display of

Moheb R. Girgis, Tarek M Mahmoud & Entesar H. Eliwa This is identified only in Decision Trees viewer as shown in Figure 10(a) and (b). Figure 10(a) shows that the highest probability (100.0%) that patients with A MI disease are found in the relationship between the risk factors (nodes): “Age >=93.5 and < 60.2 and FBS >= 81.8 and < 125.4 and NHDLC = 210.” This is identified only in Decision Trees viewer as shown in Figure 10(a) and (b). Figure 10(a) shows that the

Figure 10: Decision Tree Viewer for the Two Heart Diseases Medical Decision Support System Using Data Mining Techniques Figure 10(b) shows the Decision Tree viewer of HTN disease which indicated that the highest probability Information on least impact (49.98%) is found in the relationship between the risk factors: “MBP < 106.8 or >=

(a) Naive Bayes Attribute Characteristics Viewer in (b) Naive Bayes Attribute C haracteristics Viewer in Descending Order for Patients with HTN Disease Descending Order for Patients without HTN Disease Figure 11: The Attribute C haracteristics of the AMI Patients (a) Naive Bayes Attribute Characteristics Viewer in (b) Naive Bayes Attribute Characteristics Viewer in Descending Order for Patients with AMI Descending Order for Patient’s without AMI Figure 12: The Attribute C haracteristics of the HTN Disease

(a) Attribute Discrimination Viewer in Descending (b) Attribute Discrimination Viewer in Descending Order for Naive Bayes of AMI Disease Order for Naive Bayes of HTN Disease and 159.598” with 72 (20.45%) patients etc. favor predictable state for patients with no HTN disease. This query can be answered by analyzing the results of attribute discrimination viewer of Naive Bayes model. The