Lessons Learned from Data Mining of WHO Mortality Database

Objectives: The objectives of this research were to test the ability of classification algorithms to predict the cause of death in the mortality data with unknown causes, to find association between common causes of death, to identify groups of countries based on their common causes of death, and to extract knowledge gained from data mining of the World Health Organization mortality database. Methods: The WEKA software version 3.5.3 was used for classification, clustering and association analysis of the World Health Organization mortality database which contained 1,109,537 records. Three major steps were performed: Step 1 - preprocessing of data to convert all records into suitable formats for each type of analysis algorithm; Step 2 - analyzing data using the C4.5 decision tree and Naive Bayes classification algorithm, K-means clustering algorithm and Apriori association analysis algorithm; Step 3 - interpretation of results and hypothesis testing after clustering analysis. Results: Using a C4.5 decision tree classifier to predict cause of death, we obtained 440 leaf nodes that correctly classify death instances with an accuracy of 40.06%. Naive Bayes classification algorithm calculated probability of death from each disease that correctly classify death instances with an accuracy of 28.13%. K means clustering divided the data into four clusters with 189, 59, 65, 144 country-years in each cluster. A Chi-square was used to test discriminate disease differences found in each cluster which had different diseases as predominant causes of death. Apriori association analysis produced association rules of linkage among cancer of the lung, hypertension and cerebrovascular diseases. These were found in the top five leading causes of death with 99-100% confidence level. Conclusion: Classification tools produced the poorest results in predicting cause of death. Given the inadequacy of variables in the WHO database, creation of a classification model to predict specific cause of death was impossible. Clustering and association tools yielded interesting results that could be used to identify new areas of interest in mortality data analysis. This can be used in data mining analysis to help solve some quality problems in mortality data.