In Africa, experimental studies have shown that most cancers are diagnosed at an advanced stage of the disease which usually contributes to its complications and mortality rate. This is due to a limited awareness of the early signs and symptoms of the disease among the public and healthcare providers. According to Lasebikan, Nwadinigwe & Onyegbule, (2014), the mortality rates of this disease is always compounded by the later stage at which the disease is diagnosed, presenting a ticking time bomb of life expectancy and lifestyle changes such as women having fewer children, as well as hormonal intervention such as postmenopausal hormonal therapy [1]. To reduce this harm caused by the disease, an effective way is to detect it early [2]. However, early detection and prognosis requires an accurate information, reliable analytic procedure and efficient algorithm. Therefore, the researcher's direction in this work is to present a reliable analytic procedure and efficient algorithm suitable for the prediction of cancer of the kidney through data analytic approach, in which medical practitioners and patients can gain valuable knowledge and help for proactive intervention strategies in order to save lives from this harmful disease.
Data analytic has proven to be a multi-dimensional discipline that uses descriptive techniques and predictive models to gain valuable knowledge from data warehouses for recommendations and decision making. It is the discovery of patterns and communication of meaningful insight in data [3]. According to Berson, Smith and Thearling (1999), data analytics is the science of examining raw data with the purpose of drawing conclusions from it [9]. It focuses on inference, identify undiscovered patterns and establish hidden relationships [4]. Figure 1 depicts the process of data analytics. The science is generally divided into exploratory data analysis (EDA), where new features in the data are discovered and confirmatory data analysis (CDA) where existing hypotheses are proven true or false. Typically, it is used to describe the technical aspects of data analysis, especially predictive modeling, machine learning techniques. Data Analytics has been commonly apply to business data, marketing mix modeling, web analysis, risk analysis and fraud analysis to communicate insights from data. It is very good in recommending action and guide decision making,
Dataset pertaining to this research work was collected from selected health centres and hospitals in the south western part of Nigeria using purposive and selective sampling techniques. The researcher collected a sample data totaling, 1,006 records from fifty-two selected health centres in six ( 6) different states. The data collected was cleaned, normalized and organized in a form suitable for data analytic process. Table 1 shows the data format for the research data collection while Figure 1 and Figure 2 show the visualized information about selected states and health centres respectively.
Table 1 shows the data format for the research data collection
Statistically, out of the 1,006 patient's data captured, 44.8% were male while the remaining 55.2% are female, (See Table 2). The analysis further revealed that 57.1% of the patients are exposed to chemical and industrial contents while 32.7% of the population as gender and hereditary disorder. The patient's life style data collected also indicated that the people around this region are addicted to smoking and drinking of alcohol, regular use of nonsteroidal anti-inflamatory drug (NSAIDs) such as ibuprofen and naproxen, which can double the risk of the disease by 51%. Other factors include obesity; faulty genes; a family history of kidney cancer; having kidney disease that needs dialysis; being infected with hepatitis C; and previous treatment for testicular cancer or cervical cancer. There is an indication also, that High blood pressure is a possible risk factor though still under investigation.
Waikato Environment for Knowledge Analysis (WEKA) platform was used for the data analytic experiment. It is a powerful data mining tool that has a GUI Chooser from which any of the four major WEKA application environments (Explorer, Experimenter, KnowledgeFlow and Simple CLI) can be selected. The Explorer Application is selected for this experiment because it has a workbench that contains a collection of visualization tools, data processing, attribute ranking and predictive modeling with graphical user interface (GUI) for easy access to this LMT from the family of Decision Tree. The Decision Tree also known as "white box" classification model can provide explanation for their models, and could be used directly for decision making [5], while the Rule Induction is one of the fundamental tools of data mining, in which formal rules are extracted from a set of observations. The rules extracted represent a full scientific model of the data [6]. According to Kapil et al., (2013), rule induction is a popular and well researched method for discovering interesting relations between variables in large database. These abilities and aptitudes of rule induction are suited and of good requirement for any effective and efficient intelligent system. A major paradigm of the Rule Induction is the Association Rules [7].
As shown in Figure 3, the patient's databank component is responsible for the data collection, updating and storing patient's data from different sources. The classifier system component is responsible for the data modeling based on the algorithms in the layers. The performance evaluation component is responsible for the evaluation of the performance of the algorithms considered in the layers using standard metric to produce the best (optimal) algorithm. The rule generated from this algorithm is to be incorporated into the prediction system. Since the objective of the research work is to present a suitable algorithm for the cancer of the kidney prediction system, which the work has achieved. Hence the prediction system processes is not discussed in the work, but will be discussed in the future work of this research.
Ten (10) classification algorithms from the family of classifiers implemented in this work were used to model the patient's dataset. The datasets for the experiment was first divided into two, which includes the training and testing datasets. 66% of the datasets was devoted to training while the remaining 34% was used for testing of randomly selected data. JRip, PART and Decision Table in layer 1 of the classifier system were first used to model the patient's data and later the Decision Tree classifiers.
The 10-fold cross validation test and percentage split modes were also considered in the modeling. Since they are from different classifiers family, they yielded different models that classify differently on some inputs. The algorithms were tested on the datasets in order to determine that which best models the data with best predictive accuracy.
The comparison of the performance of the various algorithms in layer 1 and layer 2 based on the output from the percentage split (hold-out) and 10-fold cross validation modes was carried out. The results of the models from the two modes and the performance evaluations are presented in Table 3. The 10-fold cross-validation test mode was considered good since it produced the best model both in layer 1 and 2 of the classifier system. Moreover, the 10-fold cross validation mode have been widely used, and it is described a better option to determine the performance of a classifier [8]. Table 4 shows the standard metric accuracy details from the 10-fold cross validation mode considered for all the algorithms in the experiment. Figure 4 and Figure 5 show the graphs of predictive accuracy and time taken to build the models by the classifiers respectively. From the experimental results and analysis, it shows that the J48 decision tree and LMT rules outperform all other algorithms in the layers. However, J48 decision tree was chosen as the best algorithm in this work because it has the correctly classified instances of 74.7%, ROC Area of 0.78 and recall of 0.714 respectively. It has a lower FP rate of 0.153, F-Measure of 0.614 and took lesser time of 0.03 seconds to build the model compared to LMT and other classifiers as shown in Table 4. Additionally, J48 decision tree algorithms generally have this ability that can produce a simple tree structure with high accuracy in term of classification rate, even with huge volume of data [9]. Pruning methods have been introduced to reduce the complexity of tree structure without any decrease in classification accuracy. The J48 decision tree structure and rules as generated by WEKA are presented in Figure 6. The rules generated from the best algorithm (J48 pruned decision tree) are as stated in rules 1 to 20. The rules were tested in a prediction system framework and their prediction levels are classified as follows: (PL) -One, Two and Three. This show the status of patients and by interpretation: Level One and Two indicates a risk level or status of the disease manifestation in the patients that needs to be attended to urgently. While, level Three indicates that the patient is not manifesting any symptoms of kidney cancer disease, but may suffer from other diseases. A back-end for updating the rules as the situation arises will be incorporated into the system to match other conditions.
The research work was focused at presenting an efficient algorithm suitable for predicting the status of kidney cancer in patients. To achieve the objectives of the research work: (i). Dataset pertaining to patient was acquired from fifty LGA (52) selected Health Centres in the south western region of Nigeria using purposive and selective sampling techniques. (ii) the researcher developed a two-layered classifier system consists of Rule Induction and Decision Trees implemented on Waikato Environment for Knowledge Analysis (WEKA) to build the data model using data analytic approach, and (iii) different machine learning algorithms were used in search for the algorithm that produced the best model with predictive accuracy. In the experiment, ten (10) classification model algorithms from different classifier family were implemented on the patients'dataset. Since they are from different classifiers family, they yielded different models that classify differently on some inputs. The comparison of the performance of the various algorithms in layer 1 and layer 2, and the standard metrics of accuracy, precision, recall and f-measure for the best classifier considered in this work was carried out as shown in Table 3 and Table 4 respectively. The results show that the J48 decision tree outperform all other algorithms in the layers with predictive accuracy of correctly classified instances of 74.7 % in 0.03 seconds, ROC Area of 0.78, FP rate of 0.153, TP rate of 0.714, precision and recall of 0.614.