مکان پذیر بودن ساخت مدل عواقب پیش بینی شده برای سرطان سینه با استفاده از ابزار داده کاوی
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|22104||2008||11 صفحه PDF||سفارش دهید||7688 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Expert Systems with Applications, Volume 34, Issue 1, January 2008, Pages 108–118
A Predictive Outcome Model (POM) for breast cancer was built, and its ability to accurately predict the (5 year) outcome of an incidence of cancer was assessed. A wide range of different feature selection and classification methods were applied in order to find the best performing algorithms on a given dataset. A special Model Selection Tool, MST, was developed to facilitate the search for the most efficient classifier model. The MST includes programs for choosing different classification algorithms, selecting subsets of features, dealing with imbalance in the data and evaluating the predictive performance by various measures. These steps are important in most data mining tasks and it would be time consuming to conduct them manually. The dataset, Rose, was assembled retroactively for this study and contains data records from 257 women diagnosed with primary breast cancer in Iceland during the years 1996–1998. An extra feature, containing the risk assessment of a doctor was added to the dataset which initially contained 400 features, both to see how much that could enhance the performance of the model and to investigate to what extent such a subjective assessment can be predicted from the remaining features. The main result is that similar performance is achieved regardless of which algorithm is used. Furthermore, the inclusion of the doctor’s assessment does not appear to significantly enhance the performance. That is also reflected in the fact that the models are in general more successful in predicting the doctors risk assessment than the actual outcome if resulting Kappa values are compared.
The process of data mining is to find patterns and relationships in the data. A relatively large amount of data on cancer patients has been collected over the last few years, but the results of data mining are directly affected by the quantity and quality of the data. The Nobel price winner Joshua Lederberg stated: “Data are the building blocks of knowledge and the seeds of discovery. They challenge us to develop new concepts, theories, and models to make sense of the patterns we see in them” ( Lederberg, 1999). Successful application of data mining to cancer patient data can result in new knowledge which can assist in cancer diagnosis and in the choice of treatment. More than one million people were diagnosed worldwide with breast cancer in the year 2000, according to the International Agency for Research on Cancer’s (IARC) extensive databases (Ferlay, Bray, Pisani, & Parkin, 2004). The number of new cases has been increasing for the last few decades, especially in the western part of the world. In Iceland, the number of newly diagnosed patients has been steadily increasing since 1958, whilst the number of patients dying of breast cancer has remained nearly the same (Jonason & Tryggvadottir, 2004). Survival at five years after the initial diagnosis has changed from being less than 50% during the years 1959–1963 to about 85% during the years 1994–1998, making the prognosis of patients with breast cancer one of the best among all cancers. The purpose of this study was to build a Predictive Outcome Model (POM) that could accurately classify newly diagnosed patients into either of the following two classes: no-event and recurrence-event of cancer, five years after diagnosis. The research is based on a data set Rose, which was assembled in cooperation with the Cancer Centre of Research and Development at the University Hospital in Iceland, during the years 1996–1998. The Rose database includes a relative small number of instances (257) but a large number of features (400). The features that could be used for this study had to be facts collected from the time when the patient was first diagnosed and treated. In clinical practice, patients are classified into risk groups when diagnosed with breast cancer. It was therefore of interest to be able to conduct an experiment where the result of the classifier model would be compared to the results obtained from the clinical practice. This resulted in the introduction of a new three-valued feature named Risk to the Rose dataset. The Risk feature was the estimate, evaluated by a doctor, of the risk for a newly diagnosed patient to show marks of the disease within five years of diagnosis. The medical doctor grouped the patients into three risk groups: high, intermediate or low risk of recurrence. It was expected that the performance of the POM could be improved by adding this feature to the data set. A secondary objective was to use the specialist’s risk estimate for each patient as a class attribute to see whether the POM could simulate the pattern implicitly used by the medical doctor for estimating this risk. For the prediction, a POM was built from a training set of instances, whereas each instance was characterized by some set of given features. Building a valid and reliable POM can both be difficult and time consuming. Firstly, the modeller needs to clean the data, and select the most appropriate features and class attribute. Secondly, the data instances which the learning process will be based upon have to be chosen, and a learning method has to be selected from a range of algorithms currently available. Finally, the modeller needs to assess the reliability of the obtained results. Herein, a Model Selection Tool (MST) was constructed in order to ease the process of building an effective POM. As already mentioned, the resulting POM was to be used to predict the five years outcome for a newly diagnosed breast cancer patient using appropriate information about the patient. This tool was implemented on top of the data mining package WEKA (Witten & Frank, 2000). An important aspect of the study was to gain better understanding of the relative importance of the features included and to prepare the data for the data mining task. Specific questions that were addressed are how many and what type of features have to be selected to reach a satisfactory prediction, whether there is a learning algorithm that is significantly better than others for this type of data, and whether a subjective evaluation from a doctor has a marked influence on the results.
نتیجه گیری انگلیسی
In the case study, only 257 instances were available. The results of this study give rise to continue this work by collecting data from patients diagnosed in the year 1999 and later, when five years have passed from the diagnosis of the respective patient. It would also be valuable to collect selected information from the years before 1996 and to create a new class attribute, using 8 or 10 years instead of only 5 years. Using survival status at 60 months could also be used as a class attribute for bigger datasets instead of the outcome of the disease. There is great need to continue the work conducted in this research study. The Model Selection Tool that has been developed should prove useful in the construction of a more reliable Predictive Outcome Model. Such a model would have to be built from a larger standardized database than the one used in this study. If new biological data features are also included, there are good prospects for a practical Predictive Outcome Model.