Akosa, Josephine Kelly, Shannon SAS Analytics Day

Transcription

1 Application of Data Mining Techniques in Improving Breast Cancer Diagnosis Akosa, Josephine Kelly, Shannon 2016 SAS Analytics Day

2 Facts and Figures about Breast Cancer

3 Methods of Diagnosing Breast Cancer Surgical Biopsy: Has sensitivity close to 100%. Cost associated with this method is high. Mammography: Sensitivity fluctuates between 68% and 79%. Reporting sensitivity varies with radiologists experience. Limitations including variation in age and breast density prevent researchers from significantly improving the sensitivity of this method.

4 Methods of Diagnosing Breast Cancer (contd.) Fine Needle Aspiration (FNA) with visual interpretations: Sensitivity varies between 65% and 95%. FNA biopsies are minimally invasive and can be completed within minutes. Limitations associated with mammography are less severe with FNA with visual interpretation. Computer decision aids can improve radiologists ability to correctly diagnose the malignancy of breast tumors.

5 Data Description and Preparation The study utilizes the Wisconsin Breast Cancer data, originally compiled by Dr. William H. Wolberg and available within the UCI Machine Learning Repository. The dataset contains 699 clinical case samples (65.52% benign and 34.48% malignant) assessing the nuclear features of fine needle aspirates taken from patients breasts. There are 11 attributes per observation including the ID and the binary target variable. The input variables are measured on an interval scale (1-10), with 1 indicating a normal state and a value of 10 indicating a highly abnormal state. There were 16 missing values but due to the small percentage (2.3%), these cases were excluded from the analysis. Weight of evidence approach (WOE) was employed to convert the categorical variables into numerical values.

6 Data Description and Preparation To ensure honest assessment of the models built, the data was partitioned into training (70%) and validation (30%) subsets. Prior probabilities were set to account for oversampling. Variable Label Mean Standard Deviation Minimum Median Maximum Skewness Kurtosis WOE_BC Uniformity of Cell Size Uniformity of Cell WOE_BN Shape WOE_CT Bland Chromatin WOE_MAdh Bare Nuclei WOE_Mit Clump Thickness Single Epithelial Cell WOE_NN Size WOE_SECS Marginal Adhesion WOE_UCSh Normal Nucleoli WOE_UCSz Mitoses Table 1. Weight of evidence variable summary statistics

7 Methods A variety of data mining techniques were considered for model building. All models were built in SAS Enterprise Miner Support Vector Machines Gradient Boosting Logistic Regression Data Mining Techniques Random Forest Decision Tree Neural Networks

8 Process Flow Chart

9 Model Comparison Fit Statistics Misclassification rate KS Statistic Gini Coefficient ROC Index Sensitivity Specificity SVM (Linear) % 95.52% Decision tree (3 branches) % 94.78% Autoneural (default) % 97.01% Random Forest via PLS % 96.27% Random Forest via regression % 96.27% Linear Logistic regression % 96.27% Autoneural via regression % 97.01% Random Forest via PC % 96.27% Decision tree via PC % 96.27% Boosting via PC % 96.27%

10 Explaining the Best Model With regards to the selected gradient boosting model, the first 5 principal components (PC) were used in the model building. These components account for 90.48% of the total variability in the data. PC_1 PC_2 PC_3 PC_4 PC_5 PC_6 PC_7 PC_8 WOE_UCSz WOE_UCSh WOE_SECS WOE_BC WOE_NN WOE_BN WOE_MAdh WOE_CT WOE_Mit

11 Summary of Findings The gradient boosting model turned out to be the best model for diagnosing breast cancer using data from fine needle aspiration. Uniformity of cell shape and size, bare nuclei, and bland chromatin were identified as the best FNA characteristics with respect to breast cancer diagnosis. Outcome prediction can be further improved by refining the methods used to identify and measure the FNA characteristics. Finally, utilizing this model would help decrease interpretation errors by radiologists.

12 Acknowledgement & Contact We wish to express our sincere gratitude to Dr. Goutam Chakraborty, Department of Marketing and founder of SAS and OSU Data Mining Certificate program Oklahoma State University for his support and guidance throughout this study. CONTACT INFORMATION Josephine Sarpong Akosa 320-C Math Sciences (MSCS) Stillwater, OK