Sunday, October 18, 2009
Grand Ballroom, Salons 1 & 2 (Renaissance Hollywood Hotel)
Mehmet Ayvaci, MS1, Oguzhan Alagoz, PhD1, Jagpreet Chhatwal, PhD2, Edward A. Sickles, MD3, Elizabeth S. Burnside, MD, MPH, MS1 and Houssam Nassif4, (1)University of Wisconsin-Madison, Madison, WI, (2)Merck Research Laboratories, North Wales, PA, (3)University of California San Francisco School of Medicine, San Francisco, CA, (4)University of Wisconsin, Madison, WI
Purpose: To develop a Bayesian Network (BN) for prospective differentiation of ductal carcinoma (DCIS) versus invasive breast cancer using demographic risk factors and breast imaging features and to demonstrate the superior performance of the BN in older women. Methods: We used 1,582 diagnostic mammograms that were associated with either invasive or DCIS breast cancer which is evidenced by pathological results. Each mammogram consisted of two parts: first, structured record of demographic factors and principal clinical findings and secondly, a text report from the Radiology Information System-RIS. The mammographic descriptors extracted from the text report through Natural Language Processing and the factors in the structured part were used in constructing a BN to predict the probability of invasive cancer versus DCIS. We used 10-fold cross-validation, a common method to ensure independent validation, to train and test the model and calculate the area under (AUC) the receiver operating characteristic (ROC) curves to measure the performance of our BN for 1) patients in all age groups, 2) patients <50 years old and 3) patients > 65 years old. We compare ROC curves for older and younger cohorts using Delong’s Method as well as report on significant descriptors in all data groups. Results: The area under the ROC curve for all patients was 0.86 for predicting invasive cancer versus DCIS. We found that our model predicted invasive versus in situ disease better in older women, AUC = 0.86, versus younger women, AUC = 0.79 (p-value<0.01). Conclusion: A Bayesian Network model developed based on demographic risk factors and BI-RADS descriptors reported by radiologists in both structured and free-text mammography reports can differentiate invasive from in-situ malignancy and perform superiorly in older women. A computerized decision aid, such as ours, that can accurately predict the type of breast cancer can empower older women to better manage their breast health in the context of their co-morbidities and life expectancy.
Candidate for the Lee B. Lusted Student Prize Competition
