New SVM model predicts near-term breast cancer risk

Risk scores computed by a new support vector machine (SVM) model that works with bilateral mammographic feature asymmetry could predict a woman’s near-term risk of developing breast cancer, according to a study published in the December issue of Academic Radiology.

The prevalence of breast cancer in women necessitates uniform, population-based mammography screening protocols for those of qualifying age. However, interpreting mammograms can be difficult for radiologists thanks to the multitude of breast abnormalities that exist, the overlap of dense tissues, and low cancer detection rate.

Personalized screening recommendations based on individualized risk assessment could bypass these hinderances. 

“The prerequisite of reaching this goal of establishing a new and more effective personalized cancer screening paradigm is to identify cancer risk factors and/or develop risk prediction models with improved discriminatory power, which aim to stratify women into different risk groups whereby different screening methods and intervals can be recommended,” wrote Maxine Tan, PhD, of the University of Oklahoma, and colleagues.

The authors designed a study to investigate the feasibility of predicting near-term risk of breast cancer development in women after a negative mammography screening exam based on a statistical learning model that combines computerized image features related to bilateral mammographic tissue asymmetry and other clinical factors.

Negative digital mammograms from 994 women were retrospectively collected to conduct this study. Following the next sequential screening exam, which was 12 to 36 months later, 283 women were positively diagnosed for cancer, 349 required additional diagnostic workups and were later proved benign, and 362 remained negative.

Once a pool of 183 features was established, a Sequential Forward Floating Selection feature selection method was used to search for effective features. The researchers used 10 selected features to develop and train a SVM classification model to compute a cancer risk or probability score for each case. The two performance assessment indices were the area under the receiver operating characteristic curve and odds ratios (ORs).

Results indicated that the area under the receiver operating characteristic was 0.725 ± 0.018 for positive and negative/benign case classification. Cases were correctly predicted with the SVM model 71.3 percent of the time. The ORs revealed an increasing risk trend with increasing model-generated risk scores. Additionally, regression analysis of ORs showed a significant increased trend in slope.

“The task of achieving a good balance in risk-benefit and cost-benefit analysis for breast cancer screening remains a difficult and unresolved issue,” remarked the study’s authors. “It is thus important to investigate and develop more effective or higher discriminative risk factors and/or risk prediction models, in particular those that enable more effective prediction of near-term risk after one or a series of negative screening examinations.”