The Center for Screening and Preventive Medicine in the Tel Aviv Sourasky Medical Center provides medical screening services to adult men and women. The program is voluntary and provides general health screening services. Most participants receive the service as an employer-provided benefit. During the visit they undergo multiple screening exams including a complete physical exam, gynecological exam and different tests. Breast cancer screening includes a CBE by a surgeon (either general or breast surgeon on a rotating basis, with a wide range of experience). If the CBE is normal, based on the woman's age, recent imaging history and insurance plan, a screening mammography is performed as well, usually during the same visit. The program has no upper age limit for screening mammography. Women with abnormal CBE are recommended for further work-up with imaging based on their age: women under 30 are recommended to undergo a breast ultrasound; women aged 30 and up are recommended to first undergo a bilateral mammogram with further work-up as needed. Further testing with ultrasound and magnetic resonance imaging (MRI) is selectively recommended per treating surgeon, based on the CBE and risk assessment. In order to expedite the screening process, the tests are performed in parallel with some women being first examined by the surgeon and some undergoing imaging first. Institutional Review Board was obtained and informed consent was waived.
All women undergoing CBE with or without screening mammography in our center between December 2007 and October 2016 were included. Women with a personal history of cancer (except for skin cancer) or known BRCA pathogenic mutation were excluded, as were medical tourists and men. We excluded women with suspicious breast complaints (detailed in Additional file 1: Appendix 1). Women with non-specific complaints (Additional file 1: Appendix 1) were not excluded.
We collected data including demographics (date of birth, health and screening history), and breast cancer risk factors (family history, parity, menopausal status, etc.) Data on mammograms done before the visit was routinely obtained from the women at the time of the visit. In cases in which this data was not available the chart was reviewed and data on previous visits to the screening center with breast imaging was retrieved.
Findings on CBE, and on mammography and breast ultrasound when performed were extracted. The CBE comprised a free-text written by a physician and/or four binary values that represent the exam findings in the breast and axilla (i.e., normal or abnormal exam in each breast and axilla). We considered the CBE result positive if one of the binary values was positive. When no values were reported, a breast surgeon manually reviewed the physician's text and decided if there was a pathological finding. A positive CBE was considered when the description included one of the words detailed in the appendix (Additional file 1: Appendix 1).
Before 2015, mammography was performed using Hologic Selenia digital mammography system (Bedford, USA). Standard four-view mammographic examinations are obtained. Additional views are performed as required, breast ultrasound is recommended in all cases with dense breast. Tomosynthesis (Hologic Selenia Dimensions) was introduced in 2015 to our breast imaging center. Single reading of mammograms by one of several dedicated breast radiologists was completed without the routine use of computer aided detection. Mammograms completed within 60 days of the CBE visit were considered to be associated with the visit. The result of the mammography test was provided in free text written by the physician and transformed into binary labels (normal/abnormal) by natural language processing. A script was created in order to extract the result and recommendation from the mammography reading using a rule-based algorithm. A mammogram was considered abnormal if the Breast Imaging-Reporting and Data System (BIRADS) classification was 0, 3, 4, or 5, or if the text included a recommendation for further work up. When the ultrasound was completed together with the screening mammogram, and the imaging was reported as normal, this was considered a normal exam. A list of terms and phrases was compiled and used to create a pattern detection script. It included an action verb or noun followed by a recommended test (detailed in Additional file 1: Appendix 1). Variations of the terms were used in order to include all possible synonyms and inflections. When the pattern was identified by the script, the study was considered abnormal, otherwise it was considered to be normal. The accuracy of the script was manually reviewed and more action verbs and recommendations were added. False positives were identified and addressed. Finally, we randomly sampled 100 cases and manually reviewed the cases to confirm the efficacy of the pattern recognition script. The analysis was performed using Python programming language (version 3.5) and packages NumPy, SciPy and scikit-learn.
The outcome was defined as diagnosis of breast cancer within 3 months of the most recent CBE visit, and was ascertained by linking the women's records to the Israel National Cancer Registry (INCR). The INCR collects data on all incident cancers since 1982; completeness of data collection on solid tumors is estimated at 95–96% . In order to assure complete catchment of all cancers diagnosed within 3 months of the visit, the data included visits of women up to October 1st 2016.
Descriptive statistics of the study population included: age, breast cancer risk factors, and number of visits per woman. The findings on CBE and mammography were summarized according to age group (under 45, 45 years and up). The groups were compared using chi-square test for categorical data, all tests were 2-sided and significance was set at 0.05.
Cancer detection rates within 3 months from the most recent CBE visit were calculated. These rates were calculated separately for the whole cohort; for different age groups; for women undergoing regular screening mammography (defined as a CBE visit less than one year from last mammogram) and for women under the screening age (under 45). These cancer cases were further scrutinized to assess the modality by which the cancer was diagnosed (CBE or imaging).
Positive predictive value (PPV) and negative predictive value (NPV) were calculated using the percentage of women diagnosed with breast cancer within 3 months of the visit. Rate precision was determined with 95% confidence intervals (CI), which were derived by using the Wilson score interval for binomial parameters.