Study design and study population
In the present study, we applied a longitudinal research design based on a retrospective cohort study of patients who had undergone breast cancer surgery between January 1, 1996 and December 31, 2010 in Taiwan, China. The inclusion criteria were patients older than 18 years, who had received breast cancer surgery, and who were identified by database searches using ICD-9-CM 174.× diagnosis codes 174.0–174.9 and procedure codes 85.20–23, 85.33–36, 85.4×, 85.5×, 85.6×, 85.7×, 85.8×, and 85.95.
Data collection
The present study analyzed data obtained from “the Bureau of National Health Insurance (BNHI)” in Taiwan, China. The BNHI database provided detailed administrative data regarding healthcare services, including outpatient visits, hospitalizations, and prescriptions, and has become extremely comprehensive [12]. The Longitudinal Health Insurance Database for year 2005 was established using a random sample with one million beneficiaries of all residents aged ≥18 years enrolled in the “National Health Insurance” program. The data source in this retrospective study was “the National Health Insurance Research Database” in Taiwan, China.
Ethical considerations
The present study, which solely analyzed aggregate secondary data without identifying specific patients, was exempt from full review by the internal review board of Kaohsiung Medical University Hospital. Nevertheless, the study protocol still conformed to the ethical standards established with the 1964 Declaration of Helsinki, which waived the requirement for written or verbal patient consent in data linkage studies.
Potential confounders
The research variables were categorized into patient and hospital characteristics based on the research covariates. The patient characteristics comprised age, comorbidity (circulatory system comorbidity and genitourinary system comorbidity), and treatment methods (chemotherapy, radiotherapy, and hormone therapy). Breast cancer surgery type was categorized as breast-conserving surgery, modified radical mastectomy, and mastectomy with reconstruction. Comorbidities identified according to ICD-9-CM codes for primary and secondary diagnoses were used to calculate Charlson comorbidity index (CCI) [13]. The hospital characteristics were surgery volumes of hospital and surgeon, and hospital level. For each hospital or surgeon, the surgery volume was defined as the number of breast cancer surgeries performed by the respective hospital or surgeon each year. Hospital level was recorded as medical center (>500 beds), regional hospital (301–500 beds), or district hospital (<300 beds) according to accreditation by the Taiwan Joint Commission on Hospital Accreditation.
The database was randomly separated into three datasets for training, testing, and external validation in a 7:2:1 ratio. In a probabilistic view of neural networks, this randomization can be viewed as a form of statistical sampling, such as Monte Carlo sampling. Once the optimization algorithm reached a certain level of precision, the stability and generalizability of the results obtained with a given ANN should be investigated using a jack-knife validation [14].
Statistical analysis
The unit of analysis in the present study was the individual patient who underwent breast cancer surgery. The primary analytical methods were descriptive and inferential statistical analyses. The descriptive analyses had two objectives: (1) to describe the distribution of continuous variables using mean ± standard deviation (SD) and median in interquartile range (IQR); and (2) to describe the distribution of categorical variables using the number of total samples (N) and percentage (%). The inferential analysis comprised univariate and multivariate analyses using the ANN, MLR, and Cox models. The independent variables were age, CCI, surgery type, circulatory system comorbidity, genitourinary system comorbidity, chemotherapy, radiotherapy, hormone therapy, hospital level, surgery volumes of hospital and surgeon, and the dependent variable was the 5-year mortality of breast cancer patients after surgery. The discriminatory power of the models was also analyzed using the area under the receiver operating characteristic curves (AUC). Here, discriminatory power refers to the ability of a model to distinguish individuals who died from those who survived. A perfectly discriminatory model would assign a higher probability of death to patients who died than to patients who survived. For categorical variables, an overall test was applied to calculate the global P value, ensuring that the assumption of proportional hazards was not violated, and to identify any time-varying covariates.
The sensitivity analysis was performed to assess the importance of variables in the prediction models. The training process was simplified by introducing key variables and excluding all unnecessary variables. The sensitivity analysis was also performed to assess the relative significance of input variables in the prediction model and to rank the variables according to the order of importance. The sensitivity of the input variables against the output variables was expressed as the ratio of the network error (sum of squared residuals). A variable sensitivity ratio (VSR) of 1 or lower indicates that the variable diminishes network performance and should be removed.
The SPSS Version 20.0 statistical software (IBM SPSS Inc., Chicago, IL, USA) was used for statistical analyses.