Study design
This study used a retrospective cohort study design. Five tertiary hospitals in Henan Province were selected, including 3 hospitals that integrate Chinese and Western medicine and 2 hospitals that use Western medicine. The data are real-world clinical EMR data based mainly on the structured EMR system of the hospital, and they were extracted from the medical record homepage in the medical record system and from the hospitalization information. With the approval of the ethics committee, the patient’s written informed consent was not required for this study.
Patients
This study mainly included inpatients with primary liver cancer from 5 tertiary hospitals from 2015 to 2017. A total of 2,067 patients were enrolled. The inclusion criteria were as follows: ① patients diagnosed with primary liver cancer consistent with the "Guidelines for the Standardized Pathological Diagnosis of Primary Liver Cancer (2011 Edition)" [14]; ② patients aged ≥18 years old; ③ patients with complete relevant examination results and hospitalization information; and ④ patients and their families who were willing to cooperate with follow-up visits and follow-up calls, the data from which were included in the information collected. The exclusion criteria were as follows: ① patients with other primary tumors who were receiving treatment; ② patients and their family members were unwilling to follow up or patients whose original data were severely insufficient; and ③ patients who died unexpectedly.
Data source
The data of this study are based mainly on real-world clinical EMRs. Through data mining, the information in the clinical EMRs was extracted, and the data were cleaned and standardized. Finally, a relatively standardized database was established. The patient information collected in the study included the medical records of patients admitted for primary liver cancer at 5 tertiary hospitals from 2015 to 2017, as well as the telephone follow-up information for patients from August 2018 to March 2019. The EMR information mainly included basic demographic information, disease-related information (family history, drinking history, history of illness), indicators related to disease progression (Child-Pugh classification, Barcelona Clinic Liver Cancer (BCLC) staging, Chinese staging, complications, etc.), admission/discharge information, and disease treatment-related information (surgery treatment and Chinese medicine intervention). The patient follow-up information obtained by telephone follow-up mainly included the patient's final outcome, the time of diagnosis, the time of death and information related to the patient's out-of-hospital medication use.
Exposure
In this study, adjunctive TCM treatment was regarded as the exposure factor, and the cohort was divided according to the degree of adjunctive TCM treatment received by primary liver cancer patients. Patients who received adjunctive TCM treatment and had a cumulative treatment time of more than 1 month were classified as the adjunctive TCM intervention cohort, i.e., the exposure group, and patients who have not receive adjunctive TCM treatment or whose cumulative adjunctive TCM treatment time was less than 1 month were classified as the non-Chinese medicine adjuvant intervention cohort, i.e., the non-exposure group [15]. TCM treatment included TCM decoction treatment, Chinese patent medicine treatment, and TCM characteristic therapies (acupuncture, massage, external treatment, etc.).
Outcome variables
The main observation indicators in this study were the survival outcome and survival time of primary liver cancer patients. The survival outcome and survival time of patients during the study period were recorded. The survival time starting point was the time at which primary liver cancer was diagnosed, and the end point was the time when the patient died of liver cancer or the end of the study period.
Statistical analysis
According to the inclusion and exclusion criteria and the TCM exposure criteria, this study analyzed the data of patients exposed and not exposed to TCM. First, we described the baseline conditions of all patients statistically. Quantitative data were described by the mean ± standard deviation or median (upper and lower quartiles), and comparisons between groups were performed using a t test or Wilcoxon rank sum test; qualitative data were analyzed using frequencies or percentages. Comparisons between groups were performed using the Chi-square test or Fisher's exact test.
Screening for confounding factors was mainly based on the results of the comparison between the baseline groups, removing variables with P<0.05, and combining the relevant literature and the recommendations of clinicians to choose individual variables. To balance confounding factors, the GBM propensity score weighting method was adopted. The GBM algorithm is achieves equilibrium between the confounding variables of the treatment group and the control group. It does not perform significance tests between the means and standard deviations of the confounding factors of the two groups but uses the best commonly used tools to measure balance or matching: average standardized absolute mean difference (ASAM) and the Kolmogorov-Smirnov test statistic (Kolmogorov-Smirnov test statistic). In this study, the difference in propensity scores before and after weighting is represented by the KS statistic, and a KS statistic <0.05 is regarded as equality between groups [16-18].
The impact of the adjunctive TCM intervention on the survival outcome of primary liver cancer patients was analyzed by logistic regression. The Kaplan-Meier method was used to calculate the survival rate and draw a survival curve, and the survival rate was compared between groups by the log-rank test. Multivariate Cox proportional hazard regression analysis was performed to observe the influence of the adjunctive TCM intervention on the survival outcome and survival time of primary liver cancer patients and related risk factors.
Sensitivity analysis of potential confounding recognition: We usually only build models to estimate the propensity scores for the observed variables. The models do not include unobserved confounding factors or potential biases. Therefore, we needed to identify possible potential confounding factors; sensitivity analysis tests whether a model is sensitive to potential confounding bias by sequentially removing confounding variables from the model.