Malignant cancer may increase the risk of all-cause in-hospital mortality in patients with acute myocardial infarction: a multicenter retrospective study of two large public databases

MIMIC is a database of critical care medicine on the United States created by a team of critical care physicians, emergency physicians, and computer specialists at the Beth Israel Deaconess Medical Center (BIDMC), Massachusetts General Hospital, Massachusetts Institute of Technology (MIT), and the University of Oxford. The version we utilized is MIMIC-IV 1.0, which contains case information on patients admitted to an ICU or emergency room at BIDMC between 2008 and 2019 [18]. The eICU Collaborative Research Database (eICU-CRD) was made available by Philips Healthcare in partnership with the MIT Laboratory for Computational Physiology [19]. As a multicenter database, eICU-CRD contains data on more than 200,000 admissions to ICUs from 208 hospitals in the United States between 2014–2015.

Data extracted from the MIMIC IV and eICU-CRD databases do not require individual informed consent because research data is publicly available and all patient data are de-identified. These databases were approved by the Massachusetts Institute of Technology (Cambridge, MA) and Beth Israel Deaconess Medical Center (Boston, MA), and consent was obtained for the original data collection. The author (L.Z.) attended a series of courses offered by the NIH and was granted access to these databases after passing the required assessment (ID: 38601114).

All patients diagnosed with AMI in the MIMIC-IV database and the eICU-CRD were included. The exclusion criteria were patients younger than 18 years, duplicates, had unclear outcomes, and lacked troponin test results. Finally, 3,034 and 5,968 patients were included from the MIMIC-IV database and the eICU-CRD, respectively (Fig. 1).

Missing data is a common phenomenon in the two databases. In order to achieve the consistency of variables as much as possible, we extracted the same variables with less than 20% missing values between the two databases. We use structured query language to extract data from the database. The data collected were age, gender, weight, Acute Physiological Score (APS), first care unit, intervention measures within 24 h of ICU admission, ventilator use, vasopressor use, dialysis use, whether they were undergoing PCI, whether CABG (coronary artery bypass grafting) was performed; main site of cancer; and comorbidities including congestive heart failure (CHF), hypertension, cerebrovascular disease (CD), chronic pulmonary disease (CPD), liver disease (LD), diabetes,renal disease (RD). Peak troponin: troponin t in MIMIC IV and troponin i in eICU-CRD; The results of the first laboratory examination after ICU admission were obtained for, white blood cell (WBC) (K/μL), hemoglobin (g/dL), platelets (K/μL), potassium (mEq/L), sodium (mEq/L), bicarbonate (mEq/L), creatinine (mg/dL), BUN (mg/dl), and glucose (mg/dL). Vital signs within 24 h of ICU admission were also obtained: heart rate, blood pressure (BP), respiratory rate, and temperature. Figure S1 shows the exact situation for continuous variables in both databases. The correlation matrix shows the Pearson's correlation between any two variables in the datasets [20].

The outcome assessed in this study was all-cause in-hospital mortality. Follow-up began with the admission of these patients and ended with their discharge or death.

Patients with AMI were divided into two groups according to whether they had concomitant malignant cancer: cancer (not including patients with a previous history of cancer) and noncancer groups. After the Shapiro–Wilk test, the continuous variables in this study are all non-normal distributions, so the median and interquartile ranges were used to described. Kruskal–Wallis tests analyzed the differences in continuous variables. Categorical variables were described using counts and percentages, and χ2-test or Fisher’s exact test was used to test the distribution differences between groups. The association between the two groups and all-cause in-hospital mortality was analyzed using Kaplan–Meier and Cox proportional-hazards regression models. The log-rank test was conducted for nonparametric analysis to compare the survival distributions of the two groups. Cox multiple regression was used to control for confounding variables.

Propensity score matching (PSM) and inverse probability of treatment weighting (IPTW) were used to further verify the stability of the results and reduce the influence of data bias and confounding variables. When running PSM and IPTW with R software, missing variables are not allowed. So, the “mice” package of R software was used to account for missing covariate values using multiple imputation. In the PSM analysis, the cancer group included both the patients diagnosed with AMI and cancer after admission. Patients in the cancer group were matched with patients without cancer through nearest-neighbor matching at a 1:1 ratio. IPTW is an approximation method based on PS construction to deal with confounding variables, which applies the inverse of the propensity scores as weights to the original population and constructed two hypothetical populations. PSM reduces the sample size so that the populations of groups with the same PS have approximately equal sample sizes, while IPTW increases the sample size, which is similar to the formulation of a unified standard population, and adjusts the average level of the observed effects in the two groups according to the weights of confounding variables in the standard population, so as to eliminate the influence of the different distribution of internal confounding variables on the effect values between the two groups. Standardized mean differences (SMDs) before and after matching were used to determine whether PSM and IPTW reduced the differences in covariates between the two groups. Finally, variables that are still unbalanced and may be confounded by clinical judgment were further incorporated into the Cox multiple regression model for adjusting and verifying the results.

Further subgroup analyses were performed to assess the potential correction effects by age (< 55 years or ≥ 55 years), gender (male or female), PCI (no or yes), CABG (no or yes), ventilator use (no or yes), vasopressor use (no or yes). The potential interactions were evaluated by adding cross-product terms of groups with the above stratified variables to the model.

In the MIMIC-IV database and the eICU-CRD, 3,034 and 5,968 patients, respectively, met the selection criteria and were included in this study. Dividing the patients with AMI into two groups according to whether or not they had malignant cancer resulted in the noncancer groups of the MIMIC-IV database and the eICU-CRD including 2,737 and 4,368 patients, respectively, and the cancer groups including 297 and 578. Table 1 lists the general condition, disease severity, comorbidities, and laboratory parameters of the patients. Patients were older in the cancer group than in the noncancer group in the MIMIC-IV database (73.00[65.00,80.00] vs. 70.00 [60.00–80.00] years) and in the eICU-CRD (75.00 [66.00, 82.75] vs. 65.00 [56.00, 76.00] years). In the MIMIC-IV database and eICU-CRD, male patients in the cancer group accounted for 63.0% and 57.6%, respectively. while the non-cancer group accounted for 62.7% and 63.4%; Those in the cancer group weighed less than those in the noncancer group in the MIMIC-IV database (72.00[62.50,87.60] kg vs 80.00[68.00,94.43] kg) and in the eICU-CRD (76.94 [64.46, 90.07]kg vs 83.00 [70.00, 98.30]). Disease severity scores were higher in the cancer group than in the noncancer group in the MIMIC-IV database (52.00[40.00,70.00] vs 43.00[30.00,62.00]) and in the eICU-CRD (38.00 [27.00, 52.50]vs 32.00[23.00,47.00]). Fewer patients in the cancer group underwent PCI than those in the noncancer group in the MIMIC-IV database (31.6% vs 40.8%) and in the eICU-CRD (19.4% vs 25.8%). The remaining baseline characteristics are listed in detail in Table 1.

The Kaplan–Meier survival curve indicated that the probability of in-hospital survival of patients in the cancer group was lower than that of patients in the noncancer group (Fig. 2). In this study, variables include age, gender, weight, APS, first care unit, ventilator use, vasopressor use, dialysis use, PCI, CABG; comorbidities (CHF, CD, CPD, LD, diabetes, and RD); laboratory indicators (troponin, WBC, hemoglobin, platelets, potassium, sodium, bicarbonate, creatinine, BUN, and glucose); vital signs within 24 h of admission to the ICU (heart rate, BP, respiratory rate, and temperature), the Pearson’s coefficients between the two random continuous variables were all less than 0.5 (Figure S2), indicating that there was no significant correlation between them. After adjusting for potential confounders by multiple Cox regression, the risk of all-cause in-hospital mortality was significantly higher in the cancer group than in the noncancer group, and the HR (95% CI) values for the cancer group were 1.56(1.22,1.98) and 1.35(1.01,1.79) in the MIMIC-IV database and the eICU-CRD, respectively, corresponding to 1.56- and 1.35- fold higher risks of all-cause in-hospital mortality in the cancer groups, respectively (Table 2).

It can be observed in Table 1 that many variables such as age, weight, disease severity score, and PCI were all considerably unbalanced. However, after PSM these were all well balanced. The Kaplan–Meier survival curve indicated the same trend as the original population (Fig. 2). However, there were still some variables that differed between the groups (Table S1, Table S12). These variables were then again subjected to the Cox multiple regression, and the final results were consistent with those for the original population. In the MIMIC-IV database and the eICU-CRD, the all-cause in-hospital mortality risks of patients in the cancer groups were 1.83- and 1.54-fold higher, respectively, than those in the noncancer groups (Table 2).

The virtual populations obtained from the IPTW data set using multiple logistic regression were well balanced in the two databases. The results of the Kaplan–Meier survival curve were also consistent with the original population and the population obtained after matching; that is, the probability of all-cause in-hospital mortality was lower in the cancer than in the noncancer group (Fig. 2). These virtual populations were again subjected to Cox multiple regression, and a trend similar to that of the population is obtained. In the MIMIC-IV database and the eICU-CRD, the all-cause in-hospital mortality risks of the cancer groups were 1.54- and 1.51- fold higher than those of the noncancer groups, respectively (Table 2).

There was no significant interaction between the groups in the two databases (Fig. 3).