Medical and pharmacy data, notes, laboratory results, and cardiovascular imaging were obtained from the Indiana University (IU) Health Data Warehouse and INPC. The INPC, managed by the Indiana Health Information Exchange, is one of the largest health information exchanges in the country (24). Over 123 separate clinical entities contribute data representing 19 million patients, billions of clinical observations, and over 300 million clinical text reports. The INPC contains data for two-thirds of Indiana's population—a proportion that rises to 80%–100% in central Indiana where the major health systems are located (25, 26). INPC data were complemented with data from the IU Health Data Warehouse, which contains structured data from IU Health's Cerner electronic health records (EHR). As of late 2022, this database captures data on 5.6 million patients, representing 16 hospitals and numerous outpatient clinics.

Eligible patients over the age of 35 at the time of thymic cancer diagnosis were identified by ICD-9 (164.0, 212.6) and ICD-10 (C37, D15.0) codes for thymoma, thymic carcinoma, and thymic neuroendocrine tumor (Figure 1; see Supplementary Material for the complete ICD codes list). Patients with fewer than 2 ICD diagnoses of thymic tumor in the EHR were excluded, and data were collected for two years prior to the index date. We identified 1,212 patients with thymic tumors from January 1, 2003, to December 31, 2020. Selecting patients residing in Indiana or a neighboring state allowed for closer follow-up and ensured adequate data collection but reduced the number to 680. Cancer characteristics were extracted from EHR data, including WHO staging, Masaoka staging, size of tumor, cancer therapies, and presence of a paraneoplastic syndrome. The tumor size was based on a CT scan at the time of diagnosis or final pathologic evaluation. Cardiovascular risk factors evaluated included diabetes mellitus, hypertension, dyslipidemia, obesity classified as BMI >30, and a history of tobacco use. Patients were noted to have pre-existing coronary disease if they had a diagnosis of coronary atherosclerosis/calcification from a CT scan or coronary angiogram within 1 year of thymic cancer diagnosis. Cardiac events included heart failure (including preserved and reduced ejection fraction), myocardial infarction or ischemic heart disease, stroke or transient ischemic attack, and atrial fibrillation. Patient mortality was ascertained from hospital discharge summaries and the Indiana State Department of Health's death certificate files.

All 680 patient records were individually reviewed by the first author (AK). Records were ascertained for the accuracy of diagnoses and cancer characteristics based on pathology data, imaging, and oncology notes. The list was cross-referenced with a separate internal pathologic database of patients from 2010 to 2020 (n = 155, 52.36%) to confirm thymic neoplasm. To verify the accuracy of the chart review, a blinded reviewer (SCC) was given records of 20/680 patients, half of whom had tissue biopsy confirming diagnosis and half did not, and accurately verified 19/20 (95.0%) of the diagnoses. Unclear diagnoses (n = 22) were adjudicated by two authors (PJL, SB) based on pathologic data. Cardiovascular risk factors and outcomes were captured from various settings including inpatient and outpatient clinical encounters, and at different Indiana institutions through the INPC. Although there were cancer patients from neighboring states, they continued to receive subspecialty care at the Simon Cancer Center years after their cancer diagnosis, providing reliable estimates of long-term outcomes.

A control cohort consisted of patients meeting the following criteria: no prior cancer diagnoses and at least 2 years of clinical data in the system prior to an encounter occurring the same year as a thymic cancer patient (TCP) index date. The study index date of a cancer patient was defined as the date of the first documented diagnosis of thymic neoplasm. The index date of a control patient was the first visit date in the same calendar year as the index date of the matched patient. Controls were randomly selected and were age and sex-matched to TCPs at a 10:1 ratio. Data extraction was repeated for the control cohort.

Demographic and clinical characteristics of both cohorts are presented using the mean, standard deviation, or count unless otherwise specified, and between-group comparisons used the unpaired t-test or chi-squared test, respectively. Incident events (cardiac or death) were observed from the index date and censored at the last recorded encounter at IU Health. These events could be predicted by the multiple CV risks at baseline, which were then grouped into three risk levels (low, medium, high) to streamline subsequent analyses (see Supplementary Material Table S5). Time-to-event comparisons between groups were graphed using Kaplan-Meier curves and analyzed using causal-specific Cox regression (results presented as hazard ratio [HR] with 95% confidence interval [CI]), adjusted for age, sex, race, and the 3 CV risk levels. The cumulative incidence plots (Figure 3) showed potential change-points of the proportional hazards at around three years and 7 years. With the data censored at 7 years due to insufficient number of later events, the model with time-dependent hazards (with change-point at three years) showed significantly better fit than the constant hazards model (p < 0.025).

Of the 1,212 patients identified with a thymic cancer diagnosis, 680 met the inclusion criteria. After excluding patients misclassified as having thymic cancer, duplicate entries from different health systems, patients with insufficient follow-up data, and patients without pathology-confirmed thymic cancer, the final cohorts consisted of 296 TCPs and 2,960 noncancer controls (Figure 1). Most TCPs had thymoma (74.0%), followed by thymic carcinoma (22.3%) and thymic neuroendocrine tumor (3.7%) (Table 1). Of 271 TCPs with Masaoka staging, 17.9%, 20.3%, 12.8%, and 40.5% were in stages 1–4, respectively. Paraneoplastic syndromes were present in 112 (37.8%) TCPs and 72 (24.3%) had myasthenia gravis. Treatments included surgical resection (76.4%), chemotherapy (50.3%) with 98 (33.1%) receiving anthracyclines, chest radiotherapy (32.4%), and immunotherapy (21.0%) (Figure 2).

Age and sex distributions were similar for TCPs and controls (due to matching), and both groups were about 82% White (Table 2). Of the 296 patients, 198 (67%) were in Indiana and 98 (33%) were from neighboring states.

Compared to controls, TCPs had similar rates of baseline obesity, tobacco use, and pre-existing CVD but significantly lower rates of hypertension, dyslipidemia, and diabetes mellitus (Table 2) and lower total number of CV risk factors per person. Preliminary analyses of TCPs showed these risk factors were predictive of subsequent cardiac events, and TCPs could be grouped into three levels of overall CV risk: 1) low-risk (no risk factors or pre-existing CVD); 2) medium risk (1–2 risk factors and no pre-existing CVD); and 3) high risk (>2 risk factors and/or pre-existing CVD).

Over a mean follow-up period of 4.7 years (median = 4.4, range 0.5–16 years), the following events occurred: 31 TCPs and 346 controls developed heart failure (10.5% vs. 11.7%); 33 TCPs and 286 controls developed atrial fibrillation/flutter (11.2% vs. 9.7%); 14 TCPs and 195 controls had a stroke (4.7% vs. 6.6%); 15 TCPs and 259 controls had an ischemic event/myocardial infarction (5.1% vs. 8.8%); 63 TCPs and 686 controls had any cardiac event (21.3% vs. 23.2%); and 19 TCPs and 91 controls died (6.4% vs. 3.1%) (see Supplementary Material Table S5).

The Kaplan-Meier curves (Figure 3) show that within TCPs and within noncancer controls, the incidence of CV events increases with the levels of baseline CV risk (low, medium, high) and within each risk level, the incidence of CV events was higher in TCPs than in controls. Results of the final time-dependent proportional hazards model showed that higher-incident cardiac events could be predicted by increasing age (HR = 1.05 per year) and male sex (HR = 1.37 relative to female) (Table 3). The 3 CV risk levels at baseline were also strongly predictive of the incidence of cardiac events, with HR of 1.45 (95% CI 1.11–1.90, p = 0.007) for medium risk and 3.73 (95% CI 2.88–4.83, p < 0.001) for high relative to low risk. The hazard ratios of TCPs relative to controls were 1.38 (95% CI 1.01–1.87, p = 0.042) in years 0–3 and 0.77 (95% CI 0.47–1.25, p = 0.29) in years 4–7.

To assess the validity of our data source and approach, we evaluated the known association of heart failure with anthracycline therapy in our dataset. In our cancer cohort (12 heart failures among 98 anthracycline-treated patients and 19 heart failures among 198 non-treated patients), anthracycline-treated patients had significantly higher overall heart failure events (HR = 2.77, 95% CI 1.26–6.06, p = 0.011) after controlling for demographics and baseline CV risk. Among classes of cancer therapies, immunotherapy showed no significant association with cardiac events among the cancer patients with time-dependent analysis. The effects of radiotherapy and chemotherapy were time-dependent. There was a significantly higher incidence of any cardiac event with radiotherapy and chemotherapy treatments compared to untreated cancer patients, but this difference did not manifest until after the first three years (Table 4). The data were too limited to investigate the effects of treatment combinations.

Secondary analyses comparing TCPs with and without a paraneoplastic process after cancer diagnosis showed no significant difference in adverse CVD or death (HR 0.71, 95% CI 0.41–1.23, p = 0.22; HR 1.24, 95% CI 0.5–3.09, p = 0.64).

Our study had several limitations. It featured a single quaternary referral center's experience and included many patients with advanced-stage tumors. This may have skewed the number of cardiac events. Less complete historical data from referral patients may have biased our results. There may have been bias based on the geographical distribution of referrals as rural areas tend to provide less data to the INPC. We may have limited data prior to the index date and missed pre-existing conditions. Selection bias may have resulted from evaluating only patients over 35 and limiting the cohort to patients from Indiana or a neighboring state. By limiting the cohort geographically, we aimed to obtain comparatively more complete data with longitudinal follow-up due to the high penetration of our health information exchange. To ensure accuracy, we minimized potential recall and misclassification biases using both ICD coding and physician review of EHR data. We strove to have accurate diagnoses by only including patients with pathologic diagnoses of thymic neoplasm and confirmed with a subset of patients with a second pathology database. This led to smaller numbers of patients in the study, decreasing the power for multivariable analyses. Patients with thymic cancers in our institution have not traditionally had extensive cardiac workups prior to initiating treatment or undergoing surgical resection, which may have led to underestimation of baseline CVD. We could not separate exposure from treatments, such as concurrent chemotherapy and immunotherapy, to analyze cardiac events which likely led to confounding. We had similar distribution of thymic cancer subtypes and similar incidence of cardiac events compared to the published literature. Our subanalysis for heart failure in TCPs who received anthracyclines also reinforces validity. Since this was an exploratory study, we did not perform propensity score matching. We cannot make causal inferences but were able to highlight associations between baseline CV risk factors and higher incidence of adverse CVD in TCPs.