Novel Risk Model Predicts Long-Term Outcomes in Recurrent Pericarditis

A Cleveland Clinic-led team of clinicians and researchers has developed the first risk-stratification model for predicting long-term outcomes in patients with recurrent pericarditis. The model, developed from a retrospective study of 365 patients using both traditional and machine learning methods, showed high discriminative ability in an internal validation study. Its development and validation results were reported in the Journal of the American College of Cardiology (2024;84[13]:1193-1204) along with an accompanying editorial.

“We are hopeful that this novel machine learning model will improve clinicians’ ability to identify patients at elevated risk for recurrent pericarditis and initiate earlier anti-inflammatory therapy, with the end goal of reduced morbidity and better quality of life,” says the paper’s senior and corresponding author, Allan Klein, MD, Director of the Pericardial Diseases Center at Cleveland Clinic.

A debilitating condition with no prognostic tool

Recurrent pericarditis is a typically debilitating condition that affects as many as 30% of patients who experience acute pericarditis. The condition has heterogeneous phenotypes that demand a personalized treatment approach with close follow-up monitoring to prevent recurrences and keep symptoms at bay. In some cases, management may require immunosuppressive therapy.

“Identifying patients who are at risk for recalcitrant pericarditis is key for taking effective preventive measures and for timely delivery of needed treatment, including identifying candidates for novel immunomodulatory therapies or surgical pericardiectomy,” Dr. Klein notes. “However, there has been no established tool for predicting outcomes in the setting of recurrent pericarditis, so we undertook this effort to change that by leveraging Cleveland Clinic’s considerable experience base in caring for patients with this condition.”

Study design essentials

He and colleagues retrospectively reviewed 497 consecutive patients with recurrent pericarditis cared for at Cleveland Clinic from 2012 through 2019. After exclusion of patients with prior pericardiectomy, follow-up of less than six months. or secondary pericarditis (due to radiation, cancer or bacterial, metabolic or fungal causes), 365 patients were included in the analysis.

Across this cohort, the researchers assessed rates of clinical remission, which was defined as absence of additional recurrences, complete cessation of anti-inflammatory therapy, resolution of symptoms and absence of evidence of active inflammation.

They first performed manual feature selection to identify variables most associated with outcomes in this cohort, basing choices on clinical experience, prior studies and data availability. This yielded 34 clinical variables (demographics, medical history, physical exam findings, lab/imaging data, medications, etc.) that were then evaluated with five different machine learning survival models to calculate a score predicting the likelihood of clinical remission within five years.

“We decided to use machine learning-based models because of the heterogeneous nature of recurrent pericarditis cases and the data complexity inherent in the interplay amongst the many clinical, laboratory and imaging variables considered,” explains co-investigator Tom Kai Ming Wang, MBChB, MD, a Cleveland Clinic cardiologist who is part of the Pericardial Diseases Center team.

The overall data set was divided into a training set (70% of the cohort) and a test set (30% of the cohort), with the former used to develop and optimize the risk-stratification model and the latter used to assess and validate performance of the final model.

Results and the risk model at a glance

Patients in the cohort had a median follow-up of 35 months (interquartile range, 16-88 months) and a median of three recurrences at baseline. Breakdown by recurrent pericarditis etiology was 61% idiopathic, 21% postcardiac injury syndrome and 18% autoimmune.

Clinical remission was achieved by 118 patients, or 32% of the cohort, over the course of follow-up.

Ten variables were found to be most prognostically predictive: age, sex, number of recurrences at baseline, etiology, heart rate, severity of late gadolinium enhancement (LGE) on cardiac magnetic resonance (CMR), left ventricular ejection fraction (LVEF), steroid dependency, colchicine use and use of a disease-modifying antirheumatic drug (DMARD).

These 10 variables were included in the final risk-stratification model with the following score values to indicate the nature and relative value of their contribution to risk of recurrences:

• Age ≥ 55 years, –2 points (advanced age is protective)
• Female sex, 1 point
• > 3 baseline recurrences, 2 points
• Autoimmune etiology, 3 points
• Heart rate: 1 point for ≤ 60, 2 points for 61-80; 3 points for >80
• Moderate or severe LGE, 2 points
• LVEF > 60%, 1 point
• Colchicine use, 1 point
• Steroid use, 2 points
• DMARD use, 2 points

Patients were stratified into three risk groups based on their total score across all 10 variables:

• Low-risk group (score ≤ 3) (n = 62)
• Intermediate-risk group (score of 4 to 7) (n = 154)
• High-risk group (score ≥ 8) (n = 149)

The model predicted clinical outcomes with good accuracy, achieving a C-index of 0.800 on the test set. It also exhibited strong performance in risk stratification, as complete remission rates decreased proportionally with increasing risk scores. Kaplan-Meier curves showed significant differences in clinical remission rates among the high-, intermediate- and low-risk groups (log-rank test, P < .0001).

Cox regression analysis showed that, relative to the high-risk group, patients in the low-risk group had a 27-fold higher incidence of complete remission and patients in the intermediate-risk group had a sevenfold higher incidence of complete remission.

Big step forward, but more validation needed

In their study report, the researchers note that many of the prognostic variables of their risk model are consistent with features identified as prognostically important in prior research of recurrent pericarditis, although their study is bolstered by the largest sample size to date and a substantial follow-up period.

They also point out that their model identified a novel prognostic feature not recognized before — heart rate.

“Our use of multiple machine learning-based methods allowed us to determine the most influential combination of prognostic factors and define clinically significant thresholds for continuous parameters that are crucial for predicting outcomes,” Dr. Wang says.

Despite the model’s promising performance on internal validation, the authors note, its wider applicability still needs to be evaluated externally in multicenter studies, particularly since the model was developed based on data from a single tertiary center. Additionally, much of the study period predated recent advances in the use of CMR in patients with pericarditis and the emergence of targeted therapy with interleukin-1 blockers, which may have further improved rates of clinical remission.

“If external validation is provided by future prospective studies, our risk stratification model has potential to facilitate early and appropriate targeted intervention in patients at high risk of further recurrences,” Dr. Klein concludes. “That would be highly useful in easing the challenges of managing this complex patient population.”