Development of a prognostic model for Macrophage Activation Syndrome in Systemic Juvenile Idiopathic Arthritis

Introduction: Macrophage activation syndrome (MAS) is a potentially life-threatening complication of systemic juvenile idiopathic arthritis (SJIA) characterized by heterogeneous organ involvement and severity. Early identification of patients at high risk of complicated clinical course may improve outcome by helping initiate prompt, appropriate immunosuppressive and supportive treatments. Yet, despite recent progress in clarifying the underlying immunological mechanisms, factors driving organ damage and severe outcome are not entirely understood, nor has the prognostic value of routinely gathered clinical and laboratory factors been fully explored. Objectives: To develop a prognostic model for SJIA-MAS based on routinely available parameters at disease onset, accounting for patient heterogeneity, possible latent factors, non-linear relationships and confounders. Methods: We examined a retrospective multinational cohort of 362 patients diagnosed with SJIA-MAS. The relationships between demographic, laboratory features at MAS onset (such as hemoglobin, whole blood cells, platelets, ERS, CRP, AST, ALT, bilirubin, fibrinogen, d-dimer, ferritin and creatinine), therapeutic interventions and outcomes were analyzed. Outcomes of interest included a “severe course” (defined as ICU admission or death), occurring of organs failure and CSN dysfunction. To identify potential phenotypes related to clinical features and outcome, we explored laboratory parameter patterns at MAS onset through Latent class modeling, which detects multiple unobserved clusters in heterogeneous populations. A structural causal approach was then used for investigating causal pathways leading to severe outcomes. Directed acyclic graphs (DAGs) were employed to depict possible causal relationships between the candidate biomarkers, potential confounding variables, and the outcomes, and inform the choice of adjustment sets in multivariate regression models. We assessed the possible relationships between variables and outcomes by penalized likelihood logistic regression and identified optimal cut off points for prognostic factors using Multiple Adaptive Regression Splines (MARS) and Classification and Regression Trees (CART). To account for possible treatment confounders, the effect of cyclosporine and etoposide use on outcomes was estimated using augmented inverse probability weighting (IPW) with double robust methods. Finally, results from previous analyses were incorporated in a probabilistic framework through a Bayesian network (BN) model, which provides risk estimates for specific clinical scenarios and quantifies the amount of information contributed from the identified prognostic variables. Results: The latent class model revealed six clusters based on biomarkers at MAS onset, characterized by the following features: mild alterations of white blood cells, platelets, fibrinogen, d-dimer and ferritin values, considered the baseline type (cluster 1, n =115); hyperferritinemia with low organs involvement (cluster 2, n = 101); elevation of inflammatory markers (cluster 3, n =51); hepatobiliary involvement (cluster 4, n = 41); severe pancytopenia, liver and kidney failure with higher elevation of LDH, d-dimer, ferritin (cluster 5, n = 30); biliary and renal dysfunction (cluster 6, n = 24). Cluster 2 and 3 presented lower age and SJIA duration at MAS onset compared to other subgroups. Cluster membership was predictive of severe course (p<0.001), CSN involvement (p<0.001), Hemorrhagic complications (p <0.001) and Heart failure (p<0.001), with patients in cluster 5 showing the highest risk of severe course and heart failure, and increased occurrence of CNS and Hemorrhagic manifestations in both cluster 5 and 6. In multivariate regression models, parameters at onset associated with risk of severe course were creatinine (OR 1,6 [95% CI 1.13–2.3]; p = 0.008) and albumin levels (OR 0,65 [95% CI 0.44–0.98]; p = 0.044) Higher risk of CNS involvement was found for patients younger at MAS onset (OR 0,62 [95% CI 0.42–0.92]; p = 0.018). Na (OR 0.0,89 [95% CI 0.82–0.96]; p = 0.006) and creatinine values (OR 1.69 [95% CI 1.14–2.5]; p = 0.009) were identified as independent predictors of mortality. There was no evidence for an effect of etoposide (OR 1.03 [95% CI 0.91–1.12]) and cyclosporine (OR 1.04 [95% CI 0.92–1.19]) on severe course. BNs defined distinct groups with different probability of severe outcomes, achieving a c-index of 0.76 for mortality, 0.81 for severe course and 0.81 for CNS involvement. Adding the obtained latent clusters to the BN model increased the prediction accuracy for severe course up to a c-index of 0.83. Based on information theory metrics (mutual information) from the BN model, decision algorithms for each outcome and a web-based decision support tool for external users were implemented. Conclusions: We developed a probabilistic prognostic model of SJIA-MAS based on routinely available data. This stratification tool may facilitate informed decision-making about the clinical management of these patients. The probabilistic and information-theoretic approach offers a framework for further validation, expansion and integration of the model with emerging molecular biomarkers.