Machine Learning to Assess Macrophage Activation Syndrome in Still’s Disease: GIRRCS AOSD Study

14 Apr 20266 min read0 viewsJournal Feed

GIST

ObjectivesThis study aims to explore the application of machine learning techniques in assessing macrophage activation syndrome (MAS) in Still’s disease.MethodsA multicenter, observational, prospective study was conducted, including patients with Still’s disease enrolled in the Gruppo Italiano di Ricerca in Reumatologia Clinica e Sperimentale (GIRRCS) AOSD Study Group and the AutoInflammatory Disease Alliance (AIDA) Network Still’s Disease Registry.ResultsA total of 737 patients (age: 35.5 ± 17.8, male sex: 44.7%) with Still’s disease were assessed; 11.4% were affected by MAS, and 3% had a poor prognosis. First, random forest imputation was applied to the original dataset.

Subsequently, a machine-learning-driven assessment was developed to explore MAS occurrence. Collectively, regression models, an exploration decision tree, and a random forest were applied, suggesting the importance of ferritin, age, C-reactive protein (CRP), and systemic score.

A logistic regression model accounting for data leakage concerns was then generated using these variables, and missing values were imputed using random forest imputation. This analysis supported the role of the selected variables, which were further combined across different clinical scenarios to estimate the probability of MAS.

Clinical Editorial

Event-focused overview and study rationale

The report examines how machine learning approaches can be applied to assess macrophage activation syndrome (MAS) within Still’s disease, using data from multi-center cohorts.

Study design and population

A prospective, observational, multicenter study drew from the GIRRCS AOSD Study Group and the AIDA Network Still’s Disease Registry.
The cohort included 737 individuals with Still’s disease, with mean age 35.5 years and 44.7% male.

MAS burden and prognosis

MAS occurred in 11.4% of patients; 3% had a poor prognosis according to the report.

Analytical methods and modeling approach

Initial data handling employed random forest imputation for missing values.
A machine-learning framework was constructed to evaluate MAS risk, incorporating regression analyses, an exploration-oriented decision tree, and a random forest model.
A logistic regression model was subsequently developed with attention to data leakage concerns, using the same imputed dataset for stability.
The modeling effort explored the contribution of key variables and their interactions across varying clinical scenarios to estimate MAS probability.

Key variables driving MAS risk

The analyses highlighted several variables as influential: age, ferritin level, C-reactive protein (CRP), and systemic involvement as captured by a systemic score.
Notably, higher ferritin and CRP values, together with older age and greater systemic involvement, were repeatedly emphasized in the risk stratification.

Quantified risk in defined profiles

The highest MAS probabilities emerged for patients with: age ≥ 45 years, ferritin ≥ 4,178.1 ng/mL, CRP ≥ 27.15 mg/L, and a systemic score ≥ 7, with an estimated MAS probability of 34.7%.
A related profile—ferritin ≥ 4,178.1 ng/mL, CRP ≥ 27.15 mg/L, and systemic score ≥ 7—also yielded a substantial MAS probability of 33.5%.
These estimates derive from combining selected features across clinical scenarios to yield a clinician-oriented risk estimate.

Contextual interpretation and limitations

The study supports a potential role for age at onset, hyperferritinaemia, elevated CRP, and multiorgan involvement in MAS risk within Still’s disease.