A decision tree-based algorithm for structured risk stratification of rare rheumatic diseases in a tertiary referral...

Rare inflammatory rheumatic diseases are often characterized by heterogeneous, multisystemic symptom patterns, complicating early diagnostic differentiation. This study aimed to model a structured, symptom- and laboratory-based decision approach for risk stratification of rheumatologic diagnoses in patients referred to a tertiary Center for Rare Diseases (CRD) with unclear systemic complaints.

We conducted a retrospective cross-sectional analysis of 173 patients evaluated at a CRD. Patients were classified into rheumatologic (RHEUMA) and non-rheumatologic (OTHER) diagnostic groups based on final diagnostic outcomes. A standardized questionnaire capturing 52 symptoms was aggregated into domain-specific and composite scores. Laboratory data were summarized into predefined indices. Group differences were analyzed descriptively using inferential statistics. A decision tree model based on Chi-Square Automatic Interaction Detection (CHAID) was constructed to support structured risk stratification.

A confirmed rheumatologic diagnosis was established in 52.0% of patients. Symptom and comorbidity patterns were broadly similar across diagnostic groups, with fatigue and generalized pain being the most prevalent complaints. The RHEUMA group showed significantly higher scores in selected symptom domains and in an immunoserological laboratory index (all p < 0.05). The CHAID-based decision model, integrating symptom scores, laboratory markers, and autoimmune history, showed an apparent classification of 81.5% (AUC = 0.893), compared to 76.3% (AUC = 0.823) for logistic regression.

The proposed decision tree model provides a transparent framework for structured risk stratification in a highly preselected tertiary referral population. Given the monocentric, retrospective, and exploratory study design, the findings should be interpreted as hypothesis-generating. External validation in independent cohorts is required to assess model generalizability, calibration, and robustness before clinical implementation can be considered. Within these constraints, the approach illustrates the potential of transparent, rule-based symptom aggregation to support structured clinical reasoning and prioritization in complex multisystem presentations.