JMIR Publications

ABSTRACT

Background:

Estimating the magnitude of effect of immunosuppressive therapy in autoimmune disease and transplantation medicine as a single numeric score in tabular data is highly valuable, especially when developing statistical or machine learning models for prognosis, risk stratification, and other tasks.

Objective:

We aimed to derive a single continuous score that represents a patient’s overall immune status at a given time after exposure to immunosuppressive therapy.

Methods:

We developed an immunosuppressive intensity (ISI) score model to estimate point-in-time immunosuppressive state as a continuous cumulative score ranging from 0 to 1. Model structure and parameterization were informed by a structured expert elicitation process using a modified Delphi approach across commonly used immunosuppressive therapies. A base ISI model was implemented as a scaled and shifted sigmoidal ISI score function incorporating three parameters: A (starting intensity), n (decay rate), and d (half-life, 50% pharmacodynamic effect). Age and lymphocyte/CD19 B-cell counts were then incorporated to generate a biomarker-informed adjusted ISI score. Finally, we developed a cumulative ISI score to model the immunosuppression state when multiple medications are active contemporaneously. We evaluated the model in three international ANCA-associated vasculitis cohorts (RITA Ireland vasculitis (RIV) registry, IDIBELL registry, and Chapel Hill registry) for biological alignment, clinical plausibility across disease phases, and utility in relapse-risk modeling compared with conventional categorical treatment encoding, using a generalized estimating equation (GEE) model. The biological alignment was further evaluated by correlating the ISI score with Torque Teno virus (TTV) count, a marker of immunosuppression.

Results:

Following a Delphi process, we defined parameters for the base and adjusted the ISI score across a range of intravenous and continuous oral immunosuppressive medications. The adjusted cumulative ISI score showed stronger biological alignment and tracked disease stage appropriately. The median adjusted cumulative ISI scores were close to 1 during the peri-diagnosis phase (except for pre-treatment encounters), dropped to around 0.5 in remission, and around 0.3 in relapse encounters across the three AAV cohorts, thereby supporting clinical plausibility. The correlation between TTV count and cumulative ISI score was slightly stronger for the adjusted score than the base (unadjusted) (r=0.37 vs 0.35; both p<0.001). Therefore, subsequent clinical analyses focused on the adjusted score. Among the GEE models, the model including the adjusted cumulative ISI score had the lowest QIC, compared with the unadjusted ISI score model and the categorical treatment indicator model, indicating better relative model fit.

Conclusions:

We describe, for the first time, a pragmatic ISI score to represent a patient's overall immunosuppressive treatment status at a specific time point. This provides a reusable treatment-state variable for clinical analytics and prognostic modeling and represents a first step toward biomarker-enriched digital twins of immunosuppressive state for future decision support and translational digital medicine applications.