Optimal Stopping Theory With Pharmacogenomic State Augmentation Minimizes Cumulative Disease Burden in Sequential Biologic Switching for Rheumatoid Arthritis

Hypothesis

We propose that sequential biologic switching decisions in rheumatoid arthritis (RA) can be formalized as an optimal stopping problem on a pharmacogenomic-augmented state space, and that the resulting policy significantly reduces cumulative disease burden (time-integrated DAS28) compared to current treat-to-target (T2T) heuristics.

Background

Current T2T guidelines recommend switching biologics after 3–6 months of inadequate response, but the decision of when to switch and to what remains largely empirical. Meanwhile, pharmacogenomic markers (HLA-DRB1 shared epitope alleles, FCGR3A-158V/F polymorphisms affecting anti-TNF clearance, NAT2 acetylator status, CYP3A4 variants) carry predictive information about biologic-specific response trajectories that is rarely integrated into switching decisions.

Formal Framework

Define patient state at time t as S(t) = (X_clin(t), G, H(t)) where X_clin captures current disease activity (DAS28, CRP, ESR, swollen/tender joint counts), G is the static pharmacogenomic profile, and H(t) is treatment history. The switching decision becomes an optimal stopping problem: at each decision epoch, the clinician chooses to continue the current biologic or stop and switch.

The optimal policy minimizes:

J = E[∫₀ᵀ DAS28(t)dt + Σᵢ cᵢ · 𝟙(switch at tᵢ)]

where cᵢ represents switching costs (washout flares, new adverse event risk, immunogenicity reset). The pharmacogenomic vector G modifies transition probabilities between disease activity states under each biologic, enabling prospective estimation of response likelihood before switching.

Testable Predictions

1. Primary: In simulation using published pharmacogenomic effect sizes (FCGR3A on rituximab response, HLA-DRB1 on anti-TNF), the optimal stopping policy reduces time-integrated DAS28 by ≥15% over 5 years compared to fixed 6-month T2T switching.
2. Secondary: The policy identifies a subgroup (~20–30% of patients) where early switching at 8–12 weeks — before conventional 3-month assessment — is optimal due to pharmacogenomic contraindications.
3. Tertiary: Adding pharmacogenomic state augmentation to the stopping rule improves the policy value by ≥8% over clinical-only optimal stopping, quantifying the information value of genotyping.

Methods to Test

• Phase 1 (in silico): Construct a semi-Markov model of DAS28 trajectories under 5 biologic classes (anti-TNF, anti-IL6, anti-CD20, CTLA4-Ig, JAKi), parameterized from published RCT data and pharmacogenomic substudies. Solve the optimal stopping problem via backward induction on discretized state space.
• Phase 2 (retrospective): Apply the derived policy to longitudinal registry data (e.g., CORRONA, BIOBADAMEX) with available pharmacogenomic data, comparing counterfactual cumulative DAS28.
• Phase 3 (prospective): Pragmatic cluster-randomized trial of pharmacogenomic-informed optimal stopping vs standard T2T.

Limitations

• Pharmacogenomic effect sizes for newer biologics (JAKi, IL-17i) are less well-characterized than for anti-TNF/rituximab.
• The semi-Markov assumption may not capture non-linear dynamics in disease trajectory; extensions to continuous-time stochastic control (Hamilton-Jacobi-Bellman) could address this.
• Immunogenicity and anti-drug antibody development introduce path-dependent state transitions not fully captured by Markovian models.
• Registry data may have selection bias in switching patterns.

Clinical Significance

If validated, this framework provides a mathematically rigorous, personalized switching algorithm that integrates genomic information at the point of care. The information-value calculation (prediction 3) directly quantifies whether pharmacogenomic testing is cost-effective for a given patient profile — bridging precision medicine theory with health-economic decision-making.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology