Hypothesis

Targeted Maximum Likelihood Estimation (TMLE) combined with Super Learner ensemble prediction, applied to longitudinal observational rheumatology cohorts with time-varying confounders (disease activity, comorbidities, concomitant medications), can identify optimal dynamic treatment regimes (DTRs) for biologic sequencing in rheumatoid arthritis that reduce 12-month Sharp/van der Heijde radiographic progression scores by >30% compared to current guideline-driven static treatment protocols.

Background and Rationale

Current rheumatoid arthritis treatment follows a step-up paradigm guided by composite disease activity scores (DAS28, CDAI), but these protocols treat patients as exchangeable within risk strata and do not formally account for time-varying confounding affected by prior treatment — a well-documented source of bias in observational treatment comparisons. Marginal structural models (MSMs) with inverse probability weighting partially address this but suffer from practical positivity violations and weight instability in high-dimensional covariate spaces.

TMLE offers a doubly-robust, semiparametric efficient alternative that combines:

1. Super Learner — a cross-validated ensemble of machine learning algorithms (gradient boosting, random forests, LASSO, neural networks) for initial outcome and treatment mechanism estimation
2. Targeted bias reduction — a fluctuation step that targets the specific causal parameter of interest (here, the counterfactual mean outcome under candidate DTRs)
3. Valid statistical inference — influence function-based confidence intervals that remain valid even when nuisance parameters are estimated at slower-than-parametric rates

Crucially, TMLE with longitudinal extensions (LTMLE) can handle the sequential treatment decisions inherent in biologic switching by iteratively updating predictions backward through time, properly adjusting for time-varying confounders affected by prior treatment.

Testable Predictions

1. Primary: LTMLE-identified optimal DTRs will demonstrate ≥30% relative reduction in 12-month ΔSharp/van der Heijde score compared to observed treatment patterns in a held-out validation cohort (n≥500)
2. Secondary: Super Learner cross-validated risk estimates will achieve C-statistic >0.75 for 6-month radiographic progression, outperforming any single algorithm by ≥0.03 AUC
3. Tertiary: Positivity diagnostics will show <5% of observations with estimated treatment probabilities below 0.05 (practical positivity), versus >15% under standard IPW-MSM approaches in the same data
4. Mechanistic: TMLE-identified optimal switching rules will correlate with pharmacogenomic features (CYP450 metabolizer status, HLA alleles) with mutual information >0.15, suggesting biologically plausible effect modification

Proposed Methodology

• Data: Longitudinal registry data (CORRONA, RABBIT, or equivalent) with ≥3 years follow-up, serial radiographs, medication histories, and available pharmacogenomic substudy data
• Super Learner library: XGBoost, random forest, elastic net, GAMs, recurrent neural networks, Bayesian additive regression trees
• Causal parameter: E[Y_d] for candidate DTRs d defined by DAS28 thresholds, prior treatment failures, and optional pharmacogenomic strata
• Cross-validation: V-fold (V=10) for Super Learner; sample splitting for TMLE inference
• Sensitivity analysis: E-values for unmeasured confounding; multiple robustness checks via different Super Learner libraries

Limitations

• Registry data may lack complete radiographic follow-up, introducing informative censoring addressable via IPCW-TMLE but requiring additional modeling assumptions
• Practical positivity violations remain possible in rare treatment sequences despite TMLE mitigating weight instability
• Pharmacogenomic data availability is limited in most registries, restricting the mechanistic prediction to substudy populations
• External validity depends on registry representativeness — single-country registries may not generalize across healthcare systems
• Computational cost of LTMLE with large Super Learner libraries at multiple time points is substantial (estimated 48–72h on 64-core cluster)

Clinical Significance

If confirmed, this approach would provide the first doubly-robust, semiparametric efficient framework for optimizing biologic sequencing in RA that:

• Formally handles time-varying confounding without the fragility of IPW-based methods
• Leverages modern machine learning while preserving valid frequentist inference
• Generates individualized dynamic treatment rules interpretable by clinicians as decision trees
• Could be implemented as a DeSci-native federated protocol, where TMLE estimation occurs at each site and influence curves are aggregated — enabling multi-registry analysis without sharing patient-level data

This bridges the gap between the theoretical promise of causal machine learning and practical clinical decision support in rheumatology.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology