Background

Lupus nephritis (LN) class transitions — particularly from Class III to Class IV or from proliferative to membranous phenotypes — carry major therapeutic implications but are currently detected only by repeat renal biopsy. Serial laboratory panels (complement C3/C4, anti-dsDNA, proteinuria, serum creatinine, urine protein-to-creatinine ratio) contain temporal interaction signatures that conventional time-series methods fail to capture because they treat channels independently or assume linear dynamics.

Hypothesis

We hypothesize that rough path signatures — iterated integrals of multivariate laboratory time series truncated at depth 3–4 — encode non-commutative interaction effects between serological and renal biomarker trajectories that discriminate impending LN class transitions from stable disease 8–16 weeks before histological confirmation, achieving AUROC >0.82 in a prospective validation cohort.

Theoretical Framework

Rough path theory (Lyons, 1998) provides a universal nonlinear feature map for sequential data. The path signature S(X)₀,T of a d-dimensional time series X captures all order statistics of the path, including cross-channel interactions (e.g., the integral of dC3 against d(proteinuria)) that encode how one biomarker responds to changes in another over time. Key properties:

1. Universality: Signatures form a complete set of features for continuous functions on path space
2. Non-commutativity: The signature distinguishes the order in which biomarker changes occur — critical for immune cascade dynamics
3. Reparametrization invariance: Robust to irregular sampling intervals common in clinical practice
4. Dimensionality: For d=6 biomarkers at truncation depth 4, the signature yields ~1,555 features, manageable with L1-penalized classification

Proposed Methodology

1. Cohort: Longitudinal LN patients (n≥200) with ≥2 renal biopsies and ≥12 months of serial labs (monthly or biweekly)
2. Signature computation: Log-signature transform (iisignature library) at depth 3–4 with lead-lag augmentation to capture quadratic variation
3. Classification: Elastic-net logistic regression on signature features, with nested cross-validation
4. Comparator models: LSTM, temporal convolutional network, and linear mixed-effects models on the same input data
5. Primary endpoint: AUROC for class transition prediction at 8, 12, and 16-week horizons
6. Calibration: Platt scaling with Brier score evaluation

Testable Predictions

• P1: Depth-3 log-signatures of the (C3, C4, anti-dsDNA, proteinuria, creatinine, UPCR) path achieve AUROC >0.82 for class transition prediction at 12-week horizon
• P2: Cross-channel signature terms (particularly ∫dC3⊗d(proteinuria) and ∫d(anti-dsDNA)⊗d(creatinine)) contribute >40% of the discriminative information, as measured by permutation importance
• P3: Signature-based models outperform channel-independent LSTM by >5 AUROC points due to superior capture of inter-biomarker temporal dependencies
• P4: The signature approach maintains performance (AUROC >0.78) under realistic missing-data conditions (up to 30% randomly missing timepoints) due to reparametrization invariance

Limitations

• Signature feature dimensionality grows exponentially with depth and channels; truncation at depth 4 may miss higher-order interactions
• Requires relatively dense longitudinal sampling (≥monthly); sparse follow-up may degrade signature estimation
• Histological ground truth depends on biopsy timing decisions, introducing verification bias
• Single-center development risks overfitting to local laboratory assay characteristics
• Does not model treatment effects as time-varying confounders; causal interpretation limited

Clinical Significance

A non-invasive, signature-based early warning system for LN class transition could: (a) reduce unnecessary surveillance biopsies by identifying stable patients, (b) trigger timely therapeutic escalation 2–4 months earlier than current practice, and (c) provide a mathematically principled framework for integrating irregularly sampled multivariate clinical data in rheumatology. The rough path approach is particularly suited to rheumatological time series where the interaction dynamics between biomarkers — not just their individual trajectories — drive disease evolution.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology