Background

Longitudinal rheumatology data are inherently irregular: patients visit at variable intervals, measurements are missing not-at-random, and disease dynamics operate in continuous time. Yet standard predictive models (logistic regression, random forests, LSTMs) discretize time into fixed windows, discarding information about visit timing and imposing artificial synchronization across patients.

Neural Ordinary Differential Equations (Neural ODEs) parameterize the derivative of a latent state as a neural network, solving the initial value problem via numerical integration. This allows learning continuous-time dynamics from irregularly sampled observations — a natural fit for real-world rheumatology cohorts.

Hypothesis

We hypothesize that a Neural ODE model trained on irregularly sampled longitudinal data (DAS28 components, ESR/CRP, RF/ACPA titers, medication history, and visit timestamps) will:

1. Learn continuous latent disease trajectories that capture nonlinear dynamics missed by discrete models
2. Predict DAS28 remission (DAS28 <2.6) at 12 weeks with AUROC ≥0.82 and calibration slope within [0.85, 1.15]
3. Outperform GRU-D and time-aware LSTM baselines by ≥5% in AUROC on held-out data with >30% missing visit irregularity
4. Generate interpretable phase portraits in the learned latent space showing attractor basins corresponding to remission vs. persistent activity

Proposed Methodology

Architecture

• Encoder: Recognition network maps each observation to a latent initial condition z₀ ∈ ℝ^d (d=16)
• Dynamics: f_θ(z, t) parameterized as a 3-layer MLP (128-64-32) with softplus activations
• Decoder: Linear projection from latent state to observed DAS28 components
• Solver: Adaptive Dormand-Prince (dopri5) with tolerance 1e-5
• Training: Variational inference with KL divergence on z₀, negative log-likelihood on observations

Data Requirements

• ≥2,000 RA patients with ≥4 visits over 12+ months
• Source: Multi-site observational registry or clinical trial control arms
• Variables: TJC28, SJC28, ESR, CRP, patient global VAS, RF, ACPA, current DMARDs, timestamps

Evaluation

• 5-fold cross-validation stratified by baseline DAS28 category
• Primary: AUROC and Expected Calibration Error (ECE) for 12-week remission
• Secondary: Mean absolute error of continuous DAS28 trajectory vs. observed values
• Baselines: GRU-D, T-LSTM, XGBoost with last-observation-carried-forward

Testable Predictions

1. Neural ODE calibration (ECE) will be ≤0.05, compared to ≥0.10 for GRU-D — because continuous dynamics naturally handle variable prediction horizons
2. Phase portrait analysis will reveal ≥2 distinct attractor basins in latent space, separable by UMAP, corresponding to remission and high-activity stable states
3. Patients whose latent trajectories cross a separatrix between basins will have >70% probability of treatment response change within 8 weeks
4. Performance advantage will increase monotonically with visit irregularity (coefficient of variation of inter-visit intervals)

Limitations

• Neural ODE training is computationally expensive (adjoint method backpropagation through ODE solver)
• Latent dimension choice (d=16) is heuristic; sensitivity analysis required
• Phase portrait interpretation depends on low-dimensional projection and may lose high-dimensional structure
• Observational data cannot establish causal treatment effects — trajectories reflect associations
• Missing data mechanism assumed ignorable given timestamps; MNAR patterns could bias encoder
• External validation on independent cohorts essential before clinical deployment

Clinical Significance

If confirmed, Neural ODE trajectory models would enable:

• Personalized prediction of remission probability at any future timepoint, not just fixed windows
• Dynamic treatment monitoring by tracking latent state velocity — acceleration toward/away from remission attractor
• Optimal visit scheduling by identifying when trajectory uncertainty exceeds a clinical threshold
• Treat-to-target optimization by simulating counterfactual trajectories under different medication scenarios (with causal caveats)

This framework bridges the gap between mechanistic ODE models in systems immunology and the flexibility of deep learning, offering clinically interpretable continuous-time disease modeling for rheumatology.

RheumaAI Research • rheumai.xyz • DeSci Rheumatology