Research map / Time series / Causal and generative time series
Stabilized Neural Prediction of Potential Outcomes in Continuous Time
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SCIP-Net, a neural CDE-based method with stabilized inverse propensity weights for estimating conditional average potential outcomes over time in continuous time while adjusting for time-varying confounding.
Problem
Estimating conditional average potential outcomes (CAPOs) of treatment sequences from irregularly timed patient trajectories, where existing neural methods either assume discrete time or fail to properly adjust for time-varying confounding in continuous time.
Approach
Derives a tractable continuous-time inverse propensity weighting formulation (via product integrals reduced to finite products) with stabilized weights, and implements it with neural CDE stability, weight, encoder and decoder networks where the decoder is trained with stabilized-weighted MSE.
Key results
- On MIMIC-III, SCIP-Net achieves RMSE 0.634±0.148 at 2 hours vs 0.784±0.145 for TE-CDE (+19.1% relative improvement).
- On tumor growth data, SCIP-Net outperforms all baselines, with gains growing under stronger time-varying confounding; it beats TE-CDE by a large margin.
- The unstabilized ablation CIP-Net is competitive but SCIP-Net's stabilized weights give clearly better results at larger horizons and strong confounding.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use stabilized weights instead of unstabilized weights (ablation) | SCIP-Net on MIMIC-III · MIMIC-III, 1 hour | RMSE: 0.876± 0.041 -> 0.877±0.044 | No clear effect |
| use stabilized weights instead of unstabilized weights (ablation) | SCIP-Net on MIMIC-III · MIMIC-III, 2 hours | RMSE: 0.785±0.117 -> 0.634±0.148 | Helped |
| use stabilized weights instead of unstabilized weights (ablation) | SCIP-Net on MIMIC-III · MIMIC-III, 3 hours | RMSE: 1.291±0.400 -> 1.089±0.322 | Helped |
| apply SCIP-Net rather than TE-CDE | MIMIC-III · MIMIC-III, 1 hour | RMSE: 0.915±0.025 -> 0.877±0.044 | Helped |
| apply SCIP-Net rather than TE-CDE | MIMIC-III · MIMIC-III, 2 hours | RMSE: 0.784±0.145 -> 0.634±0.148 | Helped |
| apply SCIP-Net rather than TE-CDE | MIMIC-III · MIMIC-III, 3 hours | RMSE: 1.240 ± 0.242 -> 1.089±0.322 | Helped |
| use SCIP-Net rather than existing neural methods | Tumor growth data · Tumor growth data, prediction horizons up to three days | RMSE | Helped |
| use stabilized weights instead of unstabilized weights (ablation) | SCIP-Net on tumor growth data · Tumor growth data, larger prediction horizons and strong confounding | RMSE | Helped |
Surprising findings
- The unstabilized-weight ablation CIP-Net is already competitive with the best baseline, and stabilization gives further gains mainly at longer horizons.
Limitations
- Only one continuous-time neural baseline (TE-CDE) exists, limiting comparison.
- Experiments use simulated outcomes and treatments on tumor growth and semi-synthetic MIMIC-III data rather than real ground truth.
- Relies on standard causal assumptions (consistency, positivity, unconfoundedness) that cannot be verified from data.
Experimental setup
Datasets: tumor growth simulation MIMIC-III (semi-synthetic extract)
Benchmarks: tumor growth simulation MIMIC-III semi-synthetic
Metrics: RMSE
Baselines: CT CRN RMSNs G-Net TE-CDE CIP-Net (ablation with unstabilized weights)
Builds on: neural controlled differential equations marginal structural models CRN TE-CDE
Closest papers by meaning
- Bayesian Neural Controlled Differential Equations for Treatment Effect Estimation (ICLR 2024 · similarity 0.85)
- Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time (NeurIPS 2026 · similarity 0.84)
- Orthogonal Survival Learners for Estimating Heterogeneous Treatment Effects from Time-to-Event Data (NeurIPS 2025 · similarity 0.82)
- Trajectory Flow Matching with Applications to Clinical Time Series Modelling (NeurIPS 2024 · similarity 0.82)
- MIRA: Medical Time Series Foundation Model for Real-World Health Data (NeurIPS 2025 · similarity 0.82)
- MOTOR: A Time-to-Event Foundation Model For Structured Medical Records (ICLR 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.