Research map / Time series / Causal and generative time series
IVP-VAE: Modeling EHR Time Series with Initial Value Problem Solvers
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 IVP-VAE, which models irregular EHR time series purely by solving parallel initial value problems (no RNN), reusing one invertible IVP solver as both VAE encoder and decoder.
Problem
Continuous-time VAE models for irregularly sampled EHR time series (Latent-ODE, Latent-Flow) process observations sequentially with RNNs plus IVP solvers, making training and inference slow and hard to parallelize.
Approach
Replace the IVP-RNN hybrid with purely parallel IVP solves: each observation is propagated independently to a shared latent z0 via an IVP solver, and the same invertible solver is reused in the decoder by solving forward in time, sharing parameters between encoder and decoder.
Key results
- IVP-VAE-Flow is about 42x faster than Latent-Flow in T-forward on MIMIC-IV forecasting and more than 28x faster in T-epoch, with at least one order of magnitude lower total training time.
- On PhysioNet 2012, IVP-VAE-Flow achieves the best forecasting MSE (0.567) and AUROC (77.1) among compared models.
- IVP-VAE-Flow reaches AUROC 0.805 and AUPRC 0.427 on MIMIC-IV mortality prediction, above Latent-Flow (0.786 / 0.404).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · MIMIC-IV forecasting | MSE: 0.755 -> 0.727 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · MIMIC-IV mortality classification | AUROC: 0.786 -> 0.805 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · PhysioNet 2012 forecasting | MSE: 0.584 -> 0.567 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · eICU forecasting | MSE: 0.594 -> 0.581 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · MIMIC-IV mortality classification | AUPRC: 0.404 -> 0.427 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · PhysioNet 2012 mortality classification | AUPRC: 0.327 -> 0.362 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · eICU mortality classification | AUROC: 0.781 -> 0.786 | Helped |
| Replace Latent-Flow with IVP-VAE-Flow | Latent-Flow · eICU mortality classification | AUPRC: 0.482 -> 0.472 | Hurt |
| Process time points independently by solving IVPs in parallel instead of using recurrent sequential computation | Latent-based continuous-time models · MIMIC-IV forecasting | T-forward: 1.063 -> 0.025 | Helped |
| Share one invertible IVP solver between encoder and decoder | Latent-based continuous-time models · MIMIC-IV classification | # Epochs: 57.8 -> 10.8 | Helped |
| Train with limited samples using IVP-VAE | Baseline models · MIMIC-IV small-data forecasting | MSE | Helped |
Surprising findings
- Imputation-based regular-series baselines (GRU, Informer, DLinear) underperform IVP-VAE on raw irregular data, which the authors attribute to imputation burying informative observations.
Limitations
- Evaluated mainly on EHR datasets and two tasks (forecasting and mortality classification); other tasks like imputation and regression left to future work.
- Mixing coefficients and loss weights were chosen empirically, and the per-observation KL averaging is a heuristic.
- Some MSE gains are within standard deviations; on several datasets (e.g., MIMIC-IV AUROC) GRU-Δt or GRU-D are equal or better.
Experimental setup
Datasets: MIMIC-IV PhysioNet 2012 eICU Climate Activity synthetic sinusoid dataset
Benchmarks: MIMIC-IV PhysioNet 2012 eICU Climate Activity
Metrics: MSE AUROC AUPRC T-forward T-epoch number of epochs number of parameters accuracy
Baselines: GRU-Δt GRU-D mTAN Raindrop GRU-ODE-Bayes CRU Latent-ODE Latent-Flow
Builds on: Latent-ODE Neural Flow Neural ODE variational autoencoder GRU-D
Closest papers by meaning
- XTSFormer: Cross-Temporal-Scale Transformer for Irregular-Time Event Prediction in Clinical Applications (AAAI 2025 · similarity 0.83)
- MIRA: Medical Time Series Foundation Model for Real-World Health Data (NeurIPS 2025 · similarity 0.82)
- Stabilized Neural Prediction of Potential Outcomes in Continuous Time (ICLR 2025 · similarity 0.81)
- One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models (NeurIPS 2026 · similarity 0.81)
- Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs (ICLR 2025 · similarity 0.81)
- Chimera: Effectively Modeling Multivariate Time Series with 2-Dimensional State Space Models (NeurIPS 2024 · similarity 0.81)
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.