Research map / Time series / Causal and generative time series
Physiome-ODE: A Benchmark for Irregularly Sampled Multivariate Time-Series Forecasting Based on Biological ODEs
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.
Builds Physiome-ODE, a benchmark of 50 irregularly sampled multivariate time-series forecasting datasets generated from biological ODE models in the Physiome Model Repository, selected by a new joint gradient deviation score, and evaluates five IMTS forecasters on it.
Problem
Existing irregularly sampled multivariate time series (IMTS) forecasting evaluation relies on about four datasets (PhysioNet2012, MIMIC-III, MIMIC-IV, USHCN), which may be biased; a constant-prediction baseline is competitive with neural ODE models on them.
Approach
Generate IMTS datasets by sampling parameters, initial states and durations of biological ODE models from the Physiome repository, select the most challenging ones by maximizing a new Joint Gradient Deviation (JGD) difficulty score, and drop irregular observations to create IMTS instances.
Key results
- LinODEnet is the best-performing model on Physiome-ODE, winning 25 of 50 datasets with average rank 2.10, ahead of GraFITi (2.20).
- On the traditional datasets, the constant GraFITi-C baseline beats neural ODE models GRU-ODE, Neural Flows on PhysioNet-2012 and MIMIC-IV.
- JGD correlates with the best test MSE across datasets with Spearman coefficient 0.77.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use constant forecasts instead of time-varying GraFITi forecasts | GraFITi on USHCN · USHCN | test MSE: 0.636±0.161 -> 0.875±0.204 | Hurt |
| Use constant forecasts instead of time-varying GraFITi forecasts | GraFITi on PhysioNet-2012 · PhysioNet-2012 | test MSE: 0.401±0.001 -> 0.407±0.001 | Hurt |
| Use constant forecasts instead of time-varying GraFITi forecasts | GraFITi on MIMIC-III · MIMIC-III | test MSE: 0.491 ± 0.014 -> 0.543±0.024 | Hurt |
| Use constant forecasts instead of time-varying GraFITi forecasts | GraFITi on MIMIC-IV · MIMIC-IV | test MSE: 0.285±0.002 -> 0.324±0.002 | Hurt |
| Use constant forecasts rather than ODE-based models | IMTS forecasting on PhysioNet-2012 and MIMIC-IV · PhysioNet-2012 and MIMIC-IV | test MSE | Helped |
| Use constant forecasts rather than the best ODE-based model | IMTS forecasting on MIMIC-III · MIMIC-III | test MSE | No clear effect |
| Evaluate forecasting models on ODE-generated Physiome datasets rather than only traditional datasets | IMTS forecasting models · Physiome-ODE (50 datasets) | average rank: LinODEnet: 2.10; GraFITi: 2.20; GraFITi-C: 2.86 | Mixed |
| Use LinODEnet rather than other evaluated models | IMTS forecasting on Physiome-ODE · Physiome-ODE (50 datasets) | number of dataset wins: 25 | Helped |
| Use constant forecasts rather than Neural Flows or GRU-ODE-Bayes | IMTS forecasting on Physiome-ODE · Physiome-ODE (50 datasets) | average rank: 2.86 vs 4.82 and 5.9 | Helped |
| Use constant forecasts rather than LinODEnet | IMTS forecasting on DUP01 · DUP01 | MSE: 0.964±0.036 -> 0.951±0.036 | Helped |
| Use constant forecasts rather than time-varying GraFITi forecasts | IMTS forecasting on INA01 · INA01 | MSE: 1.004±0.010 vs 1.004±0.009 | No clear effect |
| Use constant forecasts rather than LinODEnet | IMTS forecasting on LI01 · LI01 | MSE: 0.084±0.009 -> 0.742±0.010 | Hurt |
Surprising findings
- A constant-prediction baseline (GraFITi-C) outperforms several ODE-based models on the established IMTS datasets.
- On the new benchmark, GraFITi-C ranks above Neural Flows and GRU-ODE-Bayes (average rank 2.86).
Limitations
- Most datasets still leave large room for improvement versus the Bayes error of 0.0025 MSE.
- Datasets are semi-synthetic from ODE models; no original experimental observation data is available for the source models.
- Only 2000 instances per dataset, which is deliberately small.
Experimental setup
Datasets: PhysioNet-2012 MIMIC-III MIMIC-IV USHCN Physiome ODE models (50 selected)
Benchmarks: Physiome-ODE USHCN PhysioNet-2012 MIMIC-III MIMIC-IV
Metrics: MSE average rank number of wins Spearman correlation with JGD
Baselines: GRU-ODE-Bayes LinODEnet CRU Neural Flows GraFITi GraFITi-C
Builds on: neural ODE GraFITi Physiome Model Repository
Closest papers by meaning
- Time-IMM: A Dataset and Benchmark for Irregular Multimodal Multivariate Time Series (NeurIPS 2025 · similarity 0.88)
- Stable Neural Stochastic Differential Equations in Analyzing Irregular Time Series Data (ICLR 2024 · similarity 0.87)
- Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach (ICML 2024 · similarity 0.87)
- ABC: Any-Subset Autoregression via Non-Markovian Diffusion Bridges in Continuous Time and Space (NeurIPS 2026 · similarity 0.87)
- Elucidated Rolling Diffusion Models for Probabilistic Forecasting of Complex Dynamics (NeurIPS 2025 · similarity 0.87)
- Auto-Regressive Moving Diffusion Models for Time Series Forecasting (AAAI 2025 · similarity 0.87)
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.