pith. sign in

arxiv: 2604.26070 · v2 · submitted 2026-04-28 · 💻 cs.LG · math.OC· math.ST· q-bio.QM· stat.TH

Observable Neural ODEs for Identifiable Causal Forecasting in Continuous Time

Jennifer Wendland , Nicolas Freitag , Maik Kschischo This is my paper

Pith reviewed 2026-05-14 21:42 UTC · model grok-4.3

classification 💻 cs.LG math.OCmath.STq-bio.QMstat.TH
keywords causal inferenceneural ODEscontinuous timeobservabilitylatent state-space modelsdynamic treatment effectshidden confounderscausal forecasting
0
0 comments X

The pith

Observability of latent dynamics from observed data is necessary to identify dynamic treatment effects in continuous-time models with hidden confounders.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that observability of the latent dynamics is necessary to identify causal effects of time-varying treatments when there are hidden confounders. This bridges control theory's concept of observability with causal inference by showing that reconstructible latent states allow derivation of potential outcomes. They derive a continuous-time adjustment formula based on the measurement model and filtering distributions. Observable Neural ODEs are introduced to enforce this property in neural models for forecasting. The approach matters for sequential decisions in domains like healthcare where interventions occur over time and complete observation is impossible.

Core claim

We show that, in latent state-space models with time-varying interventions, observability of the latent dynamics from observed data is necessary for identifying dynamic treatment effects, linking control-theoretic observability to causal identifiability, even when hidden confounders affect both treatments and outcomes. We derive a continuous-time adjustment formula expressing potential outcome distributions under treatment trajectories via the measurement model, latent dynamics, and the filtering distribution over latent states given observed histories. We propose Observable Neural ODEs (ObsNODEs), Neural ODE models in observable normal form for causal forecasting.

What carries the argument

Observable Neural ODEs in observable normal form, which ensure latent states are reconstructible from observations to enable causal identification via the adjustment formula.

If this is right

  • Dynamic treatment effects can be identified from observed data when the latent dynamics are observable.
  • Potential outcome distributions under alternative treatment trajectories become computable from the measurement model and filtering distribution.
  • ObsNODEs support forecasting of outcomes under different continuous-time treatment paths.
  • The models achieve strong performance on synthetic cancer data, MIMIC-IV semi-synthetic data, and real sepsis data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The observability requirement could guide architecture choices for other neural models in continuous-time causal tasks.
  • In practice one could test whether a fitted model satisfies state reconstructibility before trusting its causal forecasts.
  • The same linking of observability to identifiability may extend to discrete-time or hybrid intervention settings.
  • Integration with continuous-time reinforcement learning for policy learning under uncertainty is a direct follow-on possibility.

Load-bearing premise

The latent dynamics admit an observable normal form that can be reconstructed from the measurement model and observed histories.

What would settle it

A concrete counterexample in which the latent dynamics are not observable from the data yet dynamic treatment effects remain identifiable would refute the necessity of observability.

Figures

Figures reproduced from arXiv: 2604.26070 by Jennifer Wendland, Maik Kschischo, Nicolas Freitag.

Figure 1
Figure 1. Figure 1: (a) Temporal DAG for treatment A, outcome Y , and latent state Z at three consecutive time points 0, t and t + s. The process ϵt is a hidden confounder affecting both treatment and outcome. (b) In a dynamic treatment regime, treatment depends on treatment and outcome history, as indicated by green dashed arrows. 3.2.1 The conditional front-door criterion To connect causal identification with latent state-s… view at source ↗
Figure 2
Figure 2. Figure 2: Synthetic cancer dataset. RMSE heatmaps (mean over five runs) for standardized (Z-score) tumor volume and body weight forecasts at confounding strength γ = 4. The horizontal axis shows the assimilation time and the vertical axis the forecast horizon. Columns correspond to ObsNODE, doseAI (Wendland, Kschischo, 2025), IGC-Net (Hess et al., 2026) and SCIP-Net (Hess, Feuerriegel, 2025). RMSE values larger than… view at source ↗
Figure 3
Figure 3. Figure 3: Semi-synthetic MIMIC-IV sepsis dataset. RMSE for ObsNODE, OptAB, IGC-Net and SCIP-Net as a function of data assimilation time (observation time) and forecast horizon over a 24-hour horizon. Results are averaged over five runs and RMSE values larger than one were clipped at one. MODEL/ s = tf − tc 1 2 3 4 5 6 SEMI-SYNTHETIC OUTCOME OBSNODE 0.13±0.00 0.21±0.00 0.25±0.00 0.27±0.00 0.30±0.00 0.33±0.00 OPTAB 0.… view at source ↗
Figure 4
Figure 4. Figure 4: Antibiotic sepsis treatment. Forecast RMSE for SOFA score, creatinine, total bilirubin, and alanine aminotransferase (ALT) as a function of data assimilation time (x-axis) and forecast horizon (y-axis). 2022). Consequently, they cannot be used for individualized outcome forecasting. These approaches are typically grounded in structural causal models and identification results such as the longitudinal g-for… view at source ↗
read the original abstract

Causal inference in continuous-time sequential decision problems is challenged by hidden confounders. We show that, in latent state-space models with time-varying interventions, observability of the latent dynamics from observed data is necessary for identifying dynamic treatment effects, linking control-theoretic observability to causal identifiability, even when hidden confounders affect both treatments and outcomes. We derive a continuous-time adjustment formula expressing potential outcome distributions under treatment trajectories via the measurement model, latent dynamics, and the filtering distribution over latent states given observed histories. We propose Observable Neural ODEs (ObsNODEs), Neural ODE models in observable normal form for causal forecasting. ObsNODEs learn continuous-time dynamics with states reconstructible from observations, enabling outcome prediction under alternative treatment paths. Experiments on synthetic cancer data, semi-synthetic data based on MIMIC-IV, and real-world sepsis data show strong performance over recent sequence models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that in latent state-space models with time-varying interventions and hidden confounders, observability of the latent dynamics from observed data is necessary for identifying dynamic treatment effects. It links control-theoretic observability to causal identifiability, derives a continuous-time adjustment formula for potential outcome distributions expressed via the measurement model, latent dynamics, and filtering distribution p(z_t | history), proposes Observable Neural ODEs (ObsNODEs) in observable normal form for causal forecasting, and reports strong performance on synthetic cancer data, semi-synthetic MIMIC-IV data, and real sepsis data.

Significance. If the observability-to-identifiability link holds and the adjustment formula can be evaluated from data, the work would provide a principled bridge between control theory and causal inference for continuous-time sequential decisions, enabling reliable forecasting of outcomes under alternative treatment paths in settings with hidden confounding such as clinical decision support.

major comments (2)
  1. [Abstract] Abstract and derivation of the adjustment formula: the central necessity claim requires showing that the observable normal form is uniquely reconstructible from the measurement model and filtering distribution under time-varying confounded interventions; the presentation assumes this reconstruction step without an explicit uniqueness or stability result, which is load-bearing for the identifiability guarantee.
  2. [Method] § on ObsNODEs and continuous-time adjustment: the formula integrates over the reconstructed latent state to identify potential outcomes, but without verifying that the normal-form transformation remains recoverable when interventions affect both treatments and outcomes, the formula cannot be guaranteed to be evaluable from observed trajectories alone.
minor comments (2)
  1. Clarify how the observable normal form is enforced in the neural ODE parameterization and whether it introduces additional constraints on the learned dynamics.
  2. [Experiments] Experiments section: include sensitivity checks or ablations that violate the observability condition to test robustness of the claimed performance gains.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential bridge between control theory and causal inference. We address each major comment below and will incorporate the requested clarifications and results into the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract and derivation of the adjustment formula: the central necessity claim requires showing that the observable normal form is uniquely reconstructible from the measurement model and filtering distribution under time-varying confounded interventions; the presentation assumes this reconstruction step without an explicit uniqueness or stability result, which is load-bearing for the identifiability guarantee.

    Authors: We agree that an explicit uniqueness result is necessary to make the identifiability claim fully rigorous. The current manuscript invokes standard control-theoretic observability results but does not restate or adapt them to the time-varying confounded setting. In the revision we will add a new theorem establishing that the observable normal form is uniquely recoverable (up to a known transformation) from the measurement model and the filtering distribution p(z_t | history) under the paper's assumptions on the latent dynamics and interventions. The theorem will include a stability condition on the filter and a brief proof sketch based on the observability rank condition extended to the nonlinear case with exogenous inputs. revision: yes

  2. Referee: [Method] § on ObsNODEs and continuous-time adjustment: the formula integrates over the reconstructed latent state to identify potential outcomes, but without verifying that the normal-form transformation remains recoverable when interventions affect both treatments and outcomes, the formula cannot be guaranteed to be evaluable from observed trajectories alone.

    Authors: We acknowledge that the recoverability of the normal-form transformation must be shown explicitly when interventions influence both treatment assignment and outcomes. The present derivation assumes this recoverability follows from the model structure but does not verify it. In the revised manuscript we will add a proposition demonstrating that, under the stated latent-state dynamics and measurement model, the normal-form coordinates remain identifiable from observed trajectories even when the intervention process depends on the latent state. The argument will rely on the fact that the filtering distribution can be consistently estimated from data and that the intervention enters the dynamics as a known exogenous input. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation links external control-theoretic observability to causal identifiability

full rationale

The paper's core derivation introduces a continuous-time adjustment formula expressed via the measurement model, latent dynamics, and filtering distribution p(z_t | history). Observability is imported from control theory as an external concept required for identifiability under hidden confounders, rather than defined in terms of the paper's own outputs or fitted quantities. ObsNODEs are defined as Neural ODEs placed in observable normal form, but the identifiability result does not reduce to a self-fit or renaming of inputs; the reconstruction step is presented as a modeling choice validated on synthetic and real data. No self-citation chains, ansatz smuggling, or predictions that are statistically forced by construction appear in the provided claims. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Based on abstract only; limited visibility into parameters and assumptions.

free parameters (1)
  • neural network parameters
    Weights of the Neural ODE learned from data to approximate latent dynamics.
axioms (1)
  • domain assumption Existence of a latent state-space model with time-varying interventions
    Invoked to frame the causal inference problem in continuous time.
invented entities (1)
  • Observable Neural ODEs (ObsNODEs) no independent evidence
    purpose: Neural ODE model placed in observable normal form for causal forecasting
    New model class introduced to satisfy the observability condition

pith-pipeline@v0.9.0 · 5462 in / 1352 out tokens · 33619 ms · 2026-05-14T21:42:58.398585+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages

  1. [1]

    2. 2nd ed. 2024. 9 Bang Heejung, Robins James M.Doubly Robust Estimation in Missing Data and Causal Inference Models // Biometrics. XII 2005. 61, 4. 962–973. Bernard Pauline. Observer Design for Nonlinear Systems. 479. Cham: Springer International Publishing, 2019. (Lecture Notes in Control and Information Sciences). Bica Ioana, Alaa Ahmed M., Jordon Jame...

    work page arXiv 2024

  2. [2]

    (Dover books on engineering)

    Dover ed. (Dover books on engineering). Johnson Alistair E. W., Bulgarelli Lucas, Shen Lu, Gayles Alvin, Shammout Ayad, Horng Steven, Pollard Tom J., Hao Sicheng, Moody Benjamin, Gow Brian, Lehman Li-wei H., Celi Leo A., Mark Roger G.MIMIC-IV, a freely accessible electronic health record dataset // Scientific Data. I 2023. 10, 1. 1. Johnson Alistair E. W....

    work page 2023

  3. [3]

    1393–1512

    7, 9. 1393–1512. Robins James M.Optimal Structural Nested Models for Optimal Sequential Decisions // Proceedings of the Second Seattle Symposium in Biostatistics: Analysis of Correlated Data. New York, NY: Springer New York, 2004. 189–326. Rubanova Yulia, Chen Ricky T. Q., Duvenaud David. Latent ODEs for Irregularly-Sampled Time Series. 2019. Rubin Donald...

    work page 2004