PaperlayerView exact revision on arXiv

Source-linked AI summary

Admissible Diffusion for Multimodal Interventional Trajectories

Xing Han, Shravan Chaudhari, Jiarui Shao, Paul Pu Liang, Suchi Saria

arXiv:2609.34433v1cs.LG

TL;DR

Clinical trajectories under observed treatment do not establish what would happen under an alternative plan, because treatment and condition influence one another over time. ADMIT combines multimodal histories, treatment-conditioned latent diffusion, and explicit constraints on states or actions during rollout. Preliminary experiments report improved representation recovery and treatment-contrast error, and simulated dose-timing predictions capture most of the tumor-volume effect while exposing limits in effect size and baseline comparisons.

  • Problem

    Predictions under observed treatment need not describe outcomes under a different plan because treatment and patient condition feed back on one another.

    Source: §1 · p. 1

  • Method

    ADMIT uses multimodal histories and treatment-conditioned diffusion to generate intervention trajectories, applying physiological constraints to proposed actions or generated states during rollout.

    Source: §1 · p. 2, §3 · p. 4

  • Results

    Complementary observations improved supervised recovery of a simulated confounder and modestly improved treatment-contrast error, while leak-free dose-timing simulations captured most of the effect but showed exposure-input trade-offs.

    Source: §5 · p. 10

  • Takeaways & Limitations

    The findings motivate assessing representation recovery, intervention accuracy, and constraint enforcement separately when testing physiological priors in intervention trajectories.

    Source: §5 · p. 10

    Most evaluated encodings use the complete factual record before extracting a history prefix, so those runs are conditional-generation diagnostics rather than prospective forecasting evidence.

    Source: §3 · p. 4

Abstract

from arXiv · show

Generating a plausible clinical trajectory does not establish what would happen under a different treatment. We present ADMIT, a framework combining irregular multimodal representations, treatment-conditioned latent diffusion and explicit constraints on generated states or actions. We formulate its interventional target through sequential g-computation and distinguish causal assumptions from constraint satisfaction. Its admissibility mechanism translates physiological prior knowledge into explicit constraints on generated states and proposed actions. Treatment-exposure dynamics condition latent transitions, while state projection or action gating applies the constraints during rollout so that they influence subsequent trajectory generation. In our preliminary experiments, multimodal inputs improved supervised hidden-state recovery and reduced treatment-contrast error. In a simulated dosing-schedule experiment with leak-free history encoding, ADMIT predicted most of the tumor-volume change caused by redistributing a fixed total dose. An exposure input improved these predictions around a temporary dose reduction whether or not the assumed clearance rate was correct, but reduced the predicted size of a dose effect, and a deterministic recurrent baseline matched ADMIT's average predictions. Exposure projection reduced constraint violations, although enforcement remained incomplete. Semi-synthetic experiments using eICU context illustrated treatment-response generation under fixed and adaptive policies. Observational examples further characterize model treatment sensitivity. ADMIT provides a framework for testing whether complementary observations and physiological restrictions improve intervention trajectories, with representation recovery, effect accuracy and rule enforcement assessed separately.

1 Introduction

Clinical trajectories under observed treatment do not by themselves establish outcomes under an alternative plan, because treatment and patient condition feed back on one another. ADMIT combines multimodal histories, treatment-conditioned generation, and physiological constraints to simulate intervention trajectories, evaluated through synthetic and semi-synthetic tests and observational examples.

1 Introduction

  • Treatment and patient condition form a feedback loop, so predictions under observed treatment need not describe outcomes under a different plan.The target is a future-state distribution under specified decisions, conditional on the history available when comparison begins.Source: §1 · p. 1
  • Prior work leaves a gap in combining multimodal histories, alternative-treatment trajectory generation, and explicit physiological constraints during rollout.The cited multimodal models do not report treatment-dependent trajectory simulation, while event generators do not impose explicit physiological constraints on actions and transitions.Source: §1 · p. 2
  • ADMIT builds on sequential g-computation and conditional diffusion, applying physiological constraints to proposed actions and generated state transitions as trajectories unfold.The intended use is to compare possible consequences of care plans for the same patient and inspect consistency with specified rules.Source: §1 · p. 2
  • The evaluations test representation recovery, changed-dose predictions, and constraint enforcement synthetically, treatment-response generation semi-synthetically, and treatment sensitivity in observational examples.The experiments are designed to assess whether physiological priors improve counterfactual accuracy while reducing violations of specified constraints.Source: §1 · p. 2

2 Related work

Related work spans longitudinal causal trajectory models, conditional diffusion, and multimodal clinical representations, but these lines address different parts of ADMIT’s target. ADMIT brings them together for treatment-plan simulation with explicit physiological constraints during rollout.

2 Related work

  • Longitudinal causal models generate counterfactual trajectories, and G-Transformer explicitly updates covariates under a queried policy, but the cited demonstrations focus on structured variables.Balanced representations must still preserve information needed for adjustment.Source: §2 · p. 2
  • Prior diffusion work includes time-series imputation and longitudinal counterfactual outcomes, while ADMIT emphasizes longer rollouts and explicit constraint operations.The stated evaluation-scope difference does not imply a restriction on the other model’s generative architecture.Source: §2 · p. 2
  • ADMIT’s organization separates the intervention plan, learned transition, and imposed constraint so their roles can be inspected, including when a constraint changes the generated trajectory’s meaning.Figure 1 depicts aligned observations forming latent states, treatment-conditioned diffusion generating blocks, and optional constraints before history updates.Source: §2 · p. 3, Figure 1 · p. 3
  • Multimodal foundation models support clinical tasks such as image interpretation, question answering, and report generation; ADMIT instead targets treatment-plan trajectories with physiological constraints.The related work also includes cross-modal interaction modeling and continual adaptation across tasks and modality combinations.Source: §2 · p. 3

3 Method

ADMIT defines a history-conditional interventional target under causal identification assumptions, then learns latent transitions and rolls them out under fixed plans or adaptive policies. Optional exposure inputs and state or action constraints affect subsequent generation, but constraint satisfaction does not establish causal identification or clinical safety.

3 Method

  • Under consistency, sequential exchangeability, and positivity, the longitudinal g-formula identifies the joint interventional law conditional on available history.The target is a history-conditional interventional distribution; replacing measured history with a learned representation additionally requires adequate adjustment and predictive information and an accurately estimated transition law.Source: §3 · p. 3
  • The multimodal encoder aligns irregular observations with time attention and local kernels, fuses them through a top-two mixture of four experts, and produces a 16-dimensional latent state.Masked reconstruction, KL, and expert-balancing losses train the encoder; clinical masks represent retained rows after imputation, and optional exposure or comorbidity heads require simulator labels.Source: §3 · p. 4
  • The causal specification requires history summaries to use only observations available by the decision time, but most evaluated runs encode the complete factual record before extracting a prefix.The dosing-schedule experiment masks post-history observations and sets them to zero before encoding; other runs are conditional-generation diagnostics, not prospective forecasting evidence.Source: §3 · p. 4
  • The transition model generates residual blocks anchored at the final history state and trains a conditional denoiser; actions modulate denoiser features through learned scale and shift parameters.An optional free-running loss penalizes factual latent errors after generated states are fed back into history; disabling reverse-step noise still leaves random initialization and does not yield a conditional mean.Source: §3 · p. 4
  • Rollout repeatedly applies learned transitions to evolving histories under fixed plans or adaptive policies, with exposure features optionally conditioning future transitions.Adaptive treatment requires states and actions to alternate at decision times; a one-step transition or sequential factorization supports that ordering, while joint blocks conditioned on realized actions require stronger assignment assumptions.Source: §3 · p. 4
  • Admissibility rules constrain decoded states or actions during rollout, but finite projection iterations can leave violations and projection changes the original intervention distribution.Action gating instead suppresses doses above an exposure threshold and defines a modified policy; deterministic exposure must remain consistent with administered actions.Source: §3 · pp. 4–5
  • A toxicity threshold is an investigator-defined limit, and satisfying it neither establishes causal identification nor certifies clinical safety.Broader physiological training penalties remain an extension.Source: §3 · p. 5

4 Experiments

Across synthetic, semi-synthetic, and observational experiments, ADMIT is tested for representation recovery, intervention-effect prediction, and exposure-rule enforcement. Results show promise in controlled settings, while incomplete enforcement and information leakage limit what the clinical-context experiments establish.

4 Experiments

  • With hidden H, multimodal inputs reduced supervised recovery MSE by 74.2% and treatment-contrast RMSE by 14.2%.The representation improvement does not verify preservation of every required confounder.Source: §4 · p. 5

4.1 Predicting the effects of dose timing and drug clearance

  • Exposure input lowered break-difference RMSE from 0.081 ± 0.009 to 0.061–0.064 across tested retention values, but a GRU matched ADMIT’s average predictions.Tumor-volume RMSE improved from 0.62 without exposure to 0.48, while the GRU had lower tumor-volume error in every seed.Source: §4.1 · pp. 6, 8
  • With correct clearance, schedule-difference RMSE was 0.102 ± 0.012 for front-loaded dosing and 0.064 ± 0.007 for break dosing, 78% and 69% below the zero-effect reference.These results come from the tumor simulator and three training seeds.Source: §4.1 · p. 6, Table 2 · p. 8

4.2 Applying an exposure limit during generation

  • Projection reduced exposure-limit violations from 89.6% to 1.0% at persistence 0.9, but from 92.8% to 27.7% when persistence was zero.The single-run results measure compliance with a chosen bound; reduced violations alone establish neither correct treatment effects nor clinical safety.Source: §4.2 · pp. 8–9

4.3 Testing treatment responses with real clinical context

  • In semi-synthetic eICU context, treatment-effect RMSE was 0.488 for always treating and 0.357 for the adaptive plan, versus 0.804 and 0.621 for no-effect predictions.The evaluation used one training seed and a full-record encoder, so it does not establish prospective performance.Source: Table 3 · p. 9, §4.3 · p. 9

4.4 Examining treatment sensitivity in observed clinical outcomes

  • In observational eICU examples, predicted treatment differences were small relative to observed-outcome error, but cannot establish treatment effects or rank plans.Alternative outcomes were unavailable, and the encoder again used the full observed window.Source: §4.4 · pp. 9–10

4 Experiments

  • The experiments separately assess hidden-factor recovery, dose-timing effects, and rule violations; the clinical-context evaluations illustrate querying but leave valid forecasting unresolved.The dosing study is limited to three seeds and one test cohort, and does not test confounding adjustment or patient-adaptive plans.Source: §4.1 · p. 8, §4.4 · p. 10

5 Discussion

The discussion reports improvements in simulated confounder recovery and treatment-contrast error, and dose-timing predictions under leak-free history. Exposure input, constraint projection, and the scope of the findings qualify these results.

5 Discussion

  • Complementary observations improve simulated-confounder recovery, with a smaller improvement in treatment-contrast error.The discussion treats representation recovery, intervention accuracy, and rule enforcement as separate requirements.Source: §5 · p. 10
  • With leak-free history, ADMIT predicts most of the effect of changing dose timing in a simulated chemotherapy setting.Source: §5 · p. 10
  • Exposure input improves predictions around a temporary dose reduction and tolerates a wrong clearance rate, but shrinks the predicted dose-effect size; a deterministic recurrent model matches ADMIT’s average predictions.Source: §5 · p. 10
  • Projection reduces rule violations but changes the transition law; further work should assess generalization, richer constraints, uncertainty calibration, and clinical value using decision-time information.Source: §5 · p. 10

A Dosing-schedule experiment: settings, checks and ablations

The appendix reports dosing-schedule settings, validity checks, and ablations that support the Section 4.1 conclusions without changing them.

A Dosing-schedule experiment: settings, checks and ablations

A.1 Settings and checks

The settings and checks describe shared training conditions, leak-free history encoding, exposure agreement, dose-block coverage, and the metrics used to score predictions.

A.1 Settings and checks

  • All conditions share one Stage-1 checkpoint per seed, and transition settings differ only in the exposure input.Training uses factual trajectories only.Source: §A.1 · p. 13
  • Replacing post-history observations leaves the leak-free encoded history unchanged, whereas the full-record encoder changes substantially; both masking and zeroing were needed to prevent leakage.With λ = 0.78, model exposure matches simulator drug levels on factual doses and planned schedules.Source: §A.1 · p. 13, Table 6 · p. 13
  • Dose blocks depend on tumor states, and because the middle factor does not vary with l1, this experiment tests response to dose timing without testing adjustment for confounding.In training, 9% of four-step dose blocks fall and 6% rise by at least 0.3; these checks document coverage of evaluated dose levels and changes.Source: §A.1 · p. 14
  • Schedule-difference RMSE compares alternative-minus-steady tumor-volume differences with simulator differences; trajectory RMSE averages tumor-volume RMSE across three schedules.The dose-effect ratio compares predicted and simulated differences between sustained doses of 0.6 and 0.2.Source: §A.1 · p. 14

A.2 Ablations

The ablations show that exposure input helped ADMIT’s break predictions but not the GRU’s, while shared sampling noise and training length affected evaluation behavior.

A.2 Ablations

  • Paired-bootstrap intervals were narrow relative to between-seed differences, indicating that training seed drove most comparisons more than test-patient choice.Table 7 reports within-seed paired patient-bootstrap differences in schedule-difference RMSE.Source: §A.2 · p. 14, Table 7 · p. 14
  • For the GRU, removing exposure lowered schedule-difference errors in every seed: break RMSE was 0.061 ± 0.004 versus 0.071 ± 0.013, and front-loaded RMSE was 0.113 ± 0.015 versus 0.123 ± 0.020.Without exposure, its trajectory RMSE was 0.44 ± 0.09 and dose-effect ratio was 0.89 ± 0.03; exposure helped ADMIT’s break predictions but not the GRU’s.Source: §A.2 · p. 14
  • Sharing sampling noise across schedules reduced Monte Carlo error for ADMIT’s break difference from 0.119 to 0.005 with 64 samples; all reported results use shared noise.The earlier independent-noise evaluation had total RMSE of 0.139.Source: §A.2 · p. 14
  • On seed 0, training for 120 rather than 40 epochs raised front-loaded RMSE for all four ADMIT conditions, so the 40-epoch setting was retained.Break RMSE also rose for three of the four conditions.Source: §A.2 · p. 14

B.1 Hidden-confounder recovery

The appendix specifies a hidden-confounder recovery experiment and a semi-synthetic treatment-assignment setup. It details the simulated outcomes, intervention-arm handling, and forecast configuration.

B.1 Hidden-confounder recovery

  • Five irregular modalities, 10% missingness, and 40 steps define the hidden-confounder experiment; both models train a simulator-labeled comorbidity readout.The setup uses three seeds, with 1,024 training and 256 validation subjects per seed; context reconstruction is enabled only for the multimodal model.Source: §B.1 · p. 15

B.2 Semi-synthetic outcomes and treatment assignment

  • Untreated semi-synthetic outcomes combine a spline, a patient-specific Gaussian process, and a random function of 38 context features.The experiment also uses a fixed random-feature function in treatment assignment.Source: §B.2 · p. 15
  • For each intervention arm, treatment assignment is recomputed from that arm’s history, while real context and untreated paths remain fixed.Outcomes are standardized using factual training statistics.Source: §B.2 · p. 15
  • Forecasts cover hours 9–24 after eight history hours, but full-record encoding, backward filling, and cohort-wide imputation medians need correction before prospective evaluation.The adaptive policy uses prefix-decoded outcomes, while final grid decoding can revise earlier outputs using the generated suffix.Source: §B.2 · p. 15

C Rollout and exposure specification

Rollout generates and appends constrained state blocks before updating the history summary and selecting the next action. When an exposure model is specified, its carryover dynamics encode an assumed mechanism, not pharmacological validation or treatment support.

C Rollout and exposure specification

  • Each rollout block is optionally constrained, appended to the sequence, and used to update the history summary before the next action.Actions follow either a fixed plan or a policy evaluated on generated history, preserving At → Zt+1 alignment.Source: C Rollout and exposure specification · p. 15
  • When specified, exposure follows Ct+1 = λCt + At, with λ = exp(−∆/τ) determined by step length ∆ and time constant τ.Reported runs use guidance scale one; amplified guidance would alter the fitted distribution.Source: C Rollout and exposure specification · p. 15
  • The tumor simulator uses τ = 4 steps, retaining about 0.78 of exposure per step; this assumed mechanism does not establish pharmacological validity or treatment support.The semi-synthetic run omits exposure dynamics, and the tumor experiment also tests misspecified λ values.Source: C Rollout and exposure specification · p. 15
Loading 2609.34433v1…