Technical Summary: From Static Risk to Dynamic Disease Trajectories

1. Problem Statement

Clinical decision-making operates as a feedback loop where risk estimates influence treatment, treatment alters disease trajectories, and both shape subsequent observations. Traditional clinical prediction models often fail in this setting because they rely on static risk scores or snapshot biomarkers that conflate disease biology with clinician behavior. Specifically, standard models trained on observational Electronic Health Record (EHR) data struggle with:

• Treatment-Confounder Feedback: Time-varying confounders that are affected by past treatments and influence future treatments and outcomes.
• Irregular and Informative Observation: Clinical data is often sampled irregularly, where the timing of measurements itself carries information about patient severity (informative observation).
• Static vs. Dynamic Claims: Models optimized for factual prediction (what happened) are often incorrectly deployed for counterfactual reasoning (what would happen under a different treatment), leading to biased treatment recommendations.

The paper argues that the field must shift from "benchmark prediction" to "decision-grade clinical evidence" by explicitly modeling disease trajectories as dynamic, intervention-aware processes.

2. Methodological Framework

The review organizes the field around a unified framework that bridges forecasting, counterfactual trajectory estimation, and policy evaluation. This framework is built upon three data-generating mechanisms and three decision tasks:

A. Data-Generating Mechanisms

1. Disease Evolution: The latent progression of the patient's state.
2. Treatment Assignment: The policy or rule determining interventions, which may be deterministic or stochastic.
3. Observation Process: The mechanism determining when and what data is recorded, which may be informative (dependent on patient state).

B. Decision Tasks

1. Factual Trajectory Forecasting: Predicting the future under the observed care regimen.
2. Counterfactual Trajectory Estimation: Estimating the trajectory under alternative treatment sequences or regimes.
3. Policy Evaluation: Assessing the value of a specific treatment policy (rule) rather than a single sequence.

C. Taxonomy of Approaches

The paper categorizes existing methods into four strata based on their causal semantics and time representation:

• S1 (Conceptual): Frameworks defining the problem.
• S2 (Non-Causal): Deep sequence models (e.g., RNNs, Transformers) that forecast observed trajectories without explicit counterfactual semantics. These are useful for prognosis but insufficient for intervention claims.
• S3 (Discrete-Time Causal): Methods handling time-varying confounding in discrete steps (e.g., Recurrent Marginal Structural Networks, Causal Transformers). These utilize Inverse Probability Weighting (IPW), balancing objectives, or g-computation.
• S4 (Continuous-Time Causal): Methods addressing irregular sampling and event-driven interventions using Neural Ordinary/Controlled Differential Equations (Neural ODEs/CDEs) and continuous-time weighting (e.g., TESAR-CDE).

D. Key Technical Components

• Causal Identification: The paper emphasizes that valid counterfactual estimation requires Sequential Exchangeability (no unmeasured confounding given history), Positivity/Overlap (every treatment has a non-zero probability), and Consistency.
• Handling Confounding:
  • Weighting: Marginal Structural Models (MSM) and IPW reweight observed data to emulate a randomized trial.
  • Balancing: Representation learning (e.g., Causal Transformer) aims to learn latent states that predict outcomes but are invariant to treatment assignment.
  • Latent Variables: Methods like Time Series Deconfounder or CDVAE attempt to infer hidden confounders.
• Irregular Time: Moving from fixed bins to continuous-time formulations (Neural CDEs) to avoid preprocessing artifacts and better model the observation process.
• Uncertainty Quantification: Emphasizing distributional counterfactuals (e.g., via generative models or Bayesian approaches) to quantify epistemic uncertainty and detect when a model is extrapolating beyond supported data.

3. Key Contributions

This review provides the first unified framework that:

1. Bridges Disciplines: Integrates multistate models, joint models, temporal point processes, deep sequence architectures, and longitudinal causal inference into a single coherent structure.
2. Explicitly Treats Bias: Systematically addresses treatment assignment, time-varying confounding, and observation-process bias as first-class components of the model, rather than nuisance variables.
3. Aligns Evaluation with Claims: Proposes a "decision-grade" evaluation standard that moves beyond predictive accuracy (AUROC) to include:
   • Overlap Diagnostics: Checking positivity and propensity distributions.
   • Target-Trial Alignment: Validating models against designs that emulate randomized trials.
   • Off-Policy Robustness: Stress-testing policies under distribution shift.
   • Uncertainty-Aware Abstention: Defining conditions under which models should refrain from making recommendations due to insufficient evidence.
4. Closes the Loop: Reframes trajectory modeling as an internal simulation problem for a closed-loop system: Prediction → Decision → Execution → Feedback Learning.

4. Results and Synthesis

As a review paper, the "results" are a synthesis of the state-of-the-art and a critical analysis of current limitations:

• Discrete-Time Limitations: While deep sequence models (RNNs, Transformers) have improved representation learning, empirical studies (e.g., ICML 2024) suggest that balancing objectives do not automatically guarantee improved counterfactual accuracy and can sometimes degrade performance if they remove predictive signal.
• Continuous-Time Necessity: Discretizing irregular clinical data introduces artifacts. Continuous-time methods (Neural ODEs/CDEs) are shown to be essential for unbiased counterfactual forecasting when observation times are informative.
• The "Non-Causal" Boundary: Many high-performing models in the literature are strictly prognostic (S2). The paper clarifies that these models cannot be safely used for treatment selection without explicit causal adjustments, as they conflate correlation with causation.
• Evaluation Gaps: The field currently lacks standardized benchmarks for continuous-time causal inference and robust protocols for off-policy evaluation in healthcare, leading to fragile policy claims.

5. Significance and Claims

The paper claims that the transition from static risk scoring to intervention-aware disease trajectory modeling is critical for the next generation of clinical AI. Its significance lies in:

• Reframing the Object of Study: Shifting the focus from a single risk score to a time-resolved, individualized trajectory that accounts for the coupled dynamics of disease, treatment, and observation.
• Enabling Safe Closed-Loop Learning: By explicitly modeling uncertainty and overlap, these frameworks enable "learning health systems" that can adapt policies while knowing when to abstain from action due to insufficient evidence.
• Preventing Semantic Overreach: The review serves as a guardrail against "claim inflation," ensuring that methodological novelty is matched by appropriate causal assumptions and validation rigor.
• Decision-Grade Evidence: It argues that progress should be measured not by architecture-level benchmarks, but by the ability to generate evidence that supports safe, personalized treatment decisions under real-world constraints.

The authors position this work as a necessary step toward AI-enabled longitudinal health intelligence that can anticipate disease evolution rather than merely describe it, ultimately supporting safer and more effective clinical decision-making.