Do-Calculus Interventions in Predictive Coding Networks: Formal Correspondence, Empirical Validation, and Architectural Prescriptions for Clinical Causal AI

Technical Summary: Do-Calculus Interventions in Predictive Coding Networks

Problem Statement
Clinical decision support systems increasingly require AI capable of reasoning about the consequences of actions (interventions) rather than merely predicting outcomes from historical observational patterns. Standard predictive models trained on Electronic Health Records (EHR) often fail under prospective deployment when treatment policies differ from training data, creating a distributional shift between the observational conditional $P(Y|X)$ and the interventional distribution $P(Y|do(T=t))$. While Pearl's do-calculus provides a formal framework for such reasoning, most causal machine learning methods treat interventional reasoning as a post-hoc estimation problem layered onto architectures lacking internal mechanisms for representing interventions. It remains untested whether specific neural architectures, such as Predictive Coding (PC) networks, can natively support interventional reasoning through their internal dynamics.

Methodology
The paper proposes and validates a formal correspondence between the dynamics of Predictive Coding networks and Pearl's do-calculus.

1. Theoretical Framework: The authors derive Proposition 1, which posits that in a linear-Gaussian Structural Causal Model (SCM), "hard clamping" a PC node (fixing its state to a value $c$) combined with the suppression of the upstream bottom-up prediction error is informationally equivalent to the $do(T=c)$ operator. This suppression severs the information pathway from the clamped node to its ancestors, mimicking the graph surgery of the do-operator.
2. Architectures:
   • PCNetworkV2: Implements Proposition 1 using a shared latent representation $z$ that decodes to covariates and outcomes. Interventional prediction involves clamping the treatment feature and zeroing the corresponding reconstruction error column before computing latent gradients.
   • PCNetworkV3: A modification where the outcome decoder receives direct access to the treatment variable alongside the latent representation, designed to handle clinical graphs where treatment has a direct edge to the outcome.
   • Baseline: A standard Multi-Layer Perceptron (MLP) trained with MSE loss serves as the associational baseline.
3. Benchmarks:
   • Toy SCM: A three-node linear-Gaussian SCM (Diagnosis $\to$ Symptom $\to$ Treatment) to validate the exact mathematical correspondence.
   • IHDP (Infant Health and Development Program): A semi-synthetic benchmark ($n=747$) with real covariate structure and reconstructed counterfactual outcomes to measure the "interventional gap" (difference between observational and interventional Mean Absolute Error).
   • Synthetic SCMs: Used for ablation studies across varying topologies (Chain, Mediator, Collider) and confounding strengths.
   • MIMIC-IV: A small-sample ($n=140$) semi-synthetic benchmark using real ICU covariate structure to test Average Treatment Effect (ATE) recovery.

Key Contributions
The paper makes three primary contributions:

1. Theoretical: Derivation and validation of Proposition 1, establishing that PC clamping with error suppression exactly implements the do-operator in linear-Gaussian SCMs.
2. Empirical: Demonstration that PCNetworkV2 achieves a statistically significant reduction in the interventional performance gap compared to an MLP baseline on the IHDP dataset (gap reduction = 0.237 MAE units, $p < 0.0001$). This advantage generalizes across Chain, Mediator, and Collider topologies.
3. Prescriptive: Identification of latent dimension as the primary architectural determinant of interventional advantage. Ablation studies show that smaller latent dimensions preserve causal inductive bias by preventing the model from memorizing spurious observational associations. Additionally, PCNetworkV3 is shown to recover the true ATE within 8.3% error on MIMIC-IV data by explicitly routing treatment to the outcome decoder.

Results

• Proposition 1 Validation: On the toy SCM, clamping $S=1$ with error suppression correctly reproduced the prior distribution $P(D|do(S=1)) = N(0,1)$, whereas standard conditioning yielded an incorrect posterior ($E[D|S=1] \approx 0.640$).
• Interventional Gap: The MLP baseline exhibited a large interventional gap (3.215 MAE units) on IHDP. PCNetworkV2 reduced this gap by 0.237 units. While PCNetworkV2 had higher observational MAE than the MLP (0.461 vs. 0.106), its architectural causal structure provided a measurable benefit in interventional settings.
• Ablation Findings:
  • Latent Dimension: The interventional advantage is non-monotonic with latent size; smaller dimensions ($d_z=4$) outperformed larger ones ($d_z=32$) because they forced the model to encode a compressed causal representation rather than spurious correlations.
  • Topology: The advantage was most pronounced in Collider structures, where the causal clamping mechanism operates with minimal interference from confounded pathways.
  • Training Fraction: A small fraction of interventional training data (5-10%) improved performance, but higher fractions (25%) disrupted latent optimization.
• Clinical Application (MIMIC-IV): PCNetworkV3 recovered the true ATE (1.0) with 8.3% error, whereas PCNetworkV2 failed (100% error) due to intervention degeneracy without the explicit treatment-to-outcome path. The MLP achieved near-perfect ATE estimation (1.5% error) via associational extrapolation, highlighting that ATE estimation and interventional MAE are distinct metrics.

Significance and Claims
The paper claims that the biological mechanism of predictive coding—specifically the separation of top-down predictions and bottom-up error updates—provides a natural substrate for causal reasoning. The key finding is that architectural causal structure (clamping + error suppression) confers a genuine, measurable interventional advantage over standard associational models, even without explicit interventional training data.

The authors modestly conclude that while predictive coding networks are not definitively established as the only or best architecture for clinical causal AI, the biological mechanism they are built around has an "unplanned use" as a causal reasoning engine. The work suggests that for interventional generalization, latent capacity should be treated as a "causal hyperparameter" (often favoring smaller dimensions) rather than a capacity knob to be maximized. The findings are currently limited to linear-Gaussian SCMs theoretically and modest-sized, tabular clinical datasets empirically, with future work needed to validate these claims in nonlinear, high-confounding, and larger-scale settings.