From Signals to Causes: A Causal Signal Processing Framework for Robust and Interpretable Clinical Risk Prediction

Based on the paper "From Signals to Causes: A Causal Signal Processing Framework for Robust and Interpretable Clinical Risk Prediction" by Surajit Das and Maxine Tan, here is a detailed technical summary.

1. Problem Statement

Current learning-based signal processing systems for medical decision-making (e.g., predicting disease risk from CT scans, ECGs, or clinical records) rely heavily on statistical correlations. While these models achieve high predictive accuracy on historical datasets, they suffer from three critical limitations in real-world deployment:

• Lack of Robustness: Models often exploit spurious correlations induced by acquisition settings (e.g., scanner manufacturer, imaging protocols) or institutional practices. When these conditions change (distribution shift), model performance degrades significantly.
• Poor Interpretability: Standard interpretability tools (e.g., saliency maps) identify which input features influence predictions but fail to explain why those features are clinically relevant. They cannot distinguish between disease-causing factors and acquisition artifacts.
• Misalignment with Clinical Reasoning: Predictions cannot be reliably assessed under hypothetical interventions (counterfactuals), such as "What would the risk be if the patient stopped smoking?" or "How does a change in screening policy affect the outcome?"

The core problem is that current approaches treat biomedical signals as isolated predictive inputs rather than as effects of latent generative mechanisms, conflating biological variability with measurement noise.

2. Methodology: A Causal Signal Processing Framework

The authors propose a paradigm shift from correlation-based learning to Causal Signal Processing, integrating causal inference, representation learning, and neuro-symbolic reasoning.

A. Causal View of Medical Signals

The framework redefines clinical risk prediction as a latent state estimation problem rather than a direct mapping problem.

• Generative Model: Biomedical signals ($X$) are modeled as mixtures of:
  • Latent Disease States ($D$): Hidden causal variables (e.g., true pathology, progression stage).
  • Causal Risk Factors ($R$): Patient attributes (e.g., genetics, smoking history).
  • Acquisition Confounders ($A$): Non-causal variables (e.g., scanner type, protocol) that affect signal appearance but not the disease.
• Structural Causal Model (SCM): The process is formalized as:
  1. $D = f_D(R, U_D)$ (Disease is caused by risk factors).
  2. $X = f_X(D, A, U_X)$ (Signals are generated by disease and transformed by acquisition).
  3. $Y = f_Y(D, U_Y)$ (Clinical outcomes depend on disease).
• Goal: Learn representations that are invariant to interventions on non-causal variables ($A$) while remaining sensitive to changes in causal variables ($D, R$).

B. Neuro-Symbolic Integration

To bridge the gap between high-dimensional signal processing and clinical logic, the authors propose a Neuro-Symbolic pipeline:

1. Statistical Representation Learning: Deep neural networks extract features from raw multimodal signals (images, time series) to estimate uncertain evidence.
2. Causal Abstraction: These features are mapped to interpretable symbolic variables (e.g., "High Risk," "Nodule Burden") using domain knowledge.
3. Symbolic Reasoning: A logic-based layer encodes clinical guidelines (e.g., ACS screening rules) and causal constraints. It performs counterfactual reasoning and ensures decisions align with medical protocols.

This separation allows the system to handle the uncertainty of raw data (neural) while enforcing the rigidity of clinical rules (symbolic).

3. Key Contributions

• Conceptual Unification: The paper provides a unifying framework that treats clinical risk prediction as a signal processing problem involving source separation (disease vs. noise) and channel modeling (acquisition effects).
• Invariant Representation Learning: It defines robustness mathematically as conditional independence: $Z \perp\!\!\perp A | D$. A robust representation $Z$ must depend only on the disease state $D$, not the acquisition channel $A$.
• Neuro-Symbolic Architecture: It operationalizes causal interpretability by combining neural feature extraction with symbolic rule engines, enabling the system to output explanations grounded in clinical guidelines (e.g., "Recommended annual mammography due to age 48 and family history").
• Theoretical Guarantees: The authors present a theorem (Theorem 1) demonstrating that under specific distributional assumptions, predictive invariance across environments (e.g., different hospitals/scanners) implies that the learned representation is identifiable as a function of the latent disease state, up to an invertible transformation.
• Comparative Analysis: The paper systematically contrasts Correlation-based, Causal, and Neuro-Symbolic approaches, highlighting that only the latter supports counterfactual reasoning and maintains stability under intervention.

4. Results and Illustrative Examples

While the paper is primarily a conceptual and methodological proposal rather than an empirical benchmarking study, it validates the framework through:

• Theoretical Proofs: Demonstrating that enforcing invariance across environments allows for the identification of disease-related representations without explicit causal graphs.
• Use Cases:
  • Lung Cancer (CT Scans): A causal model conditions on smoking history and genetic predisposition (causes) rather than scanner artifacts. It remains stable when scanner types change, whereas correlational models fail.
  • Breast Cancer Screening: A neuro-symbolic system encodes American Cancer Society guidelines. If guidelines change (e.g., age thresholds), only the symbolic layer requires updating, not the neural feature extractors.
  • Cardiovascular (ECG) & Diabetic Retinopathy: The framework disentangles true physiological biomarkers (e.g., ST-segment morphology) from device-specific artifacts (e.g., electrode placement).

5. Significance and Impact

• Clinical Trust and Accountability: By providing explanations based on causal mechanisms and clinical guidelines rather than pixel-level correlations, the framework aligns AI decisions with human clinical reasoning, fostering trust.
• Regulatory Compliance: The explicit encoding of guidelines and causal assumptions makes the decision process auditable, addressing a major hurdle for FDA/EMA approval of medical AI.
• Robustness to Distribution Shift: The framework offers a principled solution to the "domain shift" problem, ensuring models generalize across different hospitals, scanners, and patient demographics without retraining.
• Future Research Direction: It calls for a shift in signal processing research toward causal representation learning, intervention-aware evaluation protocols, and uncertainty quantification in causal structures.

In conclusion, the paper argues that moving from "signals to causes" via a neuro-symbolic causal framework is essential for building medical AI that is not only accurate but also robust, interpretable, and safe for high-stakes clinical deployment.