📊 statistics

A flexible approach to sequential prediction under intervention

This paper proposes a flexible causal predictive framework for estimating risk under preventative interventions, introducing three specific models (Unexposed Mediator, Modifiable Risk Factor, and Two Component) to ensure causally consistent risk estimates across repeated visits while highlighting limitations related to initial data reliance and structural assumptions.

Original authors: Matthew Sperrin, Bowen Jiang, Joyce Huang, Niels Peek, Alexander Pate

Published 2026-03-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Matthew Sperrin, Bowen Jiang, Joyce Huang, Niels Peek, Alexander Pate

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to predict a patient's future health, specifically their risk of a heart attack in the next 10 years. You have a crystal ball (a Clinical Prediction Model) that looks at their current stats: blood pressure, weight, cholesterol, and whether they smoke.

The Problem: The "Static" Crystal Ball
Standard crystal balls are great at saying, "Based on today's numbers, here is your risk." But they fail when you ask, "What if I take this pill?" or "What if I start running?"

Why? Because standard models are like a photo taken at a single moment. They don't understand cause and effect.

The Confusion: If a patient takes blood pressure medication, their blood pressure drops. A standard model sees the low blood pressure and thinks, "Great, low risk!" But it also sees the medication and might get confused because, in real-world data, people who take meds often have higher risk to begin with (doctors prescribe them to the sickest people). The model gets tangled up, double-counting the effects, and gives a wrong answer.
The Time Travel Issue: Even worse, if you check the patient again in 6 months, the standard model gets even more confused. It sees the low blood pressure again and thinks, "Oh, they are still on meds, so I'll count the benefit again!" This leads to inconsistent predictions over time.

The Solution: A New Kind of Crystal Ball
The authors of this paper propose a new framework called "Prediction Under Intervention" (PUI). They built four different "versions" of this crystal ball to solve these problems, like upgrading from a basic calculator to a super-computer.

Here are the four approaches, explained with analogies:

1. The "Treatment Offset" Model (The Simple Fix)

The Idea: This is like taking a standard photo and manually painting over the medication part with a "magic brush" that says, "Assume this pill works exactly as science says it should."
How it works: You force the model to ignore the messy real-world data about who took the pill and instead use a fixed, proven number for how much the pill lowers risk.
The Flaw: It works okay for a one-time check. But if you check the patient again later, the model still gets confused by the changing numbers. It's like trying to drive a car with a map that doesn't update when you turn a corner.

2. The "Unexposed Mediator" Model (The Time-Traveler)

The Idea: This model solves the "Time Travel" problem by looking at the patient's past self instead of their current self.
The Analogy: Imagine a patient named Jane. She has high blood pressure (140). She starts taking meds, and her pressure drops to 120.
- Standard Model: Looks at 120 and says, "Low risk!"
- Unexposed Model: Says, "Wait, let's look at what her blood pressure was before she took the pill (140). We will use that number to calculate the risk, but we will add the known benefit of the pill separately."
Why it's better: It prevents "double-counting." It treats the medication as a separate "bonus" added to the baseline risk, rather than letting the medication change the input numbers. This keeps the prediction consistent whether you check her today or in 6 months.

3. The "Modifiable Risk Factor" Model (The Swiss Army Knife)

The Idea: This is for when you want to ask, "What if I stop smoking? What if I lose weight? What if I take a statin?" all at once.
The Analogy: Think of the patient's body as a complex machine with dials (Blood Pressure, Weight, Cholesterol).
- Instead of asking about specific tools (like "the blue wrench" or "the red hammer"), this model asks, "What happens if I turn the Weight Dial down by 5kg?"
- It assumes that how you turn the dial doesn't matter. Whether you turn it down by diet, exercise, or surgery, the effect on the machine is the same.
The Flaw: Because it forces the dials to follow strict causal rules, it sometimes loses a bit of its "predictive sharpness." It's very flexible but slightly less accurate at guessing the exact future risk.

4. The "Two-Component" Model (The Best of Both Worlds)

The Idea: This is the ultimate solution. It splits the prediction into two parts: The Baseline and The Change.
The Analogy: Imagine a video game.
- Part 1 (The Baseline): You record the player's stats at Level 1 (Visit 0). You use a super-accurate, flexible AI to predict their score based on exactly where they are right now. This part is purely for accuracy.
- Part 2 (The Intervention): When the player levels up or changes gear (takes meds, loses weight), you don't re-calculate the whole game. You just calculate the difference caused by the change.
Why it wins: It keeps the high accuracy of the standard model for the "starting point" but uses the strict causal rules to handle any future changes. It's like having a GPS that knows the exact traffic right now (Baseline) but also knows exactly how much time you'll save if you take the highway (Intervention).

The Big Takeaway

The authors tested these models on 4 million people in the UK. They found that:

Old models get confused when you try to predict the future after a patient starts treatment.
The new Two-Component Model is the winner. It allows doctors to say, "If you start this diet and take this pill, your risk drops from 18% to 6%," and it will give the same answer if you ask again next year, without getting mathematically confused.

In short: They built a smarter crystal ball that understands cause and effect, remembers the patient's starting point, and can handle any number of lifestyle changes without losing its mind. This helps doctors give better, more consistent advice to patients trying to prevent heart disease.

1. Problem Statement

Clinical Prediction Models (CPMs) are widely used to estimate individual risk (e.g., for cardiovascular disease) to guide interventions. However, standard CPMs suffer from a critical limitation: they are associative, not causal. They cannot accurately answer counterfactual questions such as, "What is my risk if I start taking statins?" or "What is my risk if I improve my diet?" because they often suffer from:

Confounding by indication: Treatment variables are correlated with unmeasured risk factors.
Over-adjustment: Adjusting for mediators (e.g., blood pressure) that are themselves affected by the intervention (e.g., antihypertensives), leading to biased effect estimates.
Inconsistency in Sequential Prediction: When a patient is re-evaluated over time after an intervention, standard models often "double-count" the intervention's effect (counting both the treatment variable and the resulting change in the mediator), leading to incoherent risk estimates across visits.
Rigidity: Existing methods often require pre-specifying specific interventions, lacking flexibility for multiple or arbitrary lifestyle/pharmaceutical interventions.

The authors aim to develop a framework for Prediction Under Intervention (PUI) that satisfies four key characteristics:

Causal Validity: Estimates must target appropriate causal estimands (e.g., Average Treatment Effects).
Sequential Concordance: Risk estimates must remain consistent over time for the same individual undergoing sequential evaluations.
Flexibility: The model must handle an arbitrary number of interventions without re-coding the model structure.
Predictive Accuracy: The model must maintain high discrimination and calibration, balancing causal constraints with predictive performance.

2. Methodology

The authors propose a family of four modeling approaches, progressing from simple to complex, to address the challenges above.

A. Treatment Offset Model (Baseline Causal Approach)

Mechanism: Fixes the coefficient of an intervention variable (e.g., antihypertensives) to a known causal value derived from external evidence (e.g., meta-analyses of RCTs).
Limitation: While it addresses causal validity at a single timepoint, it fails Sequential Concordance. If a patient starts treatment and their blood pressure (SBP) drops, the model sees both the treatment variable and the lower SBP, effectively counting the risk reduction twice.

B. Unexposed Mediator Model (Solving Sequential Inconsistency)

Mechanism: Instead of using the observed value of a mediator (e.g., current SBP) as a predictor, the model uses the unexposed value (e.g., SBP as if the patient had never taken the medication).
Logic: By holding the mediator constant at its pre-intervention level for the prediction calculation, the model isolates the intervention's effect. This prevents double-counting when the patient returns for follow-up with improved biomarkers.
Trade-off: Requires data on "unexposed" values (often requiring imputation or specific data collection) and still struggles with multiple, diverse interventions.

C. Modifiable Risk Factor (MFR) Model (Solving Flexibility)

Mechanism: Models interventions abstractly through their effects on modifiable risk factors (mediators) like SBP, BMI, and cholesterol.
Causal Structure: Uses a Directed Acyclic Graph (DAG) to define causal pathways. Coefficients for the mediators are fixed to their causal direct effects.
Consistency Assumption: Assumes that if two different interventions (e.g., diet vs. exercise) produce the same change in a mediator (e.g., 5kg weight loss), they have the same indirect effect on the outcome.
Limitation: Fixing coefficients to causal values often degrades Predictive Accuracy (calibration and discrimination) because the model is no longer optimized for likelihood maximization.

D. Two-Component Model (Solving Predictive Accuracy)

Mechanism: Decomposes the prediction into two distinct parts:
1. Baseline Component: A standard, flexible predictive model trained on the first visit data (using observed values) to establish a factual baseline risk. This ensures high discrimination.
2. Intervention Component: A separate module that calculates the change in risk based on the difference between the baseline values and the new values (or counterfactual values) of the mediators, using the causal DAG structure.
Formula: $Risk_{total} = Risk_{baseline} \times Risk_{intervention\_adjustment}$ .
Advantage: Retains the high predictive performance of the baseline model while allowing for causal adjustments for future interventions and sequential updates.

3. Key Contributions

Framework for Sequential PUI: The paper formally defines the problem of sequential prediction under intervention and proposes specific mathematical solutions to ensure risk estimates are concordant over time.
Unexposed Mediator Concept: Introduces the novel strategy of using pre-intervention mediator values for prediction to prevent double-counting of intervention effects in longitudinal settings.
Consistency of Indirect Effects: Proposes a practical assumption allowing models to estimate the impact of arbitrary interventions (even those without direct RCT data) by mapping them to changes in known modifiable risk factors.
Two-Component Architecture: Demonstrates a hybrid approach that successfully reconciles the tension between causal validity and predictive accuracy, a common failure point in previous PUI attempts.
Empirical Validation: Validates these methods on a massive dataset (4 million individuals from the UK Clinical Practice Research Datalink) for cardiovascular disease prevention.

4. Results

The models were applied to a cohort of 4 million UK primary care patients to predict 10-year cardiovascular risk.

Treatment Offset Model: Showed minor drops in performance compared to non-causal models but produced inconsistent risk estimates over time (e.g., risk estimates dropped artificially when a patient's SBP improved, failing to account for the double-counting issue).
Unexposed Mediator Model: Successfully resolved the sequential inconsistency issue. It maintained high discrimination (C-index ~0.880) and calibration (ICI 0.001) similar to the non-causal model.
Modifiable Risk Factor Model: Successfully handled multiple interventions (statins, diet, exercise) flexibly. However, it suffered from calibration issues (Integrated Calibration Index rose to 0.005) due to the rigid fixing of coefficients.
Two-Component Model: Achieved the best balance. It accommodated multiple interventions and maintained excellent calibration (ICI 0.001) and discrimination (C-index 0.871), effectively correcting the performance drop seen in the MFR model.

Case Study (Jane):

A 65-year-old woman with diabetes and high SBP.
The Non-Causal model incorrectly suggested antihypertensives increased risk (due to confounding).
The Treatment Offset model showed risk reduction but became inconsistent at follow-up.
The Two-Component model provided consistent, calibrated risk estimates for both the initial visit and follow-up, accurately reflecting the impact of hypothetical interventions (e.g., reducing SBP to 120 mmHg).

5. Significance and Limitations

Significance:
This work bridges the gap between causal inference and machine learning prediction. It provides a practical toolkit for clinicians to use prediction models not just for static risk assessment, but for dynamic decision support where patients are monitored over time and interventions are simulated. The Two-Component Model offers a pragmatic solution for real-world electronic health record (EHR) data where perfect causal data is rarely available.

Limitations:

Reliance on Causal Knowledge: The models require a known causal structure (DAG) and accurate causal effect estimates (e.g., from meta-analyses). If these inputs are wrong, the predictions will be biased.
Consistency Assumption: The assumption that "identical changes in mediators yield identical outcome changes" regardless of the intervention source is strong and may not always hold (e.g., weight loss via surgery vs. diet).
Arbitrary First Visit: The Two-Component Model anchors predictions to a specific "first visit." If this visit is arbitrary or far in the past, predictive performance may degrade for later visits.
Uncertainty Propagation: The current implementation uses point estimates for causal effects, ignoring the variance/bias in those estimates. A fully Bayesian extension is suggested as future work.

In conclusion, the paper offers a robust, flexible, and empirically validated framework for generating causal, sequential risk predictions, moving clinical decision support from static risk scores to dynamic intervention planning.