📊 statistics

Multi-Dimensional Composite Endpoint Analysis via the Choquet Integral: Block Recurrent Encoding and Comparative Advantage Mapping

This paper introduces CWOT-CE, a novel Choquet integral-based statistical framework that integrates six heterogeneous outcome dimensions with block recurrent encoding to outperform conventional methods like Cox models and Win Ratio in power and interpretability across diverse cardiovascular trial scenarios.

Original authors: Ibrahim Halil Tanboga

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ibrahim Halil Tanboga

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 trying to judge the performance of a new medicine for heart failure. In the past, doctors and statisticians had a very narrow way of doing this. They would ask a simple question: "Did the patient die or have a major heart attack first?"

If the answer was "no," the patient was considered a success, even if they spent the next three years in and out of the hospital, feeling terrible, or if their blood tests showed their heart was getting worse. It was like grading a student's entire school year based only on whether they failed their final exam, ignoring all the homework, quizzes, and class participation they did before that.

This paper introduces a new, much smarter way to grade the "report card" of a clinical trial. The author, Dr. Ibrahim Halil Tanboga, calls this new method CWOT-CE.

Here is the breakdown of how it works, using simple analogies:

1. The Old Problem: The "First Event" Blindness

Traditional methods (like the Cox model) are like a security camera that stops recording the moment a burglar breaks a window.

The Flaw: If a patient has a minor heart attack, then spends the next year in the hospital, then recovers, the old method only cares about that first heart attack. It throws away all the data about the hospital stays and the recovery.
The Result: You miss 68% of the story. You might think a drug is working because no one died, even though the patients are suffering terribly.

2. The Competitor: The "Tie-Breaker" Method

There is a newer method called the Win Ratio. Imagine a boxing match where you compare two patients head-to-head.

How it works: You look at Patient A and Patient B. If A died and B didn't, A loses. If neither died, you check who had more heart attacks.
The Flaw: This method is great, but it gets stuck easily. If two patients have the exact same history (a "tie"), the method throws that pair away. In complex diseases, ties happen often (up to 27% of the time), meaning you are throwing away a huge chunk of your data.

3. The New Hero: CWOT-CE (The "All-Seeing Score")

Dr. Tanboga's new method, CWOT-CE, is like a personalized fitness tracker that doesn't just count steps; it calculates a "Health Score" based on six different things at once:

Did they survive?
How long were they free of heart attacks?
The "Burden" (The Secret Sauce): How many times did they go to the hospital, and when? (Going to the hospital 3 times in 6 months is worse than 3 times in 3 years).
When was their last bad event?
How did their blood markers look?
Are they alive at the end?

The "Block Recurrent Encoding" Analogy:
Instead of just counting "3 hospital visits," this method looks at the shape of the visits.

Patient A: Visits in Jan, Feb, March. (High burden, fast decline).
Patient B: Visits in Jan, then not again until next year. (Low burden, stable).
Old methods see "2 visits" for both. CWOT-CE sees that Patient A is much sicker and gives them a lower score.

4. The "Magic Math": The Choquet Integral

How do you combine these six different scores into one number? You can't just add them up like a grocery bill (because some things overlap).

The Metaphor: Imagine you are making a smoothie. If you add strawberries and raspberries, the flavor isn't just "strawberry + raspberry." They mix to create a new flavor.
The Choquet Integral is the blender. It understands that "Survival" and "Being Alive" are redundant (you can't have one without the other), so it doesn't double-count them. But it knows that "Hospital Burden" and "Last Event Time" work together to tell a stronger story, so it boosts their importance.

5. The Results: Why It Matters

The author ran a massive simulation (like a video game with 5,000 different worlds) to test this new method against the old ones.

The Verdict: CWOT-CE won in 15 out of 17 difficult scenarios.
The Big Win: In situations where patients were very similar (high correlation), the old "Win Ratio" method got stuck in ties and failed. CWOT-CE kept moving and found the truth.
The "Tie" Problem: The new method solved the problem of "ties" by turning complex histories into a single, smooth score, so it rarely gets stuck.

6. The "Why" (Shapley Decomposition)

One of the coolest features is that the method can explain why it gave a certain score.

The Metaphor: It's like a coach breaking down game tape. "We won because the defense was great, not because the offense scored."
The method can tell a doctor: "This drug worked primarily because it reduced hospital visits, even though it didn't stop deaths." This helps doctors understand exactly what the drug is doing.

Summary

The Old Way: "Did you die? No? Great, you're a winner." (Ignores suffering).
The Middle Way: "Let's compare you to your neighbor. If you're tied, we ignore you." (Loses data).
The New Way (CWOT-CE): "We are looking at your survival, your hospital visits, your blood work, and the timing of everything to give you a single, perfect 'Health Score'."

This new tool allows doctors to see the full picture of a patient's life, not just the first chapter, leading to better decisions about which heart medicines actually help people live better lives.

1. Problem Statement

Cardiovascular clinical trials increasingly rely on composite endpoints to capture the multi-dimensional nature of disease processes (e.g., mortality, nonfatal events, recurrent hospitalizations, biomarker changes). However, conventional statistical methods fail to fully utilize this complexity:

Cox Time-to-First-Event (TTFE): Discards all data occurring after the first event, ignoring recurrent burdens and post-event trajectories.
Negative Binomial/Andersen-Gill Models: Treat death as non-informative censoring, a biologically implausible assumption when death precludes further events.
Win Ratio (WR) / WRrec: Uses hierarchical pairwise comparisons but suffers from high tie rates (discarding 10–27% of patient pairs) and cannot easily incorporate continuous endpoints (e.g., biomarkers).
Joint Frailty Models: Face convergence instability in complex scenarios.

The core problem is the lack of a method that simultaneously integrates heterogeneous dimensions, handles recurrent events with temporal fidelity, accounts for death as a terminal event, and provides interpretable component attribution.

2. Methodology: CWOT-CE

The authors propose CWOT-CE (Choquet-Weighted Optimal Threshold for Composite Endpoints), a framework built on three pillars:

A. Block Recurrent Encoding (The Innovation)

Instead of simple event counts, the recurrent event process is encoded into two complementary scalar summaries to preserve temporal information:

AUC Burden: The area under the cumulative event count curve ( $\int N(t)dt$ ), capturing event intensity and timing.
Last Event Time: The time of the most recent event (or end of follow-up for event-free patients), capturing the trajectory endpoint.
This "block" encoding allows the method to distinguish between a patient with 3 events in 6 months versus 3 events over 3 years, a distinction lost in simple counts.

B. Non-Additive Aggregation via Choquet Integral

Patient profiles are aggregated using the Choquet integral with a 2-additive fuzzy measure.

Input: A vector of 6 normalized dimensions ( $K=6$ ): Survival, Event-free time, AUC Burden, Last Event, Biomarker, and Alive status.
Mechanism: Unlike a weighted sum, the Choquet integral allows for interaction terms (redundancy or synergy) between components. For example, it can model the redundancy between "Survival" and "Alive" status or the synergy between "Biomarker" and "Alive" status.
Default Weights: Pre-specified weights (e.g., Survival 0.25, AUC 0.18) and interaction coefficients are validated via Möbius parameterization.

C. Inference and Attribution

Inference: Uses a permutation test (exact finite-sample Type I error control) to compute the Choquet Benefit Index (CBI). Confidence intervals are derived via shift-pivot inversion, ensuring duality with the p-value.
Attribution: Employs Shapley-style decomposition to quantify the marginal contribution of each component to the overall treatment effect, allowing researchers to identify which specific outcomes drive the composite result.

3. Study Design

Framework: Pre-registered simulation study following the ADEMP (Aims, Data-generating mechanisms, Estimands, Methods, Performance measures) guidelines.
Comparators: CWOT-CE was compared against:
- Cox TTFE
- Win Ratio (WRrec)
- Wei-Lin-Weissfeld (WLW) marginal model
Scenarios: 20 clinically motivated scenarios (1,000–5,000 replications each) covering:
- Null hypotheses (Sharp, Trade-off, Estimand mismatch).
- Uniform effects (Statin, HFrEF).
- Discordant effects (Mortality paradox, soft-only effects).
- Correlation structures (Independent vs. High frailty).
- Robustness checks (Heavy censoring, informative censoring, delayed non-proportional hazards).

4. Key Results

A. Type I Error Control

Under the sharp null, all methods maintained nominal Type I error rates within the Bradley acceptance interval (2.5%–7.5%). CWOT-CE achieved 4.8% (MCSE 0.3%), comparable to Cox (5.0%) and WR (4.9%).

B. Power and Sensitivity

Across 17 non-null scenarios, CWOT-CE demonstrated superior or comparable performance:

vs. Cox TTFE: Outperformed in 15/17 scenarios (mean advantage +28.8 percentage points).
vs. WLW: Outperformed in 14/17 scenarios (mean advantage +27.2 pp).
vs. Win Ratio: Outperformed in 10/17 scenarios, tied in 5, and lost in only 2 (mean advantage +5.6 pp).

Specific Advantages:

High Correlation: In scenarios with strong shared frailty (COR-H), CWOT-CE outperformed WR by +35.4 pp. WR tie rates reached 27% in these scenarios, causing significant information loss, whereas CWOT-CE's continuous scoring avoided this.
Mortality-Driven Effects: CWOT-CE showed a +10.7 pp advantage in scenarios where only mortality was affected (DIS-DO).
Block Encoding Impact: Replacing the block encoding (AUC + Last Event) with simple count encoding reduced power by 22–51 pp in event-sensitive scenarios, proving the encoding strategy is the primary driver of performance.

C. Interpretability

Shapley decomposition successfully identified the dominant components in calibration scenarios (e.g., correctly attributing 50.1% of the effect to mortality in a mortality-driven scenario).

5. Significance and Contributions

Methodological Breakthrough: CWOT-CE is the first method to successfully integrate the Choquet integral (for non-additive aggregation) with block recurrent encoding (AUC + Last Event) for composite endpoints.
Solving the "Tie" Problem: By converting patient histories into continuous scalar scores, CWOT-CE eliminates the high tie rates that plague the Win Ratio, particularly in high-correlation settings common in advanced heart failure.
Temporal Fidelity: The block encoding recovers temporal trajectory information that simple event counts discard, bridging the gap between pairwise comparison methods and scalar scoring.
Clinical Utility: The method offers unique component attribution, allowing clinicians to understand which aspects of a treatment (e.g., reduced hospitalizations vs. improved survival) are driving the composite benefit.
Software Availability: The authors have released an R package (cwotce, v1.0.0) to facilitate adoption.

Conclusion

The paper concludes that CWOT-CE is broadly effective across clinically relevant scenarios. It outperforms traditional Cox models and the Win Ratio in most settings, particularly when outcomes are highly correlated or when the treatment effect is distributed across multiple dimensions. The study establishes that recurrent event encoding (how data is summarized) is a more critical determinant of statistical power than the aggregation method itself.