Conformal prediction for censored life history data
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Conformal prediction for censored life history data".
Jane: Here's the gist of the paper. Conformal prediction for censored life history data develops an approach for individual prediction of general summaries of a multistate trajectory, such as time alive out of hospital or time spent in remission, which is important because right censoring complicates calibration due to partial observation The core methodology The approach involves defining a coverage determining time as the earliest time at which the observed history establishes whether the outcome lies in the region, even if the outcome itself is still unknown The estimator is weighted by inverse probability of remaining uncensored to that time and augmented with conditional coverage probabilities for unresolved histories This results in a doubly robust estimator that attains the semiparametric efficiency bound when both censoring and conditional coverage models are consistent at suitable rates Prediction targets and observed histories The paper considers general summaries of multistate outcomes, such as state-occupation times, which are functionals of a trajectory through a fixed clinical horizon supported by follow-up The principal targets are state-occupation times defined as Γ(τ) = Z τ 0 I(Y (t) ∈ A) dt The coverage determining time is defined as TΓ,q(τ) = inf for t in
0, τ: : KΓ(t; τ) ⊆ Cq(Z) or KΓ(t; τ) ∩ Cq(Z) = ∅ Calibration from censored histories To estimate coverage, the paper uses inverse weighting of observed coverage indicators weighted by the conditional censoring survival function G0(t Z) The augmented estimator is derived using a martingale term that incorporates partial histories to predict unresolved coverage, improving efficiency without changing the mean of the IPW contribution Theoretical guarantees The augmented estimator is shown to be doubly robust with respect to the censoring and conditional coverage models and attains the efficiency bound when both are consistent at suitable rates Theorem 3 provides a finite-sample coverage guarantee, showing that the prediction region satisfies pr
Γ∗τ ∈ Cb(Z∗) D: ≥ 1 − α − en(δ, c) − c−1 rCrQ Asymptotic coverage is achieved when both error terms vanish, leading to lim inf n→∞ pr
Γ∗τ ∈ Cb(Z∗): ≥ 1 − α Clinical applications The framework is applied to predict days alive and days alive out of hospital in the HF-ACTION trial, where the AIPW method was shown to be well calibrated for both joint coverage and home-time intervals across different censoring levels Similarly, relapse-free survival in node-positive breast cancer is analyzed using this methodology, demonstrating that the coverage determining time reduces variation and brings mean coverage closer to the nominal 90% target Discussion The developed conformal prediction approach for summaries of a multistate process yields a doubly robust estimator that can attain the efficiency bound for a fixed region’s coverage The numerical studies emphasize state occupation times, but the framework applies to general functionals of the outcome trajectory The procedure is robust to covariate shift if trial participants are weighted to reflect the clinic’s case mix The landmarking approach could be adapted for dynamic prediction at successive visits, while the landmarking approach could be adapted to calibrate this interval using patients alive and under observation at six months
Paper summary: Tom: So looking at the paper "Conformal prediction for censored life history data," the authors show a method that is doubly robust when predicting general summaries of multistate trajectories, like time spent in remission.
Jane: The core idea is defining a coverage determining time to handle right censoring and then using inverse weighting combined with conditional probabilities to estimate coverage even with partial histories.
Lu: It’s important because it shows that this approach can achieve the semiparametric efficiency bound for the prediction region's coverage when both the censoring and conditional coverage models are consistent at appropriate rates.
Meng: From an engineering standpoint, it means we have a framework for handling those complex time-series summaries in production settings where data is inevitably partial.
Lalam: The paper demonstrates that this method can give us a finite sample coverage guarantee, which is a concrete thing we can use to trust our predictions about the true coverage rate.
Tom: Overall, it’s about developing a robust way to predict these kinds of functional outcomes from complicated life histories without needing perfect information at every single step.
Jane: It moves beyond just predicting one state and focuses on summarizing the entire path a patient takes through their condition, which is really important for understanding long-term health trends.
Conclusion: Tom: So we're wrapping up this look at "Conformal prediction for censored life history data." It’s about building a way to make predictions when you have incomplete patient records, you know, those censored ones where we don't see the whole story.
Jane: Right. The title itself is pretty dense, but what it boils down to is using conformal prediction to predict big summaries of a patient's journey through their health—like how long they stay out of the hospital or how long they stay in remission.
Lu: The core idea is that this method creates a doubly robust estimator. That means it handles two types of uncertainty at once: the uncertainty from the censoring itself, and the uncertainty about what happens to those histories we don't see, which is called conditional coverage.
Meng: From an engineering standpoint, that sounds complicated. Does this actually run well on real-world data when you're trying to deploy a model for actual clinical prediction?
Lalam: I’ve looked at the results. The authors showed it attains the semiparametric efficiency bound if both their censoring model and their conditional coverage model are consistent at suitable rates. That means theoretically, they get as good a prediction as any other method could possibly get under those conditions.
Tom: So, the main implication is that we can trust these predictions more when we know how to model the missing pieces accurately. It moves us away from just guessing based on what’s observed so far.
Jane: Exactly. It suggests that for things like predicting time spent in a certain state, even with partial data, we have a mathematically sound way to set our confidence intervals without needing every single piece of information upfront.
Lu: And they showed it works on specific examples like days alive out of the hospital in one trial and relapse-free survival in cancer patients. It’s not just theoretical stuff; they tested it on real clinical scenarios, and it actually reduces the variation in those predictions compared to simpler methods.
Meng: But what about the practical cost? If you have to run this whole setup with these conditional coverage terms, does that mean a massive computational overhead when we're trying to scale up deployment?
Lalam: The paper focuses on the efficiency of the estimator itself, showing it performs well at scale if the underlying models are consistent. It’s about statistical power rather than just raw speed for inference.
Tom: So, the authors have essentially built a way to make smarter predictions from messy data by accounting for both what we see and what we don't see, and they proved that this method is statistically efficient under the right assumptions.
Jane: It’s about getting a more honest picture of a patient’s entire trajectory rather than just looking at the snapshot we have today. This opens up new ways to track long-term recovery or disease progression.
Lu: The real ambition here is that this landmarking approach could be adapted for predicting outcomes dynamically, meaning you can update those predictions as new data points come in during a patient's follow-up period.
Meng: That dynamic part is where I’d want to see more work before I’d sign off on production use. We need to know exactly how much extra compute that adds for every single prediction update.
University of Wisconsin–Madison
stat.ME, stat.ML
Submitted: 2026-11-11
Updated: 2026-11-11
Comments: 20 pages, 5 figures, 2 tables
Code: https://github.com/lmaowisc/lifeconformal
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 92/100
The gist: Conformal prediction for censored life history data develops an approach for individual prediction of general summaries of a multistate trajectory…
Key concepts
- Coverage Determining Time
- This is the earliest point in time where the observed history provides enough information to determine whether a specific outcome falls within a predefined prediction region, even if the final outcome is still unknown. It helps manage uncertainty arising from incomplete data.
- Doubly Robust Estimator
- The estimator used in this method is doubly robust because it remains accurate even if either the model for censoring or the model for conditional coverage is slightly inaccurate. This makes the prediction more reliable under real-world conditions where both assumptions might not be perfectly met.
- State-Occupation Times
- These are specific summaries of a patient's journey through different clinical states over time, such as how long they spend in a particular condition. The framework is designed to predict these complex functional outcomes based on the observed trajectory.
Terminology
Summary
Conformal prediction for censored life history data develops an approach for individual prediction of general summaries of a multistate trajectory, such as time alive out of hospital or time spent in remission, which is important because right censoring complicates calibration due to partial observation
The core methodology
The approach involves defining a coverage determining time as the earliest time at which the observed history establishes whether the outcome lies in the region, even if the outcome itself is still unknown The estimator is weighted by inverse probability of remaining uncensored to that time and augmented with conditional coverage probabilities for unresolved histories This results in a doubly robust estimator that attains the semiparametric efficiency bound when both censoring and conditional coverage models are consistent at suitable rates
Prediction targets and observed histories
The paper considers general summaries of multistate outcomes, such as state-occupation times, which are functionals of a trajectory through a fixed clinical horizon supported by follow-up The principal targets are state-occupation times defined as Γ(τ) = Z τ 0 I(Y (t) ∈ A) dt The coverage determining time is defined as TΓ,q(τ) = inf for t in [0, τ]: KΓ(t; τ) ⊆ Cq(Z) or KΓ(t; τ) ∩ Cq(Z) = ∅
Calibration from censored histories
To estimate coverage, the paper uses inverse weighting of observed coverage indicators weighted by the conditional censoring survival function G0(t Z) The augmented estimator is derived using a martingale term that incorporates partial histories to predict unresolved coverage, improving efficiency without changing the mean of the IPW contribution
Theoretical guarantees
The augmented estimator is shown to be doubly robust with respect to the censoring and conditional coverage models and attains the efficiency bound when both are consistent at suitable rates Theorem 3 provides a finite-sample coverage guarantee, showing that the prediction region satisfies pr[Γ∗τ ∈ Cb(Z∗) D] ≥ 1 − α − en(δ, c) − c−1 rCrQ Asymptotic coverage is achieved when both error terms vanish, leading to lim inf n→∞ pr[Γ∗τ ∈ Cb(Z∗)] ≥ 1 − α
Clinical applications
The framework is applied to predict days alive and days alive out of hospital in the HF-ACTION trial, where the AIPW method was shown to be well calibrated for both joint coverage and home-time intervals across different censoring levels Similarly, relapse-free survival in node-positive breast cancer is analyzed using this methodology, demonstrating that the coverage determining time reduces variation and brings mean coverage closer to the nominal 90% target
Discussion
The developed conformal prediction approach for summaries of a multistate process yields a doubly robust estimator that can attain the efficiency bound for a fixed region’s coverage The numerical studies emphasize state occupation times, but the framework applies to general functionals of the outcome trajectory The procedure is robust to covariate shift if trial participants are weighted to reflect the clinic’s case mix The landmarking approach could be adapted for dynamic prediction at successive visits, while the landmarking approach could be adapted to calibrate this interval using patients alive and under observation at six months
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Double robustness for coverage estimation allows for a doubly robust estimator that can attain the semiparametric efficiency bound when both censoring and conditional coverage models are consistent at suitable rates. This means the prediction method remains accurate even if either model—the censoring distribution or the conditional coverage probabilities—is slightly incorrect.
-
The augmented inverse probability weighted (AIPW) estimator is derived by augmenting the IPW estimator with predictions of coverage for unresolved histories, which improves efficiency without changing the mean of the IPW contribution. This augmentation uses
conditional coverage probabilities for unresolved histories.
-
The system can perform individual prediction of general summaries of a multistate trajectory, such as
time alive out of hospital or time spent in remission,
using a conformal prediction approach over a fixed clinical horizon supported by follow-up. -
Prediction intervals can be narrowed by using
coverage determining time as the earliest time at which the observed history establishes whether the outcome lies in the region, even if the outcome itself is still unknown.
This decision is weighted bythe inverse probability of remaining uncensored to that time.
-
The system can estimate component-specific and joint coverage guarantees, allowing for allocation of error levels to different parts of a multi-target prediction set. This allows one to determine
component-specific coverage even before joint coverage is resolved.
Sources
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes