Much of clinical prediction runs on one simple question asked of historical records: did the event (death, readmission, recurrence) happen within a fixed window such as one year? Logistic regression answers it with odds ratios and bedside nomograms. In retrospective cohorts, though, many patients leave the record before the window closes. They transfer care, stop visiting, or reach the end of the database. Analysts then either count them as event-free or drop them. Both choices bias the estimates. They become worse when the patients who leave early are systematically sicker or healthier than those who stay, which is called informative censoring. Survival methods handle censoring, but they give hazard curves rather than the horizon odds ratios that clinicians use. The existing fixes that keep the logistic form were built for clinical trials with scheduled visits.
The proposal starts from one observation. In a retrospective electronic health record, the reason a patient vanishes is usually written in the record itself, in the weeks before they vanish: falling visit frequency, changes in lab ordering, a discharge disposition, a new outside referral. We propose to use this post-baseline "care trajectory" only to model who gets censored and when. We then fit a horizon logistic model on baseline covariates alone, with inverse-probability weights and a doubly robust correction. The odds ratios keep their usual meaning, conditional on what a clinician knows at baseline, and they stay valid under censoring that is informative given baseline covariates but explainable given the care trajectory. A key part of the contribution is a way to test this. We build an artificial-truncation benchmark that imposes realistic, learned drop-out on cohorts whose horizon outcome is fully known. This provides the ground truth the field currently lacks.
Core Claim
The target is the usual horizon model, $\text{logit}\, P(T \le \tau \mid X) = \beta^\top X$. Here $T$ is event time, $\tau$ the horizon, and $X$ the baseline covariates. Let $C$ be the censoring time and $A(t)$ the auxiliary care-trajectory history up to time $t$. Write $\tilde{T} = \min(T,\tau)$, and let $\Delta = 1$ if the horizon status is observed, meaning $C \ge \tilde{T}$. We estimate $\beta$ from the weighted score
$$\sum_i \frac{\Delta_i}{\hat G(\tilde T_i \mid X_i, A_i)}\, X_i \big(Y_i - \text{expit}(\beta^\top X_i)\big) + \text{augmentation} = 0,$$
where $Y_i = 1\{T_i \le \tau\}$. The term $\hat G$ is the estimated probability of remaining uncensored up to $\tilde T_i$, given baseline covariates and the trajectory, and it is fit with a flexible discrete-time hazard model. The augmentation term is the usual doubly robust correction built from a model of $Y$ given the history at each censoring time. With it, the estimate stays consistent if either the censoring model or the outcome-given-history model is correct.
The claim has two parts. First, in real retrospective cohorts, much of the censoring that looks informative given $X$ becomes approximately ignorable given $(X, A)$. Second, putting $A$ in the weights rather than in the outcome model recovers the full-follow-up odds ratios, with near-nominal confidence interval coverage, where naive, complete-case, and baseline-only corrections do not.
Relation to Prior Work
The closest work is Estimation of Recurrence of Colorectal Adenomas with Dependent Censoring Using Weighted Logistic Regression. It fits proportional-hazards working models for event and censoring, compresses the two risk scores with PCA into strata, and weights a logistic regression. In the ursodeoxycholic acid (UDCA) trial this revealed a treatment odds ratio of 0.767 (95% CI 0.601 to 0.988) that unadjusted methods missed. In simulation, naive proportions were biased by up to 17% with 0.2% interval coverage. The paper both motivates and constrains our design. It shows that dependent censoring breaks naive binary analysis, and it shows that auxiliary information can repair it. However, it relies on low-dimensional, fixed baseline auxiliaries and protocol-scheduled visits. We replace these with high-dimensional, time-varying trajectories from irregular care, use flexible censoring models, and add double robustness so that a misspecified censoring model does not by itself break the estimate.
Two lines of survival work supply tools but answer a different question. Explicitly Modeling Censoring Produces Superior Survival Predictors separates event-specific, censoring-specific, and shared latent factors and wins 109 of 156 comparisons. It assumes censoring is independent of the event given covariates, and it outputs survival curves rather than odds ratios. Our premise is precisely that baseline covariates are not enough. Deep Survival Machines and DeepSurv improve discrimination on SUPPORT, METABRIC, and SEER, but give neither horizon odds ratios nor a treatment of censoring that depends on post-baseline information. We use them as prediction comparators, not as competitors on inference.
Random Forest for Dynamic Risk Prediction or Recurrent Events: A Pseudo-Observation Approach turns censored outcomes into jackknife pseudo-values for machine learning. Pseudo-values with a logit link are the most obvious competitor for horizon odds ratios. Standard pseudo-values are computed from a marginal Kaplan-Meier curve, so they inherit an assumption of censoring that is independent or depends only on covariates.
The applied studies show the gap in practice. Development of a machine learning-based nomogram for predicting one-year mortality in elderly patients with chronic obstructive pulmonary disease in intensive care unit dichotomizes MIMIC-IV mortality at 365 days, with a test AUC of 0.748. An Extensive Data Processing Pipeline for MIMIC-IV formats only binary targets. Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China restricts to patients who were readmitted, which conditions on a post-baseline event. These are the studies whose estimates our method would change.
Research Program
The work has three parts.
The first is the estimator. A discrete-time censoring hazard model, either gradient-boosted or a small recurrent network over binned trajectories, is fit on $(X, A(t))$. Bins follow the two-hour-to-daily windowing conventions of the MIMIC-IV pipeline, coarsened to weekly bins after discharge. A matching outcome-given-history model supplies the augmentation term. Weights are cross-fitted to avoid overfitting bias, and inference uses a sandwich variance with a bootstrap check.
The second part is the artificial-truncation benchmark. We take cohorts where horizon status is fully known for everyone. One is the MIMIC-IV elderly COPD one-year mortality cohort, rebuilt following the nomogram paper, whose one-year mortality was 40.6%. The others are the fully followed subsets of SUPPORT and METABRIC at horizons shorter than their administrative censoring. In each, the full-data logistic fit defines the true $\beta$ and true horizon risks.
We then impose censoring under three regimes. Regime A depends only on $X$, which is a sanity regime where baseline-only methods should tie with ours. Regime B depends on the trajectory $A(t)$ through a censoring hazard learned from real drop-out patterns in a separate, naturally censored MIMIC-IV cohort. This step matters because it keeps the benchmark from being hand-tuned to favor our estimator. Regime C additionally depends on a variable withheld from both $X$ and $A$, where every method should fail. In regime C we report how far a standard sensitivity analysis on the weights moves the estimates.
The third part is real-data reanalysis. Using naturally censored MIMIC-IV cohorts, we measure how far the odds ratios and nomogram calibration published in the cited studies move when the method is applied.
Baselines
The naive reference fits ordinary logistic regression with every early-censored, event-free patient labeled $y=0$, as in the MIMIC-IV COPD nomogram. The complete-case reference drops these patients. Both are run on identical cohorts and covariates. Their bias relative to the full-data $\beta$ sets the size of the problem. The file reports up to 17% bias with 0.2% coverage for naive estimates in the UDCA simulations, but gives no reported result on Lacuna for naive horizon-odds-ratio bias in MIMIC-IV or SUPPORT.
The strongest existing competitor is the weighted logistic regression of Hsu et al., reimplemented with proportional-hazards working models on the baseline covariates plus a fixed set of summary auxiliaries at baseline. We give it the same covariate set and the same cross-fitting folds. The file reports that its intervals achieved approximately 94% coverage, with bias near zero, in its own simulations. Beating it means lower odds-ratio bias and coverage closer to 95% in regime B, while staying no worse in regime A.
A second strong competitor is pseudo-observation logistic regression, with pseudo-values from Kaplan-Meier and from a covariate-adjusted censoring model, in the spirit of RFRE.PO. A third converts a Cox model, DeepSurv, or Deep Survival Machines into horizon risks. These are compared on calibration, Brier score at $\tau$, and AUC, not on odds ratios, which they do not produce. Each competitor gets the same hyperparameter search budget, set by number of configurations, on the same validation folds. For reference, DeepSurv reports a METABRIC C-index of 0.867 against 0.816 for Cox. That is a continuous-time discrimination figure and is not directly comparable to horizon metrics.
The decisive ablation is our estimator with the trajectory removed from the censoring model, keeping only baseline covariates while holding the model class fixed. A second ablation drops the augmentation term to isolate double robustness. If removing the trajectory does not increase bias in regime B or on real censoring, the central claim fails.
Why not just…?
Why not just use pseudo-observation logistic regression? It gives horizon odds ratios directly and is far simpler, so it is the most serious threat. It is valid when censoring is independent of outcome given the covariates used to compute the pseudo-values. Extending it to time-varying auxiliaries is not standard. Regime A should show it matching our estimator, and we would report that tie honestly. Regime B and the real-cohort comparison show whether the tie holds. If pseudo-values computed with a trajectory-aware censoring model match us, the contribution narrows to the benchmark and the empirical finding, and we would say so.
Why not just fit a Cox model and read off the risk at $\tau$? Cox regression handles censoring, but with baseline covariates only it assumes the same ignorability as baseline-only weighting. It also gives hazard ratios, which are not horizon odds ratios. Deep Survival Machines documents its weakness at long-horizon quantiles. The Cox-to-horizon baseline tests this directly through calibration at $\tau$ under regime B.
Why not just put the care trajectory into the outcome model? That would condition the odds ratios on post-baseline events, which are partly consequences of risk. The odds ratios would no longer be usable at baseline, and the model would gain the kind of leakage the SEER nomogram shows with concurrent metastases inflating AUC from 0.716 to 0.849. We will run this variant to quantify how much its apparent discrimination gain is leakage.
Why not just use Hsu et al.'s method as is? Its PCA-compressed proportional-hazards scores assume low-dimensional, fixed auxiliaries. The head-to-head comparison in regime B tests whether flexible, time-varying censoring models actually add anything over it.
Decisive Checks
The proposal succeeds if three things hold. In regime B, the trajectory-weighted estimator recovers the full-data odds ratios with interval coverage near 95% while baseline-only weighting and pseudo-values fall clearly short. In regime A, it ties them. In regime C, it fails visibly, which shows the benchmark is not rigged. It fails if baseline-only methods match it under drop-out learned from real data. That outcome would itself be useful evidence that retrospective electronic health record censoring is less informative than feared.
References
- Estimation of Recurrence of Colorectal Adenomas with Dependent Censoring Using Weighted Logistic Regression
- Explicitly Modeling Censoring Produces Superior Survival Predictors
- Deep Survival Machines: Fully Parametric Survival Regression and Representation Learning for Censored Data With Competing Risks
- DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network
- Random Forest for Dynamic Risk Prediction or Recurrent Events: A Pseudo-Observation Approach
- Development of a machine learning-based nomogram for predicting one-year mortality in elderly patients with chronic obstructive pulmonary disease in intensive care unit
- An Extensive Data Processing Pipeline for MIMIC-IV
- Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China
- Nomogram Development for Predicting Synchronous Lung Metastasis in Patients with T1 Colorectal Cancer: An SEER-Based Analysis