Many emergency treatments only help if they arrive in time. A drug that protects heart tissue when given before a blocked artery is reopened may do nothing, or harm, when given six hours later. Animal studies often dodge this question by dosing before or during the injury. Clinical cohorts usually answer it with crude bins, "early versus late." This proposal estimates the whole curve instead. From routine intensive-care records, we estimate how a rescue therapy's effect on the death rate changes as a smooth function of the delay between injury onset and treatment. From that curve we read off where, if anywhere, the benefit reaches zero.
The statistical obstacle is specific. The most relevant modern method for treatments that start at different times, Time-Varying Causal Survival Learning, corrects the bias that comes from patients having to survive long enough to be treated. To do so it assumes "invariance to history": once treatment starts, its effect does not depend on when it started. For rescue therapy that assumption rules out exactly the quantity we want to measure. We drop it. We replace the single treatment coefficient with a function of delay, keep the method's resistance to badly specified models of patient risk and of when doctors start treatment, and test invariance to history directly rather than assuming it.
Core Claim
Write $h(t)$ for a patient's instantaneous death rate at time $t$ after injury onset. Write $L_t$ for the patient's recorded clinical history up to $t$, $s$ for the time rescue treatment began, and $A_t$ for an indicator that treatment has started by $t$. We model
$$h(t \mid L_t, s) = h_0(t)\,\exp\big(g(L_t) + A_t\,\beta(s)\big), \qquad \beta(s) = \sum_{k=1}^{K} \theta_k B_k(s),$$
Here $h_0$ is an unspecified baseline death rate and $g$ is a flexible risk function of history, fitted by a neural network as in DeepSurv. The term $\beta(s)$ is the log hazard ratio of treatment begun at delay $s$, written as a combination of $K$ spline basis functions $B_k$ with coefficients $\theta_k$. Invariance to history is the special case where all $\theta_k$ are equal, which collapses $\beta(s)$ to a constant.
The claim is that $\theta$ can be estimated at the usual $\sqrt{n}$ rate with valid confidence bands, even though both nuisance pieces are learned by flexible machine learning. The two nuisance pieces are $g$ and the model of when clinicians initiate treatment given deteriorating vitals. The route is to extend the orthogonal score construction of TV-CSL from a scalar effect to the spline coefficients. Two derived quantities matter clinically. The first is the delay $s^\ast$ at which $\beta(s^\ast)=0$, the point of no return. The second is the slope of $\beta$ at clinically typical delays, which tells you how much each hour of delay costs.
Research Program
The work has three parts.
The first part is the estimator. It takes TV-CSL's double machine learning recipe, which subtracts out errors in two separately fitted nuisance models so that neither has to be exactly right. In our version the target is the vector $\theta$, not one coefficient. The orthogonal score is built on a discretized person-hour risk set. In each hour, a patient still untreated contributes an initiation hazard estimate, and every patient contributes an outcome hazard estimate. Both are fitted with cross-fitting.
Two pieces need new derivation and checking. The first is that the score stays orthogonal when the treatment term depends on $s$. The second is the variance of smooth functionals like $s^\ast$, obtained by the delta method on the estimated $\theta$. We also give a Wald test of the hypothesis that all $\theta_k$ are equal, which is a direct test of invariance to history.
The second part is the application, the timing of continuous renal replacement therapy (CRRT, round-the-clock dialysis for critically ill patients with failing kidneys) after acute kidney injury in MIMIC-IV. CRRT suits the question because start times vary widely across patients. It also carries heavy confounding by indication: the sickest patients are started first, so naive analyses can make the therapy look harmful. And it already has a strong prediction literature to compare against.
Injury onset is the first hour meeting kidney injury stage 2 by the standard creatinine and urine-output criteria. Patients enter follow-up at onset, and the outcome is 28-day mortality with discharge treated as censoring. The history $L_t$ includes hourly vitals, lactate, creatinine slope, fluid balance and vasopressor dose. We take the cohort definitions and the 28-day endpoint from the MIMIC-IV intra-abdominal pressure study, which combined splines with Cox regression to find a 16 mmHg risk threshold in this database, though only for averaged baseline pressure.
The third part is a short design note for preclinical work. Animal rescue studies typically test one arbitrary post-injury time point. Examples include the semaglutide study in heart ischemia–reperfusion injury, which pretreated the animals (Semaglutide attenuates myocardial ischemia-reperfusion injury), and the ferrostatin-1 study in sepsis-related heart damage (Ferrostatin-1 improves acute sepsis-induced cardiomyopathy). Using simulations from the fitted curve shapes, we will report how many delay arms, and at which delays, are needed to estimate $s^\ast$ to a given precision. These animal studies motivate the estimand; they do not supply data.
Baselines
The naive reference is a time-dependent Cox model with one treatment indicator, adjusted for the same history variables, giving one hazard ratio for "on CRRT." This is the model that confounding by indication tends to push toward an apparent harm. Any method claiming to recover a delay curve should first show it removes that artifact.
The second baseline is binned delay with weighting, as in the BCG delay study. That study weighted patients by propensity scores, binned delay into week ranges and fitted Cox models. It found progression hazard ratios of 1.62 for recurrence and 3.34 for progression at 19–25 versus 6–10 weeks. It is the standard clinical-journal approach to treatment delay. We run it with the same covariates and with hourly bins chosen by quantiles of observed delay.
The third baseline is the strongest simple competitor. It is a time-dependent Cox model with a restricted cubic spline in delay (the same $B_k$ basis), where the confounders enter as linear time-varying covariates, fitted by maximum partial likelihood. It differs from our estimator only in lacking the orthogonal correction and the flexible nuisance models. It is therefore the ablation that isolates our key ingredient. In the semi-synthetic study it and our method get identical basis functions, covariates and cross-validation budget.
The fourth baseline is TV-CSL as published, with constant effect after treatment start. Its reported Stanford Heart Transplant result is the reproduction target. A naive Cox interaction between prior surgery and transplant had coefficient −2.191 with p < 0.01, which shrank to −0.557 with p = 0.47 under TV-CSL (Time-Varying Causal Survival Learning).
For flexible nuisance and risk modeling we include Deep Cox Mixtures as a calibration reference. It reports expected calibration error, a measure of how far predicted risks sit from observed frequencies, of 0.0200 against 0.0992 for standard Cox on the FLCHAIN dataset. On MIMIC delay-curve estimation, no reported result was found on Lacuna for any of these methods.
"Beating" a baseline means two things in the semi-synthetic study: lower integrated squared error of the estimated $\beta(s)$ and 95% band coverage near nominal while the baseline undercovers. On real data, it means our estimate stays stable under the negative-control checks below while the baseline's does not.
Why not just…?
Why not just predict mortality at CRRT start with gradient boosting? Data-driven prediction of CRRT survival reached an AUROC (a 0-to-1 score of how well a model ranks patients who die above those who survive) of 0.848 on 8,986 patients this way. But it predicts outcomes for patients who were started on CRRT. It does not compare starting now with starting later. Its top predictor, surgery before CRRT, illustrates that risk ranking and treatment effect are different questions. We will use such a classifier only as a nuisance model and as a check on how well we rank risk, not as a competitor for the effect curve.
Why not just put a spline on delay in an ordinary time-dependent Cox model? This is the real threat, and it is Baseline 3. If confounding runs through a few variables with roughly log-linear effects, the spline Cox model may recover $\beta(s)$ as well as our method does. The semi-synthetic study settles this by varying the confounding from linear to strongly nonlinear, for example initiation driven by interactions of lactate and creatinine slope. If the spline Cox model matches our method across that range, we will say so. The contribution then shrinks to the estimand, the invariance-to-history test and the design guidance, and we will recommend the simpler model.
Why not just emulate a sequence of trials, "start now versus wait," at each hour? Hour-by-hour target trial emulation handles immortal time well. It yields separate effect estimates per start hour, which are noisy at sparsely populated delays and come with no joint inference on a crossing point. Our method can be seen as pooling those per-hour contrasts through a shared smooth $\beta(s)$. We will run the hourly emulation with inverse-probability weights as a further comparison and show whether the pooling buys tighter bands at equal bias.
Proposed Experiments
The decisive experiment is semi-synthetic, since real data has no true delay curve. We keep real MIMIC-IV covariate trajectories and fit an initiation model to them. We then simulate CRRT start times from that model, with deliberately strong confounding by deterioration. Deaths are simulated from known decay curves of three shapes: flat (invariance holds), monotone decline crossing zero at 24 hours, and a curve that turns harmful after the crossing.
Across sample sizes from 1,000 to the full cohort, we report three things for each method. The first is integrated squared error of $\beta(s)$. The second is the coverage of 95% bands. The third is error in $s^\ast$, together with the size and power of the invariance-to-history test. We expect the orthogonal estimator to keep coverage near 95% under nonlinear confounding while Baselines 2 and 3 undercover. If that does not happen, the method has no advantage.
Second, we reproduce the Stanford Heart Transplant numbers above with our implementation restricted to constant $\beta$. We then apply the full method there. Transplant benefit plausibly depends little on waiting time once a patient is matched, so a non-rejection of invariance would serve as a sanity check.
Third, we fit the CRRT curve in MIMIC-IV and run two falsification checks. In the first, we shift every start time by a random placebo offset; a valid method should flatten the curve toward zero. In the second, we use a negative-control outcome that CRRT should not affect, such as new pressure ulcer documentation; it should give a null curve. We report $s^\ast$ with its confidence interval, or report that no crossing is supported. With access to the UCLA–Cedars-Sinai cohort we would test whether the curve transports to another hospital system.
References
- Time-Varying Causal Survival Learning
- DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network
- Elevated IAP in critically ill patients associated with increased AKI incidence: a cohort study from the MIMIC-IV database
- Semaglutide attenuates myocardial ischemia-reperfusion injury by inhibiting ferroptosis of cardiomyocytes via activation of PKC-S100A9 axis
- Ferrostatin-1 improves acute sepsis-induced cardiomyopathy via inhibiting neutrophil infiltration through impaired chemokine axis
- Delaying BCG immunotherapy onset after transurethral resection of non-muscle-invasive bladder cancer is associated with adverse survival outcomes
- Deep Cox Mixtures for Survival Regression
- Data-driven prediction of continuous renal replacement therapy survival