Many clinical prediction problems have a few dozen to a few hundred patients described by a table of measurements, and the outcome is a yes/no diagnosis or a time until death that is often cut off before it is observed. Image tricks like cropping and flipping have no meaning on such tables. One recent tabular method, Counterfactual Residual Data Augmentation for Regression (CRDA), makes new training rows in three steps. It fits a model, keeps each patient's prediction error (the "residual"), and changes only those features that appear statistically unrelated to that error. It then adds the same residual back to the new prediction. This works for continuous targets. The authors restrict it to regression, because a yes/no label has no residual you can add back.
This note proposes a fix based on one observation. A binary label and a cut-off survival time are the same kind of partial information. Each says only that a hidden continuous residual lies in some interval. For a label, the hidden residual sits above or below the model's decision threshold. For a patient still alive at last follow-up, the hidden log-time residual sits above a known bound. If we sample the residual from its distribution inside that interval and then carry it to the counterfactual patient, we get one augmentation rule that covers classification and survival. The rule can flip a label, or turn a censored record into an observed event, exactly when the changed features push the patient across the boundary. Those boundary cases are where small cohorts are most short of information.
Core Claim
Residual-preserving counterfactual augmentation extends to discrete and censored outcomes once the residual is treated as a hidden variable known only up to an interval. In cohorts below about 200 patients, this should beat label-copying augmentation, heuristic noise and mixup, and no augmentation. The gain should come from the generated samples whose label or event status changes. The extension should also fail more gracefully than CRDA when the base model is too weak to give trustworthy residuals. CRDA's authors flag that failure as a known sample-size "sweet spot."
The proposal takes CRDA's method and generalizes it. The application regime comes from the small-cohort clinical literature, especially heart-failure survival studies with N < 200, as in Boosting Transfer Learning with Survival Data from Heterogeneous Domains.
The Mechanism
For classification, write the outcome as a thresholded hidden score:
$$y = \mathbf{1}[f(x) + \varepsilon > 0],$$
Here $f$ is a fitted base model on the logit or probit scale, and $\varepsilon$ is a patient-specific residual with a known standard distribution (logistic or Gaussian). Observing $y=1$ tells us only that $\varepsilon > -f(x)$. Observing $y=0$ tells us $\varepsilon \le -f(x)$.
To augment a patient, we first draw $\tilde\varepsilon$ from the residual distribution truncated to that interval. We then build $x'$ by perturbing only the features $X_p$ that pass CRDA's independence screen, which combines the PC algorithm with correlation tests against the residuals. The new label is $y' = \mathbf{1}[f(x') + \tilde\varepsilon > 0]$.
For survival we use an accelerated-failure-time form:
$$\log T = g(x) + \sigma\varepsilon,$$
Here $T$ is the event time, $g$ is the fitted model, and $\sigma$ is a scale parameter. An observed event pins $\varepsilon$ exactly, which recovers CRDA as a special case. A censored record at time $C$ tells us only that $\varepsilon > (\log C - g(x))/\sigma$.
The counterfactual record moves both the event time and the censoring time by the same amount, $g(x') - g(x)$. This keeps the censoring structure faithful.
We add two changes aimed at very small $n$. First, residuals are computed out-of-fold, using cross-fitting, so that $f$ does not see the patient whose residual it produces; in-sample residuals shrink toward zero. Second, the base model is deliberately simple: penalized logistic or AFT regression, or shallow boosting.
We keep CRDA's paired Wilcoxon safety filter, which rejects the augmented set if cross-validated loss gets worse. We extend it from mean squared error to log-loss and the integrated Brier score.
Research Program
The work has three parts.
The first part builds the method. It produces one implementation covering binary, multiclass (using a latent ordinal or multinomial-probit form), and right-censored outcomes, plugged into downstream logistic regression, small MLPs, XGBoost, and Cox or survival-boosting models.
The second part is analysis on synthetic data, where the true decision boundary and causal structure are known. We vary three things: sample size, how well the screen separates truly independent features, and how misspecified $f$ is. This tells us when sampling residuals inside their interval recovers the true boundary, and when it amplifies the base model's bias. That bias amplification is the confirmation-bias risk seen in self-training.
The third part is a real-cohort study on small target cohorts. It asks whether the augmentation adds anything on top of the strongest transfer method.
Baselines
The naive reference is the downstream model trained on the target cohort alone, with no augmentation. This means penalized logistic regression or XGBoost for classification and a Cox model for survival, tuned by nested cross-validation. On the MAGGIC Euro study, a target-only Cox model reaches a C-index of 0.500, as reported in Boosting Transfer Learning with Survival Data from Heterogeneous Domains. Any method that does not clearly clear this on the small studies is not useful.
The heuristic augmentation baselines are Gaussian feature jitter, random feature dropout, mixup on rows and labels, and SMOTE-style interpolation within each class. These are what a practitioner would try first. CRDA reports that comparable generic generators often failed to help on tabular regression. We give each one the same number of synthetic rows, the same augmentation-ratio search, and the same Wilcoxon filter.
The key ablation is label-copy CRDA. It perturbs the same screened features $X_p$ but simply keeps the original label, or keeps the original time and censoring status. This isolates whether sampling the residual inside its interval matters, or whether the gain comes only from feature perturbation. A second ablation replaces the interval-truncated draw with an unconstrained draw from the full residual distribution. That turns the method into model-based synthetic labeling and tests whether conditioning on the observed outcome matters. A third ablation removes cross-fitting.
The strongest existing competitor on survival is Transfer Survival Boosting (TSB). It re-weights source-cohort patients by a censoring-adjusted Brier score so that only helpful source patients count. TSB reports the best C-index in four of five MAGGIC target studies, including 0.677 on Euro, all in Boosting Transfer Learning with Survival Data from Heterogeneous Domains. We rerun TSB on our splits rather than copying its numbers, using the same source studies, folds, and tuning budget. We also run TSB combined with our augmentation applied to the target cohort.
"Beating" a baseline means a higher mean C-index (or AUROC and log-loss for classification) in at least four of five target studies, with paired bootstrap intervals that exclude zero. For TSB, the realistic and pre-registered success criterion is that TSB plus our augmentation beats TSB alone. We do not expect target-only augmentation to beat cross-study transfer.
For classification on cohorts subsampled from large repositories, no reported result was found on Lacuna for any augmentation method, so the baselines above are the reference.
Why not just…?
Why not just use mixup or Gaussian jitter? These are cheap and sometimes enough. They do not respect which features can safely change, though. Mixup also invents soft labels with no link to any patient's actual outcome. CRDA's comparison against generic generators showed frequent degradation on tabular data, with 22.9% average MSE reduction for CRDA on MLPs. That evidence is for regression only, so the same heuristics are run here with identical budgets and safety filtering. If they match our method on classification, we will report that.
Why not just perturb the safe features and copy the label? This is the most serious threat. If perturbations rarely move patients across the decision boundary, label-copy and our method generate nearly the same data. The ablation settles it directly. We log the fraction of generated samples whose label or event status changes, and we measure the downstream gain separately for flipped and unflipped samples. If the gains are equal and flip rates are near zero, the interval-residual idea adds nothing, and we would narrow the claim to survival. In survival, turning censored records into events changes the information content even without any label flips.
Why not just use a strong, well-regularized gradient-boosting model with no augmentation? CRDA's own gain shrank from 22.9% with MLPs to 6.4% with XGBoost. This suggests that much of the benefit acts as regularization for flexible models. We therefore report results for each downstream model separately. We count the method as useful for tree models only if it beats a properly tuned XGBoost, not an untuned one.
Why not just do transfer from larger cohorts instead? Where related source cohorts exist, as in MAGGIC, transfer is likely the larger effect. This is why TSB is the strongest baseline. The two methods address different gaps, however. Transfer brings in outside patients, while augmentation expands the target cohort around its own decision boundary. The TSB-plus-augmentation arm tests whether they add up. Many rare-disease settings have no source cohort at all, and there target-only augmentation is the only option.
Proposed Experiments
On synthetic data, we generate logistic and accelerated-failure-time cohorts with n from 25 to 400. Each has a known set of features independent of the noise and a controllable fraction of censoring. We measure distance to the true boundary and calibration. The decisive check is that interval-residual augmentation approaches the true boundary faster than label-copy as n shrinks, and that cross-fitting moves the n at which the safety filter starts rejecting everything to a lower value than CRDA's in-sample residuals.
On real data, the main testbed is the MAGGIC heart-failure repository. Its five target studies (DIAMO, DIG, ECHOS, Euro, IN-CH), each with N < 200, are evaluated for all-cause survival (C-index and integrated Brier score) and for a derived fixed-horizon mortality label (AUROC, log-loss, and expected calibration error). The patients with fixed-horizon labels are a subset; those censored before the horizon are handled by our censored-residual form rather than being dropped.
To confirm that classification gains do not depend on one domain, we add tabular classification tasks from the UCI and PMLB repositories that CRDA used for regression, subsampled to n of 50, 100, and 200 over 20 random draws. We also re-run CRDA's regression benchmarks to verify that our implementation reproduces its regression behavior when every residual is fully observed.
A stress test subsamples the UK Biobank plasma proteomics cohort, with 2,923 protein features. There the independence screen will likely break down because features far outnumber patients. The point is to state the method's limits honestly, not to compete with ProMeta: A meta-learning framework for robust disease diagnosis and prediction from plasma proteomics, which reports a mean 4-shot diagnosis AUROC of about 0.69.
The three decisive results are as follows. The method must beat label-copy augmentation, with the gain concentrated on boundary-crossing samples. It must hold up at sample sizes where CRDA-style in-sample residuals trigger the safety filter. And it must add to TSB on MAGGIC rather than merely match it.