Clinical trials often bundle several bad outcomes into one "composite endpoint," for example death, reinfarction, or new heart failure. A patient counts as an event if any of them happens. Bundling raises the number of events and therefore statistical power, but it quietly assumes the bundled outcomes reflect one underlying disease process. If death after a heart attack is driven by an acute immune storm and later heart failure is driven by slow metabolic remodeling, then one composite blurs two biological signals. A biomarker can look weak against the composite when it is in fact strong for one component and irrelevant to the other.
We propose to test this assumption directly. Repeated blood gene-expression measurements would be converted into a small number of pathway activity curves over time, one curve per biological program such as natural-killer-cell cytotoxicity or oxidative metabolism. We then ask whether the different components of a composite endpoint load on the same combination of these curves, or on different ones. The output is a coherence test for composite endpoints, together with a defensible way to weight or split them. The bet is that timing is what reveals discordance. Components that look molecularly identical at admission can separate once one sees which pathway is rising when each event occurs.
Core Claim
Each direction in this work plays a different role. Composite endpoints are the object under study. Temporal expression profiling, meaning repeated transcriptome measurement in the same person, supplies the signal. Functional enrichment, which summarizes thousands of genes into scores for curated gene sets, is the dimension-reduction step that makes small longitudinal cohorts tractable. Longitudinal biomarker modeling supplies the statistical machinery and its known pitfalls.
The specific claim is that a composite endpoint is biologically coherent only if a single direction in pathway-trajectory space drives the hazard of every component, up to a component-specific scale. This is a testable restriction. Our prediction for acute myocardial infarction (AMI) is that it fails in a structured way. Early components should track the acute immune pathways. Later components should track the remodeling and metabolic pathways. This is the transition Liu et al. described qualitatively in the serial cohorts GSE59867 and GSE62646, but never linked to any patient outcome.
Formally, let $z_i(t)\in\mathbb{R}^d$ be a low-dimensional summary of patient $i$'s smoothed pathway-score trajectories up to time $t$. Each component $k$ (death, reinfarction, heart failure, and so on) gets its own hazard:
$$h_{ik}(t) = h_{0k}(t)\exp\big(\beta_k^\top z_i(t)\big),$$
Here $h_{0k}$ is the baseline hazard of component $k$ and $\beta_k$ is its coefficient vector. Coherence is the hypothesis
$$H_0:\ \beta_k = c_k\,\beta \ \text{ for all } k,$$
This says there is one shared risk direction $\beta$ and only the scale $c_k$ differs by component. Under $H_0$, an unweighted composite discards little. When $H_0$ is rejected, the fitted $\beta_k$ show which pathways split the composite and when.
Research Program
Prior work on this question falls into three groups. The first motivates the gap. Higgins et al. built a composite of death or new disability after critical illness and predicted it from admission variables alone, with an AUROC of 0.76. They had no biology and did not ask whether death and disability share predictors. PESSA and interpretable meta-learning for survival link pathway scores to survival, but only from a single baseline sample and only for single-event cancer outcomes.
The second group supplies methods. ssGSEA, as used in PESSA, gives a per-sample score for each pathway. The Gaussian-process joint model of Masouleh et al. fits a smooth trajectory and the survival hazard together, so uncertainty in the trajectory carries through to the risk estimate. It also showed that within-person volatility is itself a risk factor, with a 1.2-fold mortality increase per unit of volatility in USRDS hemodialysis patients.
The third group constrains the design. That joint model was demonstrated on a single biomarker and is known to scale poorly to many. TransformerLSR handles several clinical variables with irregular visits but needs far more patients than any longitudinal transcriptomic cohort has.
Given these constraints, we would build three things.
Trajectory construction. Per-sample ssGSEA scores are computed on a curated gene-set collection, Hallmark plus a reduced Reactome set, to keep the number of pathways in the low hundreds. These scores are then smoothed with a hierarchical functional model, a population mean curve plus patient-specific deviations. The model borrows strength across patients because each patient contributes only about four samples. We then compress the pathways into $d\approx 5$–10 latent trajectory factors, each with a level, a slope, and a volatility term.
Component-specific hazards with the coherence test. We fit the hazards above jointly with the trajectory model. Joint fitting is the default. Two-stage fitting, where trajectories are estimated first and plugged in, is a declared ablation rather than an assumption. We report a likelihood-ratio statistic for $H_0$ and calibrate it by parametric bootstrap.
A semi-synthetic benchmark. This is built from the real GSE59867/GSE62646 expression trajectories, so it has realistic noise and sparse sampling. Component event times are simulated with planted links between pathways and components. The benchmark lets us measure type I error and power of the test, which no real cohort can provide.
There is a data risk we must state plainly. The file documents serial expression in GSE59867/GSE62646 but no reported outcome linkage. First, we would audit the GEO sample annotations for follow-up events. If they are absent, real-data validation of the transcriptomic arm requires a collaborating AMI or ICU cohort with serial blood RNA. Meanwhile, the coherence test itself can be validated on clinical-chemistry trajectories against a composite that is already defined. One option is DIVAT graft failure, the composite of return to dialysis or death, using serial creatinine and tacrolimus as in TransformerLSR. Another is an AMI registry of the kind analyzed by Zhang et al.. There, albumin change related differently to death, gastrointestinal bleeding (OR 0.87), and acute heart failure (OR 0.86), which already hints at component discordance from a single biomarker.
Baselines
The naive reference is the unweighted composite modeled with a single Cox model on admission covariates (age, severity) plus baseline-sample pathway scores. This is the PESSA recipe applied to a composite. It shows what a cross-sectional analysis concludes. The bar it sets is low, but it is the analysis most papers would actually run. For context, the closest reported composite-endpoint number in the file is AUROC 0.76 from Higgins et al. on the PREDICT ICU cohort. That is a different population with no omics, so it frames the discrimination ceiling of clinical scores and is not a head-to-head target.
The strongest competitor on this specific question is a two-stage, component-specific analysis. Per-sample ssGSEA scores are summarized by per-patient slope and change from baseline using ordinary least squares, then entered into separate cause-specific Cox models for each component, and the coefficient vectors are compared with a standard heterogeneity test. It is strong because it answers nearly the same question with off-the-shelf tools. With about four timepoints per patient, it may lose little. No reported result was found on Lacuna for any longitudinal-pathway model of a composite endpoint, so this baseline has no published number. We will establish it ourselves on identical data splits, the same pathway collection, and the same latent dimension $d$.
The strongest existing joint model is TransformerLSR, run on the same latent pathway factors plus clinical covariates. The file reports only that it beat joineRML, TransformerJM, and MFPCA-DS on integrated Brier score in DIVAT, with no numeric value. We give it a hyperparameter budget equal to ours, measured in fits, and cross-validated selection.
Ablations isolate the key ingredients: snapshot-only features versus trajectory features, two-stage versus joint fitting, and pathway factors versus a LASSO-selected gene panel.
Beating the baselines is defined on two fronts. On the semi-synthetic benchmark, the method must recover planted pathway-to-component links that the two-stage heterogeneity test misses, at matched false-positive rate. On real or clinical-analog data, it must improve component-specific time-dependent C-index or integrated Brier score. If the two-stage baseline matches us on both, we report that and recommend it.
Why not just…?
Why not just fit separate Cox models per component using delta-change features? This is the real threat, and it is the strongest baseline above. Masouleh et al. show that two-stage estimation biases effects by treating noisy estimated trajectories as known. They also show it discards volatility, which was an independent predictor in their data. Whether that bias matters with four sparse samples and pathway-level scores is unknown. The joint-versus-two-stage ablation on the semi-synthetic benchmark settles it, because there the true links are known. If the gap is negligible, the coherence test is the contribution and two-stage fitting becomes the recommended implementation.
Why not just use the baseline sample? A snapshot cannot express the hypothesis that components diverge over time. Biomarker studies in the file repeatedly found trajectories more informative than single values: albumin change in Zhang et al., hsCRP trajectories in Pan et al., and the neutrophil-to-lymphocyte ratio in Salciccioli et al.. Those are single clinical markers, though, not pathway scores. The snapshot ablation tests directly whether discordance is detectable only with trajectories.
Why not just select individual genes with LASSO? With roughly a hundred patients and around 20,000 genes at each of four timepoints, gene-level selection is unstable. A gene panel also does not say which biological program separates the components. Pathway summaries pool co-regulated genes, which is the case made by PESSA's cross-cohort pathway stratification. We still run the gene-panel ablation. If it matches pathway factors on discrimination, the pathway step is justified on interpretability alone, and we say so.
Proposed Experiments
The first decisive check is calibration and power on the semi-synthetic benchmark. We vary how many components share the risk direction, how far apart in time their events fall, and the sampling density. The coherence test should hold nominal type I error when components are coherent. It should detect time-separated discordance that a snapshot version of the same test cannot.
The second check is a real-data hypothesis test. In serial AMI blood transcriptomes with linked outcomes, early events should load on the immune-cytotoxicity factor and later heart-failure or remodeling events on the metabolic factor. A rejection of $H_0$ in this direction would be the substantive finding. A failure to reject is also informative: it would mean the standard AMI composite is defensible at the molecular level.
The third check is the clinical-analog replication on DIVAT graft failure, testing whether dialysis return and death share a creatinine and tacrolimus risk direction. It shows the test is not specific to omics.
Across all three, we report component-specific and composite time-dependent C-index and integrated Brier score against the baselines above. We also report test-retest stability of the smoothed versus raw pathway scores between adjacent stable-phase samples. That stability comparison checks whether temporal smoothing actually reduces ssGSEA noise.
References
- Potential biomarkers of acute myocardial infarction based on weighted gene co-expression network analysis
- Predictors of death and new disability after critical illness: a multicentre prospective cohort study
- PESSA: A web tool for pathway enrichment score-based survival analysis in cancer
- Interpretable meta-learning of multi-omics data for survival analysis and pathway enrichment
- A Flexible Joint Longitudinal-Survival Model for Analyzing Longitudinally Sampled Biomarkers
- TransformerLSR: Attentive joint model of longitudinal data, survival, and recurrent events with concurrent latent structure
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study
- The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness: an observational cohort study