CEA and Colorectal Cancer Recurrence

Why Link CEA and Recurrence?

After colorectal cancer surgery, follow-up commonly includes serial measurements of carcinoembryonic antigen (CEA) and surveillance for recurrence. CEA can rise as occult disease develops, sometimes before recurrence is confirmed. But CEA is an imperfect marker: a rise does not establish recurrence, and recurrence can occur without a marked rise. A model of CEA alone does not directly describe recurrence timing, while a time-to-event model using only baseline information ignores the evolving marker history.

A joint model connects the two outcomes. Its longitudinal component estimates each patient's underlying CEA trajectory despite measurement error; its event component relates the instantaneous recurrence hazard to that latent trajectory. The case below is hypothetical. It illustrates interpretation, not a clinical rule for acting on an individual CEA result.

A Case-Specific Joint Model

For illustration, suppose the longitudinal submodel is

Here \(A_i(t)\) indicates whether the patient is receiving adjuvant chemotherapy at time \(t\), \(b_{0i}\) and \(b_{1i}\) are patient-specific random effects, and \(\epsilon_{ij}\) is measurement error. For the usual Gaussian linear mixed model, assume \(\epsilon_{ij}\) has mean zero and variance \(\sigma^2\), and that the random effects follow a specified distribution, commonly a mean-zero multivariate normal distribution. The model makes the treatment indicator and time structure explicit; a more flexible time function may be needed if CEA changes nonlinearly.

The linked event submodel can be written

where \(h_0(t)\) is the baseline recurrence hazard, \(Z_i(t)\) contains other modeled covariates, and \(\alpha\) is the current-value association parameter. The random effects \(b_i=(b_{0i},b_{1i})^\mathsf{T}\) appear in both components, linking a patient's estimated CEA course to that patient's event risk. Conditional on the random effects and modeled covariates, the longitudinal measurement errors and event process are assumed independent. The hazard is conditional on the marker trajectory and covariates; it is not a causal effect of changing CEA.

The model requires a correctly specified longitudinal mean and event hazard, an appropriate random-effects and measurement-error distribution, and a suitable account of covariates that influence both marker values and recurrence. It also assumes censoring is noninformative conditional on the modeled information. The standard formulation treats recurrence times as observed at their recorded times. These assumptions deserve scrutiny in a surveillance setting, where clinic visits, blood tests, and imaging schedules can be related to symptoms or prior results.

What the Association Parameter Means

In this specification, increasing latent log CEA by one unit multiplies the instantaneous recurrence hazard by \(\exp(\alpha)\), holding the other modeled quantities fixed. Since one unit on this log scale corresponds to multiplying CEA by \(e\), a more directly interpretable contrast is a doubling: doubling latent CEA changes \(m_i(t)\) by \(\log(2)\), so the hazard ratio is \(\exp\{\alpha\log(2)\}\). This is a conditional hazard comparison at the same time and covariate values, not a comparison of cumulative recurrence probabilities.

Reading Marker Changes Around Treatment

Adjuvant treatment complicates the clinical story. Treatment may affect CEA, recurrence risk, and the timing or frequency of follow-up. Including \(A_i(t)\) in the longitudinal submodel lets the modeled mean shift while treatment is being received, conditional on time and the random effects. Including treatment in \(Z_i(t)\), if scientifically justified, lets the hazard also vary with treatment status. Neither inclusion by itself makes treatment comparisons causal: treatment is not necessarily assigned at random, and prognosis, treatment decisions, and monitoring can be related.

The analysis uses the latent mean \(m_i(t)\), not each noisy measurement \(Y_{ij}\), in the event submodel. This distinction matters when a single high CEA result could reflect measurement error or a transient fluctuation. In practice, the fitted trajectory is informed jointly by the observed measurements and event experience, with uncertainty carried through the model rather than treating a predicted trajectory as known without error.

Derivation and Event-Time Interpretation

The hazard ratio for a doubling follows directly from the event submodel. Compare two otherwise identical patients at a fixed time, with latent log CEA values \(m\) and \(m+\log(2)\). Their hazard ratio is

The baseline hazard and shared covariate term cancel, leaving the stated contrast. Because this is a hazard ratio, it compares instantaneous event rates among patients still recurrence-free at that time; it does not directly give an absolute recurrence risk.

For context, if a fitted model supplied a constant reference hazard of 0.040 per year and the doubling contrast were 1.625, the corresponding hazard at the doubled marker value would be \(0.040\times1.625=0.0650\) per year, at the same covariate values. If that hazard truly remained constant over one year, the one-year risk would be \(1-\exp(-0.0650)=0.0629\), or about 6.29%. The constant-hazard condition is essential for that calculation; in general the baseline hazard varies over time and absolute risk requires integration over the relevant interval.

Diagnostics, Alternatives, and Limits

A credible case analysis checks whether the data and fitted model support the interpretation. Inspect observed CEA values against fitted trajectories by patient, especially around surgery, chemotherapy, and recurrence. Look for systematic curvature, abrupt changes the model cannot represent, outlying values, and residual patterns that suggest nonconstant measurement variance or serial correlation. Check whether estimated random-effect variability is plausible and whether numerical warnings, convergence, or sensitivity to starting values undermine the fit. Event-model checks should assess the adequacy of the baseline hazard and proportional-hazards structure; compare scientifically defensible association structures when current value alone may be insufficient.

The current-value model is not the only choice. A slope association relates hazard to the rate of marker change, and a cumulative association relates it to accumulated marker exposure. A model that includes more than one of these terms can be useful only when supported by the scientific question and data; highly correlated marker summaries can make their effects difficult to distinguish. A time-dependent Cox model using the latest observed CEA is another approach, but it treats the observed value as the predictor rather than modeling measurement error and the underlying trajectory. A landmark analysis conditions on information available at a chosen time and can answer a time-specific prediction question, but it does not use the same full joint likelihood.

Recurrence is often detected at a surveillance visit rather than at its biological onset. If the event is known only to have occurred between negative and positive assessments, treating the positive-visit date as an exact onset time can distort timing. Death without documented recurrence is a competing event, not automatically independent censoring; the target question may require a competing-risks model or a clearly defined composite outcome. CEA can also be affected by nonmalignant conditions and other factors, while some recurrences do not produce elevated CEA. A model that ignores these processes can yield misleading estimates.

• Surveillance schedule: If patients with rising CEA receive more frequent tests, measurement times may be informative. Model or sensitivity-analyze the observation process when that dependence is consequential.
• Treatment interpretation: A treatment coefficient is conditional association unless the design and assumptions justify causal interpretation. Time-varying confounding may require methods beyond simply adding a treatment indicator.
• Scale and detection limits: The log transformation requires positive measurements. Values below a detection limit should be handled as censored measurements or with an appropriate measurement model, rather than replaced arbitrarily by zero before taking logs.
• Prediction uncertainty: Patient-specific predictions must account for uncertainty in fixed effects, random effects, and the event model. A point estimate of latent CEA is not a certainty about future recurrence.