Worked Example: Symptom-Triggered Biomarker Monitoring

Why symptom-triggered monitoring changes the data

In a follow-up study, biomarker measurements may be collected when patients report worsening symptoms rather than at fixed times. The resulting records are not simply a less regular version of a scheduled sample: the chance of observing a biomarker can depend on the patient’s evolving state. If symptom severity is associated with the biomarker, some biomarker states will consequently contribute more records than others.

The focus here is a hypothetical follow-up study in which patients can request an assessment when symptoms worsen. The key tasks are to specify the visit mechanism precisely and identify the sampling distortion it can create. The visit intensity describes when records are generated; it does not, by itself, establish that symptoms cause biomarker changes or that the biomarker causes symptoms.

Definition: For patient $i$, let $N_i(t)$ count visits by time $t$, let $S_i(t)$ denote symptom severity, and let $Y_i(t)$ denote the underlying biomarker level. The visit intensity $\lambda_i^V(t\mid\mathcal H_i(t))$ is the instantaneous rate of a visit at time $t$, conditional on the observed history $\mathcal H_i(t)$. In a short interval of length $dt$, the conditional probability of one visit is approximately $\lambda_i^V(t\mid\mathcal H_i(t))dt$; the probability of two or more visits is of smaller order than $dt$.

An intensity is a rate, so its units here are visits per month. To be a valid conditional intensity, it must be nonnegative and defined relative to the information available just before time $t$. A symptom report used to prompt a visit belongs in that available history. A biomarker result obtained only after the visit should not be inserted as though it were known to the clinician beforehand.

Specifying a symptom-triggered mechanism

Suppose a hypothetical study allows patient-initiated assessments, and the clinical team also contacts patients whose reported symptoms become severe. A simple working model for the visit intensity is

\lambda_i^V(t\mid\mathcal H_i(t)) =\lambda_0\exp\{\gamma S_i(t-)\},

where $\lambda_0>0$ is the rate per month at symptom score zero, $\gamma$ describes how the rate changes with symptom severity, and $S_i(t-)$ is the most recently available symptom score before $t$. If $\gamma>0$, higher reported symptom scores are associated with more frequent visits. This model is a description of the monitoring process, not a model for the biomarker trajectory.

In practice, the mechanism may instead include an explicit symptom threshold, a scheduled component, treatment status, or a patient-specific tendency to seek care. The formula must match the actual protocol and available history. A useful specification states who can initiate a visit, what information prompts it, and how the intensity changes with that information.

Worked Example: Turning Symptom Scores Into Visit Rates

Assume the hypothetical study uses $\lambda_i^V(t)=0.20\exp\{0.60S_i(t-)\}$ visits per month. For illustration, suppose each patient’s score stays constant for six months. Calculate the rates and expected visit counts for a patient at score 0 and a patient at score 2.

For score 0, the intensity is $0.20\exp(0.60\times0)=0.20$ visits per month. Over six months with that constant intensity, the expected count is $6\times0.20=1.20$ visits.

For score 2, the intensity is $0.20\exp(0.60\times2)=0.20\exp(1.20)\approx0.6640$ visits per month. The expected count over six months is $6\times0.6640=3.984$ visits.

Thus the model implies about 3.32 times the visit rate at score 2 as at score 0, since $0.6640/0.20=3.32$. These are expected counts under the stated constant-score assumption, not guaranteed numbers of visits for either patient.

From visit intensity to sampling distortion

A visit-level average gives more influence to states that generate more visits. The simplest calculation makes this visible. Suppose a hypothetical population spends equal amounts of observation time in two biomarker states: a low state with $Y=5$ units and a high state with $Y=11$ units. Suppose visits occur at respective rates of $0.5$ and $2$ per month, and the process is stable over time.

Worked Example: A Visit-Level Mean Exceeds the Time-Based Mean

Calculate the mean biomarker level across time and the expected mean among visit records under the stated two-state model.

The time-based mean gives each state its 50% time share: $0.50(5)+0.50(11)=2.5+5.5=8.00$ units.

For a randomly selected visit record, the high state contributes in proportion to its time share multiplied by its visit rate. The unnormalized visit weights are $0.50(0.5)=0.25$ for the low state and $0.50(2)=1.00$ for the high state. Their sum is $1.25$. Thus 80% of visits are expected to occur in the high state, because $1.00/1.25=0.80$, and 20% in the low state. The visit-level mean is $0.20(5)+0.80(11)=1.00+8.80=9.80$ units.

The visit-level mean is 1.80 units above the time-based mean. The difference arises because the high state produces four times as many visits per unit time. It is a sampling distortion if the target is the mean biomarker level over time; it is not a calculation error if the target is specifically the mean at a randomly selected visit.

The same logic applies when symptom severity, rather than the biomarker itself, drives monitoring. If symptoms and biomarker values are associated, then weighting records by their visit rates changes the biomarker distribution observed at visits. The distortion can occur even when clinicians never see the biomarker before deciding to arrange an assessment.

Result: For a stable process sampled through visits, if the state $Z$ has time-based distribution $f_Z(z)$ and visits occur at rate $\lambda^V(z)$, the distribution of states among visit records is proportional to $\lambda^V(z)f_Z(z)$. In particular, when $Z$ includes the biomarker and symptoms, the visit-level mean of $Y$ is $$ E_{\mathrm{visit}}(Y) =\frac{E\{Y\lambda^V(S)\}}{E\{\lambda^V(S)\}}, $$ provided the expectations are finite and the intensity is determined by the state shown. If intensity also depends on other history, that history must be included in the weighting expression.

Justification: why visit records are rate-weighted

Consider a small interval of length $dt$. Conditional on a state $Z=z$, the probability of a visit in that interval is approximately $\lambda^V(z)dt$. The probability that the state is $z$ and a visit occurs is therefore proportional to $f_Z(z)\lambda^V(z)dt$. To obtain the state distribution conditional on a visit, normalize these quantities across all states. The common factor $dt$ cancels, giving the rate-weighted distribution.

f_{\mathrm{visit}}(z) =\frac{\lambda^V(z)f_Z(z)}{\int \lambda^V(u)f_Z(u)\,du}.

Multiplying this density by $Y$ and integrating gives the visit-level expectation: $\int y f_{\mathrm{visit}}(z)\,dz=E\{Y\lambda^V(S)\}/E\{\lambda^V(S)\}$. This argument assumes the state distribution and visit mechanism are stable over the period being pooled, and that the visit intensity is correctly represented by the state and history included in the formula. With changing states over calendar time, the corresponding calculation must account for how those states and rates vary over time rather than treating the whole study as stationary.

Worked Example: Symptoms Can Create Biomarker Size Bias

Suppose, hypothetically, that over the target period 40% of patient-time is in a severe-symptom state with biomarker level 12 units, and 60% is in a mild-symptom state with biomarker level 6 units. The visit rates are 3 and 1 per month, respectively. Calculate the time-based mean and the expected visit-level mean.

The time-based mean is $0.40(12)+0.60(6)=4.80+3.60=8.40$ units.

The rate-weighted contributions to visits are $0.40(3)=1.20$ for the severe state and $0.60(1)=0.60$ for the mild state. The total is $1.80$, so the severe state contributes $1.20/1.80=0.6667$ of visit records and the mild state contributes $0.3333$. The visit-level mean is \[ \frac{0.40(3)(12)+0.60(1)(6)}{0.40(3)+0.60(1)} =\frac{14.40+3.60}{1.80} =10.00\text{ units}. \]

The visit-level mean is 1.60 units higher than the time-based mean. In this example, symptom-triggered monitoring overrepresents the severe state, which is also the state with the higher biomarker. The calculation does not show that symptoms caused the biomarker level; it shows how their association and the monitoring rule affect which states are recorded.

Checks, alternatives, and limitations

A mechanism is more credible when its implied pattern can be checked against the actual follow-up process. Useful checks include comparing observed visit rates across symptom histories, checking whether recorded triggers precede rather than follow the visit, and examining whether predicted and observed visit counts agree across follow-up intervals. The earlier material on time-rescaling diagnostics and simulation checks for informative observation mechanisms provides tools for assessing an intensity model; here the practical question is whether the specified mechanism represents this symptom-triggered study.

Failure to record symptom reports between biomarker visits is an important limitation. If a patient’s worsening symptoms prompted contact but the trigger was not documented, the observed history may not explain the visit intensity. Patient-specific access to care or unrecorded clinical concerns may also affect visit rates. A model using only the most recent recorded symptom score can then misrepresent the actual process. Sparse observations make it especially difficult to distinguish a genuine lack of visit dependence from insufficient information about the triggers.

The main alternative for the monitoring process is to model visits as scheduled or otherwise independent of the evolving state, but that description is not appropriate merely because it simplifies analysis. A scheduled component can coexist with symptom-triggered visits; the protocol determines which mechanism applies. Another approach is to model the visit process jointly with the longitudinal outcome process, allowing both to depend on shared latent features. Neither approach removes the need to define the target: a mean over patient-time and a mean over visit records answer different questions.

When the target is a time-based summary, rate weighting can in principle correct the overrepresentation shown above if the relevant intensity is adequately measured and modeled. Such correction is not automatic: near-zero visit rates in some states, inaccurate trigger records, or misspecified dependence can make estimates unstable or biased. A visit-level average may be entirely suitable when the intended target is the biomarker distribution at clinical encounters. The mistake is to interpret it as a time-based population summary without checking how visits were generated.

A common pitfall

Wrong reasoning: “The biomarker was measured at every recorded visit, so averaging the measurements estimates the average biomarker over follow-up.”
Correction: The average across visits gives more weight to states with higher visit intensity. It estimates a visit-level quantity; interpreting it as an average over time requires assumptions about the visit mechanism and, when needed, an appropriate correction.
Wrong reasoning: “The biomarker must have triggered visits because its observed mean is high.”
Correction: A high visit-level mean can arise because symptoms trigger visits and symptoms are associated with the biomarker. The monitoring process and the biomarker process must be distinguished.

Key takeaway: State the clinical trigger, encode it in a conditional visit intensity using information available before the visit, and identify whether the intended summary is over time or over visits. When visit rates vary across symptom-associated biomarker states, the raw visit-level sample is rate-weighted.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A hypothetical clinic uses $\lambda^V(t)=0.15\exp\{0.4S(t-)\}$ visits per month. What is the intensity at symptom score 1, to three significant figures?
Two equally common states have biomarker values 4 and 10 units and visit rates 1 and 3 per month, respectively. Find the time-based mean and the visit-level mean.
Why can symptoms produce a distorted visit-level biomarker mean even if the clinician does not observe the biomarker before arranging a visit?
What target does an unweighted mean of all recorded visit measurements directly summarize?

Solutions. (1) The intensity is $0.15\exp(0.4\times1)=0.15\exp(0.4)\approx0.224$ visits per month. (2) The time-based mean is $0.50(4)+0.50(10)=7.00$ units. The visit weights are $0.50(1)=0.50$ and $0.50(3)=1.50$, so the visit-level mean is $\{0.50(4)+1.50(10)\}/2.00=8.50$ units. (3) If symptoms are associated with biomarker values, symptom-dependent visit rates make some biomarker-associated states contribute more records, even when the biomarker itself is not a trigger. (4) It directly summarizes biomarker values among visit records, not necessarily among patients or across patient-time.