Reading an Unstable Trajectory
A titration trajectory records how dose and response change across successive adjustment occasions. An unstable pattern may appear as large alternating dose changes, continued adjustment after an effective dose has been reached, or apparently erratic responses. These patterns can look similar in a plot, but they do not imply the same cause. A large update can destabilize an otherwise prompt, consistent feedback loop; delayed response information can make an otherwise reasonable update act on an outdated state; and inconsistent measurements can make the algorithm chase variation that is not a reliable change in the underlying response.
The aim is not to diagnose a cause from one surprising observation. First establish what response each measurement represents, when it was obtained relative to dose changes, and how the titration rule used it. Then compare observed changes with the trajectory expected under the rule and a plausible local dose-response relationship. The examples use hypothetical data and simplified models to make those checks explicit.
For a useful first diagnostic, suppose the response is locally linear and increases with dose: \(y=a+bd\), where \(d\) is dose, \(y\) is response, and \(b>0\) is the local response change per unit dose. Suppose the target is \(T\), the measurement reflects the current dose without delay, and the rule is \(d_{k+1}=d_k+K(T-y_k)\), where \(K>0\) is the adjustment gain in dose units per response unit. This simplified setup assumes a stable response relationship, correctly timed and comparable measurements, and no other dose changes between assessments. These assumptions need checking before interpreting the pattern.
Why an Excessive Update Can Alternate or Diverge
Let \(d^*\) be the dose that gives the target under the local model, so \(a+bd^*=T\). Define response error \(x_k=y_k-T\). With an immediate measurement and the rule above, the next response error is determined by the next dose:
Thus each update multiplies the previous error by \(1-bK\). Convergence requires the absolute value of this multiplier to be less than one. For \(b>0\) and \(K>0\), the condition is \(0<bK<2\). If \(0<bK\leq1\), the error moves toward zero without changing sign (with \(bK=1\), the target is reached in one update in this idealized model). If \(1<bK<2\), the error changes sign at each update but shrinks: the trajectory oscillates while converging. At \(bK=2\), the error alternates without shrinking; if \(bK>2\), the alternating errors grow. This is a local result, not a guarantee for a nonlinear response, bounded dose rule, or delayed system.
Worked Example: Alternating Divergence From a Large Update
Suppose, hypothetically, response follows \(y=2+0.8d\), with dose \(d\) in mg and response in response units. The target is 10 response units and the adjustment rule uses \(K=3\) mg per response unit. Assume each observation reflects the current dose, the slope is locally constant, and no dose bounds intervene. The target dose is \(d^*=(10-2)/0.8=10\) mg, and the dimensionless gain is \(bK=0.8(3)=2.4\), above the stability limit of 2.
Starting at \(d_0=6\) mg, the response is \(y_0=2+0.8(6)=6.8\). The update is \(d_1=6+3(10-6.8)=15.6\) mg. The next response is \(y_1=2+0.8(15.6)=14.48\). The next update is \(d_2=15.6+3(10-14.48)=2.16\) mg, giving \(y_2=2+0.8(2.16)=3.728\). A further update gives \(d_3=2.16+3(10-3.728)=20.976\) mg and \(y_3=2+0.8(20.976)=18.7808\).
The dose and response cross the target repeatedly, and the deviations grow. The multiplier is \(1-bK=1-2.4=-1.4\), so each response error changes sign and grows by a factor of 1.4 in this model. This pattern supports excessive update size only if the measurements are prompt and reasonably consistent and the local slope estimate is credible. In practice, such large doses might be prevented by bounds; bounds would alter the trajectory but would not, by themselves, explain why the updates repeatedly cross the target.
Separating Delayed Feedback From Large Updates
A trajectory can keep moving in the wrong direction after a dose change because the measured response has not yet caught up with that change. The critical diagnostic is the time link between each dose and each response: record when the dose was changed, the assessment time, and which dose history could have produced the observed response. The earlier material on adjustment intervals for pharmacokinetic equilibration and response-guided titration with delayed measurements provides the relevant timing concepts. Here the focus is on recognizing their signature in a trajectory.
With a delay, a small or moderate adjustment can still be followed by another adjustment in the same direction because the next observation remains low or high under the previous dose. Once the response to the higher or lower dose becomes visible, the algorithm may then reverse direction. This creates overshoot or oscillation that can be mistaken for an excessively large update. Evidence for delay includes a consistent lag between dose changes and response changes, and continued response movement after the dose has been held constant. Evidence for excessive gain is stronger when measurements are correctly aligned to the current dose and prompt changes repeatedly overshoot in proportion to the update.
Worked Example: A One-Interval Response Delay
Consider a hypothetical response relationship \(y=2+0.8d\) with target 10 response units. Use \(K=1.25\) mg per response unit, so \(bK=1\); with immediate feedback, one update would reach the target in this idealized linear model. Suppose instead that response \(y_k\) reflects the preceding dose \(d_{k-1}\): \(y_k=2+0.8d_{k-1}\). The rule still updates from the latest observed response, \(d_{k+1}=d_k+1.25(10-y_k)\). Set \(d_{-1}=d_0=6\) mg so the initial observation is defined.
The initial observation is \(y_0=2+0.8(6)=6.8\), so \(d_1=6+1.25(10-6.8)=10\) mg. But \(y_1\) still reflects \(d_0\), so \(y_1=6.8\) and the rule raises the dose again: \(d_2=10+1.25(10-6.8)=14\) mg. Now \(y_2\) reflects \(d_1=10\), giving \(y_2=10\), so \(d_3=14\) mg. Then \(y_3\) reflects \(d_2=14\), giving \(y_3=13.2\), and \(d_4=14+1.25(10-13.2)=10\) mg.
The response associated with 10 mg arrives after the dose has already been raised to 14 mg. The subsequent response above target prompts a reduction, even though the trajectory included a dose that produced the target under the assumed model. This is evidence of delayed feedback, not proof that delay is the only cause. Check the actual assessment timing, whether response continues evolving after a dose change, and whether other process changes could account for the lag.
Recognizing Inconsistent Response Measurements
Measurement inconsistency is suggested when observations depart irregularly from the expected dose-response pattern, without a stable lag or a repeatable relation between the size of an update and the ensuing overshoot. It can arise from measurement error, variable assessment conditions, nonadherence, changes in disease state, or genuine within-person response variation. The trajectory alone cannot establish which of these is responsible. Examine measurement timing and method, dose administration records, and repeated observations where feasible; compare residuals, the differences between observed and model-predicted responses, rather than treating every raw value as a true change in response.
A useful comparison is to ask whether the dose-response sequence is internally coherent. If the same dose under comparable conditions produces markedly different responses, inconsistent observations deserve attention. If response consistently changes only after a delay, timing is the more direct explanation. If well-timed measurements are consistent but large dose changes cross the target repeatedly, gain is a stronger candidate. More than one mechanism can operate at once.
Worked Example: Noisy Measurements and Dose Chasing
Suppose a hypothetical local model is \(y=8+0.5(d-10)\), with target 8 response units. A dose of 10 mg therefore predicts 8 units. Use \(K=0.5\) mg per response unit. Assume the dose-response relationship is prompt, but individual observations include measurement or short-term response variation. The values below are hypothetical, and the residual is observed response minus the response predicted at that dose.
| Occasion | Dose (mg) | Predicted response | Observed response | Residual |
|---|---|---|---|---|
| 0 | 10.0 | 8.00 | 6.0 | -2.00 |
| 1 | 11.0 | 8.50 | 10.0 | +1.50 |
| 2 | 10.0 | 8.00 | 7.0 | -1.00 |
| 3 | 10.5 | 8.25 | 9.0 | +0.75 |
Starting from 10 mg, the first observation gives \(d_1=10+0.5(8-6)=11\) mg. At 11 mg, the prediction is \(8+0.5(11-10)=8.5\), but the observed response is 10; therefore \(d_2=11+0.5(8-10)=10\) mg. At 10 mg, the observed response is 7, giving \(d_3=10+0.5(8-7)=10.5\) mg. At 10.5 mg, the prediction is \(8+0.5(0.5)=8.25\); the observed value of 9 gives \(d_4=10.5+0.5(8-9)=10\) mg.
The updates reverse direction with alternating residuals, while the doses remain near 10 mg. This is compatible with the algorithm reacting to inconsistent observations, rather than demonstrating a persistent instability in the underlying dose-response relationship. It does not establish that measurement noise is the cause: four observations are limited, and changing clinical state or other unrecorded factors could also produce residuals. Repeated measurements under comparable conditions, where safe and appropriate, can help assess whether observations are reproducible.
A Practical Diagnostic Sequence
Use the following checks before changing an algorithm or concluding that the response is unstable. They help distinguish mechanisms but are not a substitute for clinical judgment or an appropriately specified model.
List each dose, dose-change time, response assessment time, and response value. Confirm that each observation is assigned to the dose history it could reflect.
Review measurement method, timing, adherence, and conditions. Ask whether repeated observations at comparable doses and conditions agree sufficiently for the intended decision.
If a plausible local slope \(b\) and gain \(K\) are available, calculate \(bK\). A value above 2 predicts divergence in the immediate, linear, unbounded model; it is a warning under those assumptions, not a diagnosis by itself.
Continued response change after a dose hold supports a response lag; irregular residuals support variable observations; repeatable, prompt overshoot following large adjustments supports excessive gain.
Depending on the evidence, consider a smaller update, an assessment interval that allows response to emerge, or improved measurement and confirmation. Changing all three at once makes it harder to learn which mechanism mattered.
Alternatives, Limitations, and a Common Pitfall
Trajectory inspection is a first-pass diagnostic. A model-based analysis can estimate local sensitivity and quantify how the observed response deviates from predictions, but its conclusions depend on the model and data. A controlled hold or repeated assessment can help distinguish delayed change from inconsistent measurement, but may not be suitable when delaying treatment is unsafe. A conservative update can reduce overshoot risk, but it does not correct a mistimed measurement. These approaches answer different questions and can be combined when appropriate.
The simple stability calculation assumes a monotone local linear relationship, immediate feedback, consistent response units and timing, a fixed gain, and no dose limits. Nonlinear dose-response curves can have different local sensitivities at different doses. Dose bounds and rounding alter the update sequence. Changing disease state, adherence, concomitant treatment, and time-varying response can all make a previously appropriate model inaccurate. Thus an observed oscillation does not uniquely identify its cause, and a stable-looking trajectory does not prove that the target is appropriate.
- Pitfall: “The dose alternates around the target, so the update must be too large.” Alternation can also result from delayed feedback or inconsistent measurements.
- Correction: Align each response to its relevant dose history, inspect measurement consistency, and then assess whether the size of prompt updates plausibly explains the overshoot.
- Pitfall: “Reducing the gain will fix any unstable trajectory.” A smaller update may damp oscillation but can leave a timing problem or measurement problem unresolved.
- Correction: Use the diagnosed mechanism to guide the intervention, and reassess the trajectory after a change rather than assuming the cause is settled.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- Under the immediate linear model, \(b=0.4\) response units per mg and \(K=2\) mg per response unit. Calculate \(bK\) and describe the predicted error pattern.
- A dose is raised twice because two consecutive response measurements remain below target. The second measurement was taken before the response to the first dose change was expected to emerge. Which cause should be investigated first, and what should be checked?
- At comparable doses and assessment times, response residuals repeatedly alternate in sign and are large relative to the expected measurement variation. Does this alone prove measurement error?
- Why is reducing \(K\) not a complete remedy if the response measurement reflects an earlier dose?
Solutions. (1) \(bK=0.4(2)=0.8\). Since \(0<bK<1\), the immediate linear model predicts errors that shrink without changing sign. (2) Investigate delayed feedback first; reconstruct dose and assessment times and determine which dose history the measurement reflects. (3) No. The pattern flags inconsistency with the assumed response variation, but changing clinical state, adherence, or a misspecified model could also explain it. (4) A smaller gain may reduce the size of updates, but the algorithm would still use outdated information; the response-to-dose timing must also be assessed.