Why Titration Histories Need to Be Read Dynamically
A titration rule can be reasonable at each decision point and still produce a poor trajectory. A dose increase may move the response past its target; several increases made before the effect of earlier changes is visible may push the response well beyond the target; or successive dose reductions and increases may cause the response to alternate around it. These patterns matter because a final dose alone does not show whether a patient spent time above a safety limit, experienced avoidable response swings, or reached the target efficiently.
The earlier material on target-based titration rules, step sizes, delayed or noisy measures, and safety holds provides the ingredients for interpreting these histories. Here the focus is diagnostic: use the sequence of measured responses and dose decisions to distinguish patterns and identify plausible algorithmic causes. A trajectory can suggest a cause, but it does not prove one; pharmacologic variability, measurement error, adherence, and other mechanisms may produce similar observations.
These definitions assume that the direction of the desired dose-response relationship is known locally. If increasing dose is expected to lower the measured response, the error sign and interpretation of an increase must be reversed. They also assume that the target is meaningful for the chosen response measure and assessment time. When the algorithm uses a target band or hysteresis, use its actual decision boundaries rather than treating every small deviation from \(T\) as a reason to adjust.
A Local Feedback Explanation
A simple local model helps organize the diagnosis. Suppose \(x_k=Y_k-T\) is the response deviation from target, \(u_k=D_{k+1}-D_k\) is the dose change, and, over the range being considered, a dose change has an approximately linear effect:
Here \(a\) describes persistence of the existing response deviation from one decision to the next, and \(b\) is the local response change per unit dose change. For a response that increases with dose, assume \(b>0\). This is a simplified model: it treats the decision interval and response measure as fixed, omits noise, and does not explicitly represent a long delay.
Consider a proportional correction \(u_k=-Kx_k\), where \(K>0\) converts the current deviation into a dose change. Substitution gives the closed-loop recurrence:
Write \(c=a-bK\). Repeated substitution yields \(x_k=c^k x_0\). If \(|c|<1\), then \(|x_k|\) decreases toward zero in this idealized model. If \(0\leq c<1\), deviations shrink without changing sign. If \(-1<c<0\), the sign alternates while the magnitude shrinks: this is damped oscillation. If \(c\leq -1\), deviations alternate without shrinking, or grow in magnitude. If \(c\geq1\), a same-direction deviation fails to shrink or grows. These conditions explain why an overly strong correction relative to response persistence can cause alternating excursions; they are not a substitute for checking the actual algorithm, delays, constraints, and data.
Worked Example: A Coarse Dose Step Crosses the Target
Suppose, hypothetically, a patient's response is assessed after it has stabilized at each dose, with the local relationship \(Y=2D\), where \(Y\) is measured in response units and \(D\) is in dose units. The target is \(T=10\) response units. The algorithm starts at \(D=2\) and increases by 2 dose units whenever the response is below target.
At \(D=2\), the response is \(Y=2(2)=4\), which is 6 units below target. The rule raises the dose to \(D=4\); the response becomes \(Y=2(4)=8\), still 2 units below target. A further 2-unit increase gives \(D=6\), for which \(Y=2(6)=12\). The response has crossed the target by 2 units.
The trajectory is \(4, 8, 12\) response units, with doses \(2, 4, 6\). In this hypothetical setting, the crossing is consistent with a dose step that is coarse relative to the local response slope: each 2-unit dose change shifts response by \(2(2)=4\) units. It does not establish that the step size is the only cause in a real patient; delayed response, dose rounding, and measurement variation would also need consideration.
Worked Example: Alternating Corrections Under Strong Feedback
Consider the local model \(x_{k+1}=0.2x_k+u_k\), with response deviations \(x_k\) in response units and dose changes \(u_k\) in dose units. Thus \(a=0.2\) and \(b=1\) response unit per dose unit. Suppose the rule uses \(u_k=-1.5x_k\), so \(K=1.5\). Start at \(x_0=1\) response unit above target and \(D_0=10\) dose units.
The closed-loop multiplier is \(c=a-bK=0.2-1.5=-1.3\). The first dose change is \(u_0=-1.5(1)=-1.5\), so \(D_1=10-1.5=8.5\). The next response deviation is \(x_1=0.2(1)+(-1.5)=-1.3\) response units: it has crossed below target.
Now \(u_1=-1.5(-1.3)=1.95\), so \(D_2=8.5+1.95=10.45\). The next deviation is \(x_2=0.2(-1.3)+1.95=1.69\) response units. The response is above target again. In this model, \(x_0=1\), \(x_1=-1.3\), and \(x_2=1.69\): deviations alternate and grow because \(|c|=1.3>1\). This is an illustration of an overly strong correction under the stated model, not a recommendation to use a continuous proportional rule in practice.
Distinguishing Delay From an Oversized Correction
A delayed response can make a moderate rule behave as though it were too aggressive. A key diagnostic question is whether the algorithm adjusts again before the expected effect of the previous adjustment can be observed. Compare adjustment times with the known or assumed response time course, and inspect whether response continues to move in the same direction after a dose change. A rise in response after a dose reduction, for example, may reflect an earlier increase whose effect is still emerging rather than failure of the reduction.
Worked Example: Several Increases Before the Response Catches Up
Suppose, hypothetically, target response is 10 units and the response follows \(Y_{k+1}=0.5Y_k+D_k\), where \(D_k\) is the dose applied during interval \(k\). This is equivalent to moving halfway from the current response toward the dose-specific equilibrium response \(2D_k\). The algorithm increases the next dose by 1 dose unit whenever the observed response is below 10, and decreases it by 1 when above 10. Begin with \(Y_0=6\); the first applied dose is \(D_0=4\).
The first response is \(Y_1=0.5(6)+4=7\), so the next dose is set to \(D_1=5\). Then \(Y_2=0.5(7)+5=8.5\), so \(D_2=6\). The next response is \(Y_3=0.5(8.5)+6=10.25\), which is above target; the next dose is reduced to \(D_3=5\). Despite that reduction, \(Y_4=0.5(10.25)+5=10.125\), still above target. The rule reduces again to \(D_4=4\), after which \(Y_5=0.5(10.125)+4=9.0625\).
The doses \(4,5,6,5,4\) and responses \(6,7,8.5,10.25,10.125,9.0625\) show two linked features: successive increases were made while the response was still catching up, and the first decrease did not immediately bring the response below target. This model makes the lag explicit. In practice, assess whether the observation schedule, dose-adjustment interval, and response delay support this explanation before changing the rule.
Reading Histories and Diagnosing Failure Modes
Plot response against time with the target and any target-band boundaries visible. On the same time axis, mark every dose change and the response measurement used to make that decision. A compact table is also useful: record the measurement time, dose in effect, response value, target error, adjustment, and time since the preceding adjustment. This makes it easier to see whether an apparent swing follows an adjustment, precedes it, or persists after it.
- Overshoot after a single large change: check whether the local dose-response slope, dose rounding, or step size makes the expected response shift large relative to the distance from target. Also check whether the response was already moving before the adjustment.
- Repeated increases before a response change is visible: compare assessment and adjustment timing with the expected response delay. The dose history may be accumulating changes faster than the response can reveal their effects.
- Alternating corrections near the target: check the correction size, decision thresholds, and whether a target band or hysteresis is implemented as intended. If changes are similar in size to measurement noise, the apparent alternation may be noise-driven rather than true instability.
- Unequal or unexplained excursions: check for missed doses, changing adherence, safety holds, dose-rounding constraints, accumulation, and tolerance. These can make a simple local feedback explanation inadequate.
Do not infer a stable algorithmic pattern from a short history containing only one crossing. Conversely, a favorable final response does not erase an earlier clinically important excursion. Review the duration and magnitude of time outside relevant limits, the sequence of decisions, and the plausibility of the assumed dose-response timing. If noisy measurements are suspected, distinguish genuine response movement from measurement variability; increasing step size to “correct” random fluctuations can worsen instability.
What to Change—and What the Diagnosis Cannot Establish
The pattern suggests which design features deserve review, but no single adjustment is universally appropriate. A smaller step may reduce crossing when the response is measured promptly and reliably. A longer interval between decisions may allow more of a delayed effect to appear, but may also prolong time below target. A target band or hysteresis can prevent repeated adjustments in response to small deviations near the target. More reliable or appropriately timed measurements can help distinguish a real trend from noise. Any modification must be considered alongside safety constraints, response delay, and the clinical consequences of waiting or changing dose.
| Approach | When it may help | Important limitation |
|---|---|---|
| Reduce the adjustment step | Crossings appear large relative to the local response change per dose step. | May slow approach to target; cannot alone resolve an unrecognized delay. |
| Allow more time before reassessment | Response is still evolving when the next decision is made. | May delay needed correction and does not remove measurement noise. |
| Use target bands or hysteresis | Small deviations trigger repeated reversals. | Does not guarantee stability when delays or large changes dominate. |
| Improve measurement timing or reliability | Measurements are noisy or poorly aligned with the expected response. | Better measurement cannot compensate for an unsuitable adjustment rule. |
The local recurrence is a diagnostic simplification, not a complete patient model. Its linearity and fixed timing may fail across a wide dose range; its assumptions may fail when response is delayed, nonlinear, noisy, or affected by accumulation or tolerance. A trajectory alone cannot identify the true cause of an excursion, nor establish that a proposed change will improve clinical outcomes. Use the history to formulate and check explanations against pharmacology, measurements, and the actual rule.
Check Your Understanding
Answer using the definitions and simplified models above. Solutions follow the questions.
- In the local model \(x_{k+1}=0.4x_k+u_k\), a rule uses \(u_k=-0.6x_k\). Find the closed-loop multiplier and state whether deviations shrink monotonically or alternate while shrinking.
- A patient is below target at two consecutive measurements, and the dose is increased after each. The response rises above target at the next measurement. Name one plausible algorithmic explanation and one additional fact needed to assess it.
- Why is one response crossing insufficient evidence of oscillation?
- Give one reason repeated dose increases followed by a delayed overshoot might not be caused solely by a large step size.
Solutions. 1. The multiplier is \(c=0.4-0.6=-0.2\). Since \(|c|<1\) and \(c<0\), deviations alternate in sign while shrinking. 2. A plausible explanation is that the second increase was made before the first increase's effect was fully visible. The response delay or time from each dose change to measurement is needed to assess that explanation. 3. Oscillation requires repeated alternation; one crossing could be a single overshoot followed by stable correction. 4. The response may be delayed, so several changes can accumulate before their effects appear; adherence, measurement error, or a changing response can also contribute.