Preventing Overshoot With Hold and Step-Down Rules

Why a Threshold Needs an Action Rule

A response at or above an upper threshold is informative only if the algorithm says what happens next. Continuing to increase dose after such a response can compound overshoot; reducing dose without specifying how long to wait or when to resume can instead leave the algorithm stalled. A useful design therefore defines a trigger, an action, a recovery condition, and a restart action in advance.

The earlier material on titration rules as state-update equations provides the general framework: each decision uses the current state and measurement to determine the next dose. Here, the key addition is a separate recovery state. While recovering, the algorithm does not use the ordinary escalation rule. This separation prevents a single below-threshold measurement from immediately undoing a safety-related reduction.

Specify the State and Timing

Let \(D_k\) be the dose administered during interval \(k\), and let \(R_k\) be the response measured at the review at the end of that interval. Let \(U\) be the upper trigger. A simple rule has two states: normal titration and recovery. In normal titration, the prespecified ordinary rule applies while \(R_k<U\). If \(R_k\ge U\), the algorithm enters recovery and sets the next dose to a reduced value. In recovery, escalation is prohibited. The dose is held at the reduced level until the restart condition is met; then ordinary titration resumes at the next decision.

For a fixed step-down amount \(\Delta_{\mathrm{down}}>0\), one possible action is \(D_{k+1}=\max(D_{\min},D_k-\Delta_{\mathrm{down}})\), where \(D_{\min}\) is the permitted minimum dose. The maximum in this expression prevents the reduction from crossing the dose floor. The recovery-state hold is \(D_{k+1}=D_k\) while the restart condition is false. If a separate minimum duration in recovery is clinically or operationally required, it must also be stated; a response threshold alone does not impose that duration.

The restart threshold \(R_{\mathrm{restart}}\) is typically below \(U\). This gap avoids switching repeatedly between escalation and reduction when the response lies near one threshold. “Restart” should not be left ambiguous: an algorithm may either permit the usual next increment at the following review, or restart with a smaller, explicitly specified increment. Those choices produce different trajectories.

The rule assumes that the measured response is relevant to the decision, that reviews occur at defined times, and that a reduced dose can eventually lower response enough to satisfy the restart condition. If response is delayed, the measurement at \(k\) may still reflect earlier doses. The action is therefore applied prospectively, and the response may remain high for one or more reviews after a step-down. This is a reason to model and trace the trajectory, not to assume that the threshold immediately controls the response.

Why a Hold Can Prevent Re-escalation

A compact way to examine the action is a first-order response model. Suppose the response approaches a dose-dependent equilibrium according to \(R_{k+1}=aR_k+(1-a)(R_0+\beta D_k)\), with \(0\le a<1\). Here, \(R_0+\beta D_k\) is the equilibrium response at dose \(D_k\), and \(a\) represents persistence from one interval to the next. Assume \(\beta>0\), a constant dose–response slope over the range considered, and no measurement error for this calculation.

After a step-down to a fixed dose \(D\), define the equilibrium \(R_{\mathrm{eq}}=R_0+\beta D\). Subtracting that equilibrium from both sides of the update equation gives the response deviation after each interval. Repeating the same update \(m\) times yields:

Because \(a^m\) approaches zero for \(0\le a<1\), the response approaches \(R_{\mathrm{eq}}\) while the dose is held. If \(R_{\mathrm{eq}}<R_{\mathrm{restart}}\) and the response starts above the restart threshold, it will cross that threshold after a finite number of intervals in this deterministic model. If \(R_{\mathrm{eq}}=R_{\mathrm{restart}}\), it approaches the threshold but generally does not reach it in a finite number of intervals. Thus, a step-down that merely makes the equilibrium equal to the restart threshold can leave a strict “restart only below” rule stuck indefinitely. The conclusion depends on this simple model; real trajectories can also be affected by delays, noise, and changes in sensitivity.

Choosing the Step-Down Size

A reduction should be large enough to change the response trajectory meaningfully, but not so large that it creates an unnecessary fall below the useful range. Under the local linear equilibrium model \(R_{\mathrm{eq}}=R_0+\beta D\), one design check is to ask what dose places the equilibrium below the restart threshold. To put the equilibrium at or below that threshold requires \(D\le(R_{\mathrm{restart}}-R_0)/\beta\). In practice, a strict margin below the threshold is preferable if restart requires an observed crossing in finite time. This calculation is only as reliable as the local slope \(\beta\) and the assumed equilibrium model.

Measurement Error and Guard Bands

When the threshold is applied to a measured response rather than a known true response, noise can cause either a delayed trigger or an unnecessary reduction. A guard band can make the trigger conservative when measurement error is bounded. Suppose \(Y_k=R_k+e_k\), with \(|e_k|\le\epsilon\), where \(Y_k\) is the observed response and \(R_k\) is the true response. Triggering when \(Y_k\ge U-\epsilon\) ensures that a true response at or above \(U\) cannot be missed under this bound: \(R_k\ge U\) implies \(Y_k\ge R_k-\epsilon\ge U-\epsilon\). However, the same rule can trigger when the true response is below \(U\).

A similarly conservative restart condition uses \(Y_k\le R_{\mathrm{restart}}-\epsilon\). Since \(R_k=Y_k-e_k\le Y_k+\epsilon\), this condition ensures \(R_k\le R_{\mathrm{restart}}\). These guarantees require a genuine deterministic error bound. If error is probabilistic rather than bounded, a guard band does not provide a certainty guarantee; its performance must be evaluated under an appropriate error model.

Diagnostics, Alternatives, and Limitations

Trace both the dose and response at every decision point, and record the state of the algorithm. A rule may be going wrong if it continues to increase dose after entering recovery, repeatedly switches states near a threshold, remains in recovery despite a sustained dose reduction, or permits restart at a dose whose predicted equilibrium exceeds the upper trigger. Check that the reduction respects \(D_{\min}\), that the restart threshold is ordered below the upper trigger, and that the response measurements correspond to the specified review times. If response is delayed, also check whether a high post-reduction measurement reflects earlier doses rather than failure of the reduced dose to work.

The main alternatives address different trade-offs. A simple hold at the current dose avoids further escalation but may not lower an excessive response. A step-down followed by a recovery hold actively changes the dose and prevents immediate re-escalation, but can take time to recover. A rule that continues reducing dose at every high-response review may recover more quickly, yet can overshoot downward if the response is lagged. A proportional adjustment can scale the reduction to the size of the threshold exceedance, but its behavior depends on the chosen gain and response sensitivity. These approaches should be compared by tracing their trajectories under the same assumed response dynamics, not by comparing their action labels alone.

No threshold rule guarantees that the response will never exceed \(U\). The response may cross the threshold between review times, measurement may be noisy, and a delayed response may continue rising after the dose is reduced. A hold-and-step-down rule is therefore a feedback design, not a substitute for a valid response model, appropriate monitoring times, or evaluation of plausible uncertainty. Its parameters should be stress-tested across relevant starting responses, dose sensitivities, delays, and measurement errors.