Why a Target Band Needs More Than One Threshold
A titration rule may change dose when a response measure falls below or rises above a threshold. But if the measure fluctuates around that threshold, consecutive assessments can call for opposite actions. Repeatedly increasing and decreasing dose in response to small changes can make the regimen difficult to implement and may move the patient away from a useful dose. A target band and a hysteresis rule make explicit when dose should be left alone and how the current titration direction affects the next decision.
The target band should represent response values considered acceptable for the decision at hand. Its endpoints are not merely convenient cutoffs: they should be justified by the clinical objective and the dose-response evidence used to construct the titration rule. Hysteresis then adds state-dependent thresholds around the actions. It can reduce needless switching, but it does not establish that the target is clinically appropriate or make a delayed or unreliable response measure reliable.
For clarity, suppose larger values of a response measure \(y\) indicate a stronger response. A simple target-band rule increases dose below \(L\), decreases dose above \(U\), and makes no change within the band. A hysteresis rule adds memory: while escalation is under way, escalation may continue until response reaches a release threshold above \(L\). While de-escalation is under way, dose reduction may continue until response reaches a release threshold below \(U\).
This tutorial assumes that the response measure, assessment time, target band, and dose-change eligibility are specified in advance. Assessments must be comparable, and the direction of response must be clear. If lower values are preferable, the rule’s direction must be reversed or the measure re-expressed so that the rule’s interpretation is consistent. Existing dose-step and adjustment-interval decisions still constrain what action can be taken at each eligible assessment.
Write the Rule as a State-Dependent Decision
Let the current titration state be \(q\in\{H,E,D\}\), where \(H\) means hold, \(E\) means escalation is active, and \(D\) means de-escalation is active. Let \(y\) be the response at an eligible assessment. Use four thresholds with the ordering
When the state is \(H\), response below \(L_{\mathrm{on}}\) starts escalation, and response above \(U_{\mathrm{on}}\) starts de-escalation. Values between those thresholds lead to a hold. In state \(E\), escalation continues while response is below \(L_{\mathrm{off}}\), then ends; an extreme response above \(U_{\mathrm{on}}\) instead initiates de-escalation. In state \(D\), de-escalation continues while response is above \(U_{\mathrm{off}}\), then ends; an extreme response below \(L_{\mathrm{on}}\) instead initiates escalation. These rules treat equality at a release threshold as release and equality at an onset threshold as no trigger. A protocol can choose different boundary conventions, but it must state them unambiguously.
| Current state | Response condition | Decision at this assessment |
|---|---|---|
| Hold \(H\) | \(y<L_{\mathrm{on}}\) | Start escalation |
| Hold \(H\) | \(L_{\mathrm{on}}\leq y\leq U_{\mathrm{on}}\) | Hold |
| Hold \(H\) | \(y>U_{\mathrm{on}}\) | Start de-escalation |
| Escalation \(E\) | \(y<L_{\mathrm{off}}\) | Continue escalation |
| Escalation \(E\) | \(L_{\mathrm{off}}\leq y\leq U_{\mathrm{on}}\) | Stop escalation; hold |
| Escalation \(E\) | \(y>U_{\mathrm{on}}\) | Switch to de-escalation |
| De-escalation \(D\) | \(y>U_{\mathrm{off}}\) | Continue de-escalation |
| De-escalation \(D\) | \(L_{\mathrm{on}}\leq y\leq U_{\mathrm{off}}\) | Stop de-escalation; hold |
| De-escalation \(D\) | \(y<L_{\mathrm{on}}\) | Switch to escalation |
A protocol should also say whether an action is one dose step per assessment or continued stepwise movement at subsequent eligible assessments. The thresholds govern direction and continuation; they do not, by themselves, specify dose increments, assessment timing, or any clinical restrictions on a dose.
Worked Example: Defining a Band and Directional Thresholds
Suppose a hypothetical response measure is considered acceptable from 40 to 60 units. Use \(L_{\mathrm{on}}=40\), \(L_{\mathrm{off}}=45\), \(U_{\mathrm{off}}=55\), and \(U_{\mathrm{on}}=60\). These thresholds satisfy \(40<45\leq55<60\). Starting in state \(H\), an assessment of 38 units starts escalation because \(38<40\). At the next eligible assessment, 42 units is still below the escalation release threshold: \(42<45\), so escalation continues. At 46 units, \(46\geq45\), so escalation stops and the state becomes \(H\). A subsequent value of 39 units starts escalation again because the state is now \(H\) and \(39<40\).
The example shows why the state must be recorded: 42 units leads to continued escalation if the current state is \(E\), but would lead to a hold if the current state were \(H\). The values 40 and 60 define the onset boundaries in the hold state; 45 and 55 are release boundaries for an action already under way.
Why the Gap Can Reduce Switching
The basic justification is geometric: a measurement must traverse a nonzero gap to change an action state. Consider a single lower threshold \(c=50\) units used to classify readings as below or above the threshold. If the true response is 50 units and measurement error is symmetric, repeated readings can fall on either side. A hysteresis illustration instead starts an action below 47 units and releases it above 53 units. Moving from an initiating reading to a reading on the other side of the release boundary requires the observed value to span 6 units.
For a numerical illustration, assume measurements at successive assessments are independent and normally distributed with mean 50 units and standard deviation 2 units. Write \(\Phi(z)\) for the cumulative distribution function of the standard normal distribution. The probability that one reading is below 47 is
The probability that an independent next reading is above 53 is also \(\Phi(-1.5)\approx0.0668\). Thus, the probability of that particular two-reading sequence is
This calculation is a limited illustration, not a general estimate of switching frequency: it assumes a fixed true response, normal errors, and independent assessments. It shows why separated thresholds can make a direct noise-driven crossing less likely than decisions based on a single boundary. If errors are correlated, the independence calculation does not apply. Hysteresis also does not prevent a switch when the response genuinely changes enough to cross the relevant thresholds.
Worked Example: Checking a Measurement-Error Calculation
In a hypothetical setting, suppose the true response is 52 units, measurement error is normally distributed with standard deviation 3 units, and escalation begins when the observed response is below 44 units. The standardized threshold is
Therefore, the probability of triggering escalation at one assessment is \(\Phi(-2.667)\approx0.00383\), or about 0.383%, under these assumptions. This is a calculation for the specified measurement model, not a probability that escalation is clinically unnecessary. A different true response, error distribution, or correlation structure would give a different probability. The result can help compare candidate thresholds, but clinical acceptability still depends on the consequences of the resulting decisions.
Design, Diagnostics, and Alternatives
Design the band and gaps in a sequence that keeps the clinical objective ahead of the control rule. First specify what response values are acceptable and why. Then specify which response values should initiate each direction of dose change. Finally choose release thresholds and verify that their ordering produces the intended behavior. The release thresholds should not be set solely to make switching rare: gaps that are too wide can prolong escalation or de-escalation after the response has entered an acceptable region.
Before implementation, inspect the rule at every boundary and over plausible sequences of observations. Check whether equality triggers or releases an action as intended; whether a state can persist indefinitely; whether a transition can unexpectedly reverse direction; and whether the sequence of dose actions remains compatible with the previously specified dose steps and assessment intervals. Also verify that all thresholds use the same response units and are valid for the same assessment definition. If a measure is transformed or reported in different units, transform the thresholds consistently.
A single-threshold rule is simpler and has no state-dependent memory, but it can be more sensitive to small fluctuations around its cutoff. Requiring repeated measurements beyond a threshold before acting is another way to avoid responding to one unusual observation; it adds a confirmation requirement rather than a release boundary and can delay a decision. These approaches are not interchangeable: the choice depends on how quickly action is needed and on the consequences of an unnecessary change versus a delayed one. Hysteresis alone does not address an inappropriate target, uncertain response direction, or a measurement process that does not support the rule’s assumptions.
Use a small decision table or explicit state-transition rule in the protocol rather than relying on phrases such as “titrate toward target.” That makes it possible to check what happens after each observation and to distinguish deliberate persistence from accidental oscillation. Thresholds should be reviewed when the response definition, assessment schedule, or evidence supporting the target changes.
Worked Example: Applying the Rule to a Response Sequence
Use \(L_{\mathrm{on}}=40\), \(L_{\mathrm{off}}=45\), \(U_{\mathrm{off}}=55\), and \(U_{\mathrm{on}}=60\), starting in state \(H\). The hypothetical sequence of eligible response measurements is 63, 58, 54, 57, and 39 units.
At 63 units, \(63>60\), so start de-escalation and enter state \(D\). At 58 units, \(58>55\), so continue de-escalation. At 54 units, \(54\leq55\) and \(54\geq40\), so stop de-escalation and enter \(H\). At 57 units, the response is between 40 and 60, so hold. At 39 units, \(39<40\), so start escalation. The state sequence after the assessments is \(D,D,H,H,E\). In particular, 57 units does not restart de-escalation: the rule starts that action only above 60 while in state \(H\).
This sequence checks direction, release, and re-entry separately. It does not specify dose amounts; each action must still be paired with the dose step and timing defined for the regimen.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For thresholds \(L_{\mathrm{on}}=30\), \(L_{\mathrm{off}}=34\), \(U_{\mathrm{off}}=46\), and \(U_{\mathrm{on}}=50\), a patient is in state \(E\) and has a response of 33 units. What is the decision?
- With the same thresholds, what happens at a response of exactly 34 units while in state \(E\)?
- Suppose \(Y\sim N(48,\,3^2)\), and escalation is triggered below 42 units. Calculate the probability of that trigger at one assessment.
- Give one reason a hysteresis gap should not be made arbitrarily wide.
- Why can the response of 47 units lead to different decisions in states \(H\) and \(D\), using the thresholds in Question 1?
Solutions. (1) Since \(33<34\), escalation continues. (2) Equality is at the release threshold, so escalation stops and the state becomes \(H\). (3) Standardize: \(z=(42-48)/3=-2\). Thus \(P(Y<42)=\Phi(-2)\approx0.0228\), or 2.28%, under the stated normal model. (4) A very wide gap can prolong dose escalation or reduction after response has reached an acceptable region, delaying the intended adjustment. (5) In state \(H\), 47 is between 30 and 50, so the rule holds. In state \(D\), 47 is above \(U_{\mathrm{off}}=46\), so de-escalation continues. The difference is the state-dependent release rule.