Constructing stopping rule for safety monitoring

In many early phase clinical trials, if there is a known risk or theoretically potential risk for patients' safety, it is very common that a DSMB (data safety monitoring board) will be established to review the data on a on-going basis. Usually, there is probably a stopping rule for safety - the stopping rule based on the accumulated safety data that is different from the stopping rule for each individual patient.

The stopping rule can be based on the expert's opinion and more commonly can be derived from the statistical standpoint based on the known background rate. Here are some articles about the safety monitoring and the stopping rule for safety:

• Zhu et al (2015) Statistical Monitoring of Safety in Clinical Trials
• Chap Le (lecture notes) “AE” MONITORING IN PHASE II TRIALS
• Lecture notes: The Sequential Probability Ratio Test
• FDA’s Clinical Investigator Course: Safety Considerations in Phase 1 Trials

In fact, an easy way to construct a stopping rule is to utilize the hypothesis testing approach. The idea is to rule out that the observed event rate is higher than the background event rate and is to stop the trial if excessive event rate is observed.

Here is an example from an early phase clinical trial for ischemic stroke where a known risk for the experimental drug is to cause excessive symptomatic intracranial hemorrhage (SICH). The plan is to enroll and dose up to 20 subjects. The literature reviews reveal that the background rate for SICH in ischemic stroke patients is about 10%. We can then construct a stopping rule based on the background rate for SICH of 10%.

The DSMB is empowered to stop the trial whenever they deem fit to protect the safety of patients. DSMB will consider adopting a stopping rule that evaluates the rate of SICH after 5 enrollments, requiring suspension of trial entry if the observed rate of SICH be sufficient to reject the null hypothesis that the true SICH proportion (P) is 10% or less in favor of the (one-sided) alternative hypothesis that the true SCIH proportion is more than 10%. The table below depicts the stopping guidelines. Exact binomial methods are used to compute the p values and confidence limits. One-sided lower 90% confidence limit is calculated.

Number of Subjects Enrolled in the Study	Number of SICH Needed to Reject Null Hypothesis	Observed Rate of SICH (%)	One-sided p values	Lower 90% Confidence Limit
~5	2	40%	0.0815	11.22%
6-10	3	30%	0.0702	11.58%
11-15	4	27%	0.0556	12.18%
16-20	5	25%	0.0432	12.69%

If 2 SICH events are observed in less than and equal to 5 subjects enrolled, the lower confidence limit of 11.22% will be exceeding the 10% background rate. The stopping rule for safety will be triggered and the study should be stopped.

The SAS program for calculating the stopping rule will be something like below. Notice that we obtain two-sided 80% confidence interval in order to obtain one-sided 90% lower limit. Try-and-error method can be employed to find the lowest number of SICHs that will have lower 90% confidence limit exceed the background event rate (10%).

data stopping;
  input scenario SICH $ count;
  datalines;
  1 Have 2
  1 No   3
  2 Have 3
  2 No   7
  3 Have 4
  3 No   11
  4 Have 5
  4 No   15
run;

proc freq data=stopping;
  weight count;
  tables SICH /  binomial (p=0.10) alpha=0.20 cl;
                    **p=0.10 option indicates the background rate to compare with, here we assume the
                        SICH rate of 10%;
                    ** Alpha=0.20 to obtain two-sided 80% confidence interval;
  exact binomial;  *Obtain the exact p-value;
  by scenario;
run;