`RESEARCH
`
`
`
`APPLICATION NUMBER:
`202155Orig1s000
`
`MEDICAL REVIEW(S)
`
`
`
`
`
`
`
`
`
`
`
`Subject:
`
`
`NDA:
`
`
`Proposed Indication:
`
`Addendum Date:
`Clinical Reviewers:
`
`Apixaban Clinical Review Addendum
`202155
`Reduction in the rate of stroke and systemic embolism in
`subjects with nonvalvular atrial fibrillation
`December 17, 2012
`Martin Rose, M.D., J.D. (efficacy) and
`B. Nhi Beasley, Pharm.D. (safety)
`
`
`Table of Contents
`1
`Introduction ............................................................................................................... 2
`2 Discussion of Dr. Marciniak’s Special Review .......................................................... 2
`3 Observed Persistence of the Effect of Apixaban on Death after end of Study
`Treatment ........................................................................................................................ 4
`
`
`
`
`
`
`
`Reference ID: 3236037
`
` Page 1
`
`
`
`1
`
`Introduction
`
`This review addendum has two purposes:
`
`0 To address apixaban efficacy and safety issues raised in a special review by Dr.
`Thomas Marciniak, a medical team leader in DCRP who is not a member of the
`apixaban NDA review team but who has a strong interest in the safety issues
`described below.
`
`0 To discuss additional mortality analyses of ARISTOTLE, the study that is the
`primary support for the safety and efficacy of Apixaban for its proposed indication
`of reduction in the risk of stroke and systemic embolism in patients with non-
`valvular atrial fibrillation.
`
`2 Discussion of Dr. Marciniak’s Special Review
`
`Dr. Marciniak on his own initiative filed a “special clinical review” of the apixaban NDA
`on December 11, 2012. The review focused on the issue of missing data (mostly
`missing follow-up information) in ARISTOTLE, primarily data for death. He also
`addressed data on bleeding and cancer in ARISTOTLE and APPRAISE-2. This
`addendum will focus on the issue of missing data for mortality.
`In connection with
`incomplete follow-up information, Dr. Marciniak recommended the following:
`
`o The indication statement should include only stroke and systemic embolism.
`
`.
`
`(b) (4)
`
`o The Clinical Studies section of the label should include a discussion of the data
`
`quality problems in ARISTOTLE. It should summarize the dispensing errors and
`provide the missing follow-up statistics for both vital status and events. It should
`report that a change in one death eliminates the statistical significance of the
`death benefit and that, because of the missing data, we cannot have confidence
`in a death benefit.
`
`In previously filed reviews, the clinical reviewers reached the conclusions concordant
`with Dr. Marciniak’s first two recommendations:
`the indication should only include
`W"
`stroke and systemic embolism, and
`. Our views on those matters have not changed. However, we do not agree
`with his third recommendation concerning medication errors and the observed death
`benefit in ARISTOTLE.
`
`We acknowledge that dispensing errors in ARISTOTLE were a major review issue.
`However, the Applicant’s response to our CR letter, which included information derived
`from bottle labels, convinced us that the likelihood that the trial results were confounded
`by clinical events relating to medication errors was acceptably low (see our clinical
`review addendum of Dec. 10, 2012). Thus, the information that Dr. Marciniak would
`include in labeling regarding medication errors would be of negligible value in
`interpreting the findings of ARISTOTLE and would be more likely to confuse than to
`
`Page 2
`
`Reference ID: 3236037
`
`
`
`edify practitioners. Consequently, we do not agree that information about the
`medication errors should be included in labeling.
`Second, we do not agree that the labeling should include information regarding how
`many additional events in the apixaban arm or fewer events in the warfarin arm it would
`take to make the death finding not be statistically significant. A p-value of 0.0465
`already implies that the results are close. One way to think about a statistically
`significant mortality finding is that there were so many fewer deaths in the apixaban arm
`than in the control arm (in the ITT analysis, 603 vs. 669, 66 fewer, Table 1) that it is
`very unlikely that the observed finding was due to chance. If this difference in deaths is
`reduced to 65, there are still many fewer deaths in the apixaban arm. In order to
`establish superiority, it is not appropriate to require both the “cushion” of fewer deaths
`needed to achieve statistical significance at the 0.05 level plus an additional cushion to
`take care of post hoc “what ifs.” Moreover, the analysis that determines the number of
`events needed to overturn the mortality study finding (or stroke/se or major bleeding) is
`exploratory, highly conservative, non-random, and somewhat unrealistic.1 It should not
`be used for labeling.
`
`
`Table 1. ARISTOTLE Death and Primary Endpoint – ITT and On Treatment
`Analyses
`
`Apixaban
`
`Warfarin
`
`∆
`
`Apixaban vs. Warfarin
`Hazard
`Event1
`95% CI
`Ratio
`%/yr
`n/N
`%/yr
`n/N
`(0.80, 1.00)
`0.89
`66
`3.94
`669 / 9081
`603 / 9120 3.52
`Death ITT
`(0.74, 1.03)
`0.87
`31
`1.94
`296 / 9052
`265 / 9088 1.70
`Death Tx
`(0.75, 1.00)
`0.87
`42
`2.42
`372 / 9052
`330 / 9088 2.10
`Death TxLD+7
`(0.78, 1.01)
`0.89
`42
`2.97
`471 / 9052
`429 / 9088 2.65
`Death TxLD+30
`(0.66, 0.95)
`0.79
`53
`1.60
`265 / 9081
`212 / 9120 1.27
`Stroke SE ITT
`(0.63, 0.93)
`0.77
`49
`1.49
`225 / 9052
`176 / 9088 1.14
`Stroke SE Tx
`(0.63, 0.93)
`0.76
`52
`1.55
`236 / 9052
`184 / 9088 1.18
`Stroke SE TxLD+7
`(0.70, 1.00)
`0.84
`37
`1.62
`255 / 9052
`218 / 9088 1.36
`Stroke SE TxLD+30
`Reviewer’s analysis: erateHR create run eff tx txn7 txn30.sas, applicant’s data: adefl, adbs2
`Stroke SE is the primary endpoint. ∆=events in warfarin arm – events in apixaban arm
`The period of analysis for the event was defined as:
`ITT = randomization to January 30, 2011 (efficacy cut-off date)
`Tx = first dose to last dose + 2 days (per protocol definition)
`TxLD+7 = first dose to last dose + 7 days
`TxLD+30 = first dose to last dose +30 days
`
`
`1 For example, to determine the number of apixaban-treated subjects needed to negate the statistically
`significant mortality finding, subjects are ordered by apixaban treatment, then censor date. Events are
`sequentially imputed to apixaban-treated subjects without events until the results are not statistically
`significant. Because the cox proportional hazards model is dependent on time, this analysis is highly
`conservative since the additional events are occurring early in the trial.
`
`p-value
`0.0465
`0.1130
`0.0555
`0.0763
`0.0114
`0.0080
`0.0060
`0.0526
`
` Page 3
`
`
`
`Reference ID: 3236037
`
`
`
`In addition, the missing data rate in ARISTOTLE cited by Dr. Marciniak (3.2% to 3.6%
`for vital status) is not especially large and more importantly, it is unclear whether the
`missingness was biased in favor of apixaban. The primary efficacy analysis in this
`study was the ITT analysis. In a large global study with many subjects who withdraw
`consent, it is possible to lose track of subjects that stop coming to a site before the
`cutoff date for the ITT analysis. Perhaps some were lost to follow-up because they had
`a stroke or died, which potentially biases the study results. However, there is no reason
`to believe that this was more likely in the apixaban arm than in the warfarin arm. We
`cannot directly address whether the missing data are biased in one direction. However,
`on-treatment analyses are less likely to have missing follow-up information during the
`period of analysis. In the on-treatment analyses shown in Table 1, events are counted if
`they occurred during the analysis period. If the ITT analyses were biased in favor of
`apixaban because of differential event rates in those whose data are missing, one
`would expect the on- treatment results for the primary endpoint and death to be less
`favorable for apixaban than the ITT results because of better follow-up while subjects
`are on treatment. If the assumed bias in ITT analysis were removed in this way, we
`would expect the hazard ratio for death or the primary endpoint to move in favor of
`warfarin compared to ITT, barring other effects. Instead, point estimates for the on-
`treatment analyses for the primary endpoint and death are both slightly lower – i.e.,
`more favorable for apixaban – than the corresponding ITT results (Table 1). While this
`is not definitive proof of a lack of bias in the ITT analysis, it is reassuring and suggests
`that bias, if present, was not large. Given this reassurance from the on-treatment
`analyses and the lack of information to suggest bias in the ITT analysis, it would be
`confusing and potentially misleading to include data on follow-up statistics from
`ARISTOTLE in labeling.
`
` 3
`
` Observed Persistence of the Effect of Apixaban on Death
`after end of Study Treatment
`One other issue regarding the mortality finding should be mentioned. In the ITT
`analysis, there were 66 fewer all-cause deaths, but this was reduced to a difference of
`31 deaths in the on-treatment analysis. However, for primary endpoint events, the
`analogous differences in event counts are 53 and 49, respectively. Thus, for primary
`endpoint events (which were mostly strokes), nearly all of the benefit of apixaban was
`established during the treatment period, as one might expect for an anticoagulant.
`However, for death, a substantial portion of the benefit of apixaban was established off
`treatment (Table 1). This might be interpreted to make the observed benefit of
`apixaban for all-cause mortality to be less credible.
`However, there is an explanation for the observed persistence of the effect of apixaban
`on mortality after study drug is discontinued that does not undercut the strength of the
`overall finding in the ITT analysis. It relates to the large effect of apixaban on the rate of
`fatal stroke, the timing of death in fatal stroke cases, and the fact that physicians
`practice medicine conservatively, i.e., when a patient in a clinical study becomes
`
` Page 4
`
`
`
`Reference ID: 3236037
`
`
`
`seriously ill, they are often taken off an investigational drug and treated with usual
`therapy for the patient’s condition.
`In ARISTOTLE, the difference between the treatment arms in fatal stroke strongly
`favored apixaban in the ITT analysis (38 vs. 65 deaths, HR=0.58, 95% CI: 0.39, 0.86).
`It is notable that while fatal stroke accounts for less than 10% of deaths in the
`ARISTOTLE ITT analysis, the difference in the number of fatal strokes between the
`treatment arms accounts for more than 40% of the overall difference in deaths. This
`relationship was also observed in RE-LY (see our review of Applicant’s Complete
`Response dated 10 Dec. 2012). This difference suggests that the observed overall
`difference in favor of apixaban is not due to chance. It thus seems useful to examine
`the timing of stroke mortality in relation to the subjects last dose of study drug and last
`known stroke event.
`Figure 1 is a display of the difference in days between the last dose of study drug and
`the last known adjudicated stroke in subjects with death adjudicated as a CV death due
`to stroke (“fatal stroke”) in 102 of the 109 subjects with a fatal stroke in the ITT
`analysis.2 Data for both treatment arms are combined in this figure and the 2 others
`that follow, but all of the trends in the data in the 3 figures discussed here were similar
`in the two treatment arms (data not shown).The majority of subjects with a fatal stroke
`(72 subjects, or 71%) had their last dose of study drug on the day of the stroke or one
`day earlier. This is consistent with the conservative practice of discontinuing an
`experimental drug when a patient becomes seriously ill, especially when other
`recommended therapies are available.
`Figure 2 is a display of the difference in days between the last dose of study drug and
`death for the same 102 subjects, and Figure 3 shows data for the days between the
`final stroke and death. For fatal stroke, which was the single largest contributor to the
`difference in deaths between the treatment arms, the data indicate only 30 of the
`subjects (29%) died “on treatment” (i.e., 0 to 2 days after their last dose of study drug,
`and 56 (55%) died within 7 days of their final last dose (Figure 2). Forty-eight subjects
`(47%) died within 2 days of their final stroke, 77 (75%) died within 7 days of their final
`stroke (Figure 3).
`Thus, the data from the sample of subjects with fatal strokes indicate that a substantial
`part of the apparent late effect of apixaban on death has an explanation that does not
`undercut the observed beneficial effect of apixaban on mortality: In 29% of subjects
`with a known date for their final and fatal stroke, the stroke occurred after the on-
`treatment period. However, over half of deaths occurred more than 2 days after the
`final stroke and 1/4 of subjects in the sample died more than 7 days after their last
`stroke. Thus, it is not surprising that the effect of apixaban on mortality is not confined
`to the on-treatment period.
`
`
`2 The remaining 7 subjects with a fatal stroke (5 in the warfarin arm and 2 in the apixaban arm) did not
`have either a “stroke” in the study database nor a date for the stroke, but they had a date and adjudicated
`cause for mortality. While these patients were included in analyses of death, they were not included in
`analyses of stroke or the primary endpoint performed by the applicant or by FDA.
`
` Page 5
`
`
`
`Reference ID: 3236037
`
`
`
`Figure 1. Distribution of Subjects with Fatal Stroke: Days from Last Dose of
`Study Drug to Final Stroke (N=102)
`
`60
`
`50
`
`40
`
`30
`
`20
`
`10
`
`0
`
`-15 -4 -3 -2 -1 0
`1
`2
`3
`4
`8 10 11 14 17 24 27 28 31 38 49 59 79 84 176212358718
`Horizontal axis is days between last dose of study drug and adjudicated final stroke event. Negative
`values mean stroke occurred before the last dose; positive values mean the last dose occurred before the
`stroke.
`Vertical axis represents the number of subjects at each time point.
`
`
`
`
`Figure 2. Distribution of Subjects with Fatal Stroke: Days from Last Dose of
`Study Drug to Death (N=102)
`
`20
`18
`16
`14
`12
`10
`
`02468
`
`
`
`724
`601
`358
`212
`189
`94
`92
`91
`65
`63
`51
`41
`35
`33
`30
`28
`24
`23
`20
`18
`16
`14
`12
`11
`10
`
`012345678
`
`Horizontal axis is days between last dose of study drug and adjudicated CV death due to stroke.
`Vertical axis represents the number of subjects at each time point.
`
`
`
`
`
`
`
`Reference ID: 3236037
`
` Page 6
`
`
`
`Figure 3. Distribution of Subjects with Fatal Stroke: Days from Final Stroke to
`Death (N=102)
`
`20
`18
`16
`14
`12
`10
`
`02468
`
`0
`
`1
`
`2
`
`3
`
`4
`
`5
`
`6
`
`7
`
`8
`
`9
`
`10 12 13 15 17 19 20 23 25 35 40 91 600
`
`
`
`Horizontal axis is days between adjudicated final stroke and adjudicated death due to stroke. Vertical
`axis represents the number of subjects at each time point.
`
`Notably, despite the narrowing of the difference in deaths between the treatment arms
`after treatment, the hazard ratio for death (apixaban vs. warfarin) is not worse (Table 1).
`Also, while the following data from other trials do not directly address the issue of the
`prolongation of the apparent effect of apixaban, they are reassuring in that they
`indicated the pattern of timing with respect to treatment of deaths vs. primary endpoint
`events in ARISTOTLE is not unique.
`In ARISTOTLE a large number of deaths occurred after the end of treatment, a pattern
`that was not observed for the primary endpoint events. The ratio of primary endpoint
`events in the on treatment period (i.e., first dose to last dose + 2 days) vs. those in the
`ITT period was 401:444, or 0.84. Because this ratio is reasonably close to unity, it
`should not be surprising that the differences in primary endpoint events between the
`treatment arms (warfarin minus apixaban) were similar in the two periods (53 and 49 in
`the ITT and on-treatment analyses, respectively). However, the ratio of deaths during
`the on treatment period vs. the ITT period was 561:1272, or 0.44.
`It is not clear why the ratio of deaths in the on-treatment period vs. ITT period is higher
`for primary endpoint events than for deaths. Better ascertainment of death than stroke
`after treatment may account for some of the difference, but it seems likely that
`discontinuation of study drug because of a serious illness (such as a stroke) that
`eventually was fatal accounted for some of the difference (see Figure 1 to Figure 3 and
`associated text above).
`A similar pattern regarding the proportion of primary endpoint events vs. deaths that
`occurred on-treatment was observed in ROCKET and RE-LY, which were warfarin-
`controlled trials of rivaroxaban and dabigatran, respectively, performed in patients with
`nonvalvular atrial fibrillation (Table 2). Primary endpoint events were defined similarly
`in all 3 trials. As in ARISTOTLE, in both ROCKET and RE-LY the ratio of total deaths
` Page 7
`
`
`
`Reference ID: 3236037
`
`
`
`on treatment vs. deaths during the ITT period was substantially lower than the
`analogous ratio for primary endpoint events: 0.37 vs. 0.75 respectively for ROCKET
`and 0.55 vs. 0.74 respectively for RE-LY.
`
`
`Table 2. Atrial Fibrillation Trials – Endpoint Events by Period
`
`
`
`ARISTOTLE
`Apixaban vs. W
`
`ROCKET
`Rivaroxaban vs. W
`
`Deaths: n in ITT
` HR (95% CI)
`Deaths: n in Tx
` HR (95% CI)
`Ratio: n in Tx / n in ITT
`
`1272
`0.89 (0.80, 1.00)
`561
`0.87 (0.74, 1.03)
`0.44
`
`1264
`0.92 (0.82, 1.03)
`458
`0.85 (0.70, 1.02)
`0.37
`
`RE-LY
`Dabigatran 150 mg
`vs. W
`923
`0.88 (0.77, 1.00)
`506
`NP
`0.55
`
`
`
`
`
`
`
`
`
`575
`0.88 (0.74, 1.03)
`432
`0.79 (0.65, 0.95)
`0.75
`
`336
`0.65 (0.52, 0.81)
`250
`0.64 (0.50, 0.81)
`0.74
`
`477
`0.79 (0.66, 0.95)
`401
`0.77 (0.63, 0.93)
`0.84
`
`Stroke/SE : n in ITT
` HR (95% CI)
`Stroke/SE: n in Tx
` HR (95% CI)
`Ratio: n in Tx / n in ITT
`W = Warfarin
`Stoke/SE = Primary endpoint: time to first stroke or systemic embolism in each study
`Death = All cause death, analyzed as time to first event
`ITT = Intent to treat period, counting events from randomization to the time that sites were notified that the
`event target had been reached and the study was to end.
`Tx = On-treatment period, counting events from date of first dose to date of last dose + 2 days for
`ARISTOTLE and ROCKET and from first dose to last dose for RE-LY
`HR is for the experimental drug vs. warfarin in each study; data for ROCKET from NDA review; data for
`RE-LY from review or from NDA for on treatment data only
`NP= not provided.
`
`
`
`
`Reference ID: 3236037
`
` Page 8
`
`
`
`---------------------------------------------------------------------------------------------------------
`This is a representation of an electronic record that was signed
`electronically and this page is the manifestation of the electronic
`signature.
`---------------------------------------------------------------------------------------------------------
`/s/
`----------------------------------------------------
`
`BACH N BEASLEY
`12/21/2012
`
`MARTIN ROSE
`12/21/2012
`
`Reference ID: 3236037
`
`
`
`
`
`
`
`
`
`
`
`
`
`Date:
`
`Reviewer:
`
`
`NDA:
`
`Drug:
`
`Indication:
`
`CLINICAL REVIEW (Updated)
`
` DEPARTMENT OF HEALTH AND HUMAN SERVICES
` PUBLIC HEALTH SERVICE
`
`FOOD AND DRUG ADMINISTRATION
` CENTER FOR DRUG EVALUATION AND RESEARCH
`DIVISION OF CARDIOVASCULAR AND RENAL PRODUCTS
`
`
`
`
`
`December 17, 2012
`Thomas A. Marciniak, M.D.
`Medical Team Leader
`
`202-155
`
`apixaban (Eliquis®)
`
`To reduce the risk of stroke, systemic embolism,
`nonvalvular atrial fibrillation
`
` in patients with
`
`Completeness of follow-up and bleeding and cancer
`
`
`Subjects:
`
`Summary and Recommendations
`Because of a special interest in and experience with two issues, completeness of follow-up and
`cancer, I reviewed the apixaban studies regarding these issues. I filed an initial review on
`December 11, 2012. I updated that review for a sponsor submission dated December 14, 2012,
`on study closeout procedures and for FDA discussions regarding the cancer issue and I corrected
`typos and minor inaccuracies. This updated review incorporates and completely replaces my
`initial review. My summary and recommendations remain unchanged:
`
` I
`
` document below that completeness of follow-up and reporting of dates were poor in
`ARISTOTLE. Our confidence in the fragile alleged death benefit (with one additional death in
`the apixaban arm eliminating statistical significance) is destroyed by the missing vital status.
`Our confidence in the superiority of the stroke benefit to warfarin is also challenged by
`incomplete follow-up. Finally, the ARISTOTLE and APPRAISE-2 trials show an association
`between bleeding and solid cancers also seen in other anticoagulant and antiplatelet drug trials.
`
` recommend the following:
`
` The indication statement should include only stroke and systemic embolism.
`
`
`
`.
`
` The Clinical Studies section of the label should include a discussion of the data quality
`problems in ARISTOTLE. It should summarize the dispensing errors and provide the
`
`
`
` I
`
`
`
`
`
`Reference ID: 3232518
`
`(b) (4)
`
`(b) (4)
`
`
`
`
`
`
`
`missing follow-up statistics for both vital status and events. It should report that a change
`in one death eliminates the statistical significance of the death benefit and that, because
`of the missing data, we can not have confidence in a death benefit.
`
` The data regarding bleeding and cancer should be presented and discussed at an advisory
`committee meeting. If the rigorous analysis of the ARB trials confirms a risk for ARBs
`and cancer, then the data regarding ARBs and cancer should also be presented and
`discussed at an advisory committee meeting. It may be advantageous to have both topics
`addressed at the same meeting.
`
`
`
`
`
`
`
` The data regarding bleeding and cancer should be included in the apixaban label and in
`the labels for all antiplatelet and anticoagulant drugs. If the advisory committee meeting
`discussing bleeding and cancer is scheduled promptly then the labeling regarding
`bleeding and cancer can be delayed until after the meeting.
`
`
`Completeness of Follow-up and Fragility of Results in ARISTOTLE
`
`Definition of Completeness of Follow-up
`The clinical study report (CSR) for the ARISTOTLE trial of apixaban vs. warfarin in atrial
`fibrillation states that vital status could not be determined for 2.0% in the apixaban group and
`2.2% in the warfarin group (380 patients in both groups total). The main study publication
`reported the same vital status statistics. (Granger, Alexander et al. 2011) However, the rates of
`discontinuation from the study were much higher, 25.3% in the apixaban group and 27.5%, with
`10.1% of apixaban patients and 10.0% of warfarin patients discontinuing at their own request.
`While these reported statistics for completeness of follow-up are not good, my recent experience
`with other outcome trials suggests that the sponsor’s reporting of completeness of follow-up is
`usually optimistic compared to analyses of the submitted datasets. Hence I analyzed the datasets
`for completeness of follow-up.
`
` I
`
` assert that there is a straightforward definition of completeness of follow-up: Most outcome
`studies have a specified global study end date or censoring date for efficacy outcomes. A few
`have a pre-specified duration of follow-up from randomization such as two years. I assert that
`follow-up is complete if the patient has documented follow-up on or after the specified end date.
`
`Per a statistical analysis plan appendix and the NEJM publication ARISTOTLE had a cutoff date
`for efficacy outcomes of January 30, 2011. However, I note that the December 14, 2012,
`submission refers to “this common efficacy cut-off date (31-Jan-2011)” and has the following
`detailed description:
`
`
`“By December, 2010 we had confirmed our view that we would reach 448 events in January,
`2011. We therefore organized our CRO partners and site monitors to prepare them for the
`efforts that would be involved in the close out process, and asked them to communicate to
`sites our expectation that the efficacy cutoff date would be January 31, 2011. Furthermore,
`all visits for the cessation of study drug (and initiation of VKA or other antithrombotic
`agents) were to occur no sooner than 31-Jan-2011.”
`
`
`
`
`
`Reference ID: 3232518
`
`2
`
`
`
`These statements make it unclear whether the pre-specified efficacy cutoff date was January 31,
`2011, or January 30, 2011. Hence for ARISTOTLE follow-up for a patient is complete if the
`patient had documented follow-up on or after January 30, 2011-—or January 31, 2011?
`
`Determining completeness of follow-up has another complication: The type of follow-up
`typically varies in outcome studies. For patients who have a face-to-face study visit with the
`investigator on or after the study end date follow-up is complete for all study outcomes or events.
`However, for some patients final follow—up may consist of a phone call with the patient or a
`spouse or a primary physician, for others a report of a hospitalization, and for still others a
`newspaper obituary or a registry report of alive or dead. The level of detail available from these
`latter, non—face—to—face follow-ups varies and, while usually adequate for determining vital
`status, may not be adequate for ascertaining endpoints or adverse events. I recommend
`estimating two levels of completeness of follow—up: (1) vital status; and (2) events. For the
`former I accept any type of documented follow-up having a date unambiguously referencing the
`patient as alive. For the latter I accept reports of documented face-to-face visits, hospitalizations
`or other events, and phone calls with documented queries regarding events. Because of
`ambiguities in case report form (CRF) design and in reporting by sites determining completeness
`of event follow-up requires subjective judgments.
`
`Completeness of Vital Status Follow—up
`For one estimate of completeness of vital status follow-up in ARISTOTLE I used the sponsor’s
`variable e_cddn described as “Censor Date for Death ITT Period” from the sponsor’s
`ADEFS.XPT (“Efficacy Summary”) dataset. Counting good vital status as either known dead at
`any time or a censor date of January 30, 2011, or later, 3.2% of patients (589) are missing vital
`status. For a second estimate I also analyzed all of the datasets for the maximum dates of events,
`procedures, vital sign recordings, and status reports to estimate a date of last follow-up for every
`patient. By this latter analysis of the raw data 3.6% of patients (659) are missing vital status.
`Either number, 589 or 659, greatly exceeds the sponsor’s and investigators’ reports of 380
`missing vital status follow-up.
`
`Date Recording Problems in ARISTOTLE
`Both the sponsor’s and my estimates of vital status follow-up are likely optimistic: Dates are not
`infrequently misrecorded or misinterpreted in ARISTOTLE. The death CRF excerpt in Figure 1
`illustrates the possibility for misrecording or misinterpretation: The ‘wisit date” for the death
`“(6’
`CRF precedes the death date. (Note that the reported death date was
`. This patient was one of the three warfarin patients reported to have died
`—compared to no apixaban patients. See the further discussion of problems
`with death reports below.)
`
`(D) (6)
`
`Reference ID: 3232518
`
`
`
`Figure 1: Death in the Future
`
`:33} Iinslnl-Myors SquilII: Campany
`H
`fl
`PROTOCCI.
`HSITE wag!
`CV185 030
`l._: |_. ._a I-I H
`
`
`
`t - JAN-201 t
`ugfi'La‘wa
`
`
`CLINICAL EVENT ASSESSMENT EVENT‘SSW'NS
`
`Did the Subject Experience a Clinical Event?
`If Yes, Camden) Below.
`
`Date 0‘ Onset of Event:
`
`CITnical Event:
`
`DEATH
`
`0M6)
`W
`
`E] NO .YES
`
`CI ”0 .YEs
`
`The "visit date” may have little connection to a date on which the patient was actually observed
`or contacted. In many cases there is no way of determining whether the ‘Visit date” is an
`observation date or the date of recording or someflijng else. However, both I and the sponsor
`based our dates of last vital status partially on “visit dates”. In particular a critical form for
`follow-up, the End Of Follow-up CRF shown in Figure 2, has this ambiguity regarding “visit
`date”.
`
`Figure 2: End of Follow-up CRF
`
`VXVABristol-Myers Squibb Company
`
`PROTOCOL CV1 85°30
`VISIT DATEQDDMMMYY
`
`
`
`End of Follow-q:
`SITE
`SUBJECT
`NUMBER D D D D NUMBER D D D E E
`
`Page 362
`
`VISIT
`
`CODEX99
`
`
`SUBJECT STATUS
`
`DSTERM DSDECOD
`DID THE SUBJECT COMPLETE THE FOLLOW-UP PHASE OF THIS STUDY ?
`
`0 |:] NO
`
`1 |:] YES
`
`IF NO, PLEASE INDICATE PRIMARY REASON (MARK ONE)
`
`DSTERM DSDEC OD
`4 D SUBJECT WITHDREW CONSENT (SPECIFY)
`
`.
`DSTERIVI
`
`12 |:l DEATH
`
`98 CI OTHER (SPECIFY) DSTERM
`
`8 D LOST T0 FOLLOW-UP
`
`DE‘STDTC
`(DATE OF LAST CONTACT) DDMMMYY
`
`Reference ID: 3232518
`
`
`
`
`
`
`
`
`
`What does the “visit date” on the End of Follow-up CRF represent? Only for lost to follow-up is
`the date of last contact to be recorded on this CRF. For death there is the death form with a field
`for date of death, but what about “withdrew consent” or “other” or even “complete”? It is easy
`to document that “visit date” for “withdrew consent” likely does not represent the date on which
`the patient visited the site or withdrew consent. For example, one patient discontinued treatment
`on 16jul00 with the last verifiable events on 18jun00. However, the disposition (DS) dataset has
`a “Start Date/Time of Disposition Event” for withdrew consent of 12apr11 and the sponsor
`counts the patient as completing follow-up, censoring on 30jan11. Withdrawing consent on
`12apr11, long after the trial ended, is not rational and would not represent a withdrawal of
`consent during the ITT period, as the sponsor classifies this patient. Another patient is similar,
`with end of treatment and last events on 06may10 but withdrawal of consent allegedly on
`24feb11 with sponsor’s censoring on 30jan11. I count both of these patients (and other similar
`ones) as having incomplete vital status follow-up, partially explaining and justifying why my
`estimate of incompleteness of vital status follow-up is higher than the sponsor’s.
`
`There are other examples of anomalous dates, e.g., at least three patients have dates of last
`contact by “direct contact with subject” long after the patient was reported dead. In fact, about
`65% of patients who died have a visit date or other date greater than the date of death.
`Furthermore, for patients who did not die during the study, we do not have an unequivocal last
`date against which to check the validity of reported dates. There is no good way to detect or
`resolve many of these date inconsistencies even with a painstaking manual review of the CRFs--
`and we do not have most of the CRFs. Any estimates of completeness of vital status follow-up
`are optimistic.
`
`Fragility of the Death Benefit in ARISTOTLE
`While the estimates of missing vital status follow-up are concerning (>3%) and likely higher,
`how do they relate quantitatively to the reported death benefit? The sponsor claims that there
`were statistically significantly fewer deaths in the apixaban arm than in the warfarin arm in
`ARISTOTLE based on the analysis in Table 1 excerpted from the clinical study report:
`
`Table 1: Summary of Adjudicated Causes of Death during the Intended Treatment Period
`– Randomized Subjects (Excerpt from Sponsor’s Table 7.2.1)
`
`
`The p-value is close to 0.05 and the upper confidence limit for the hazard ratio is 1.00 so we
`know that this result is very fragile. How fragile? A change in only one death (one more with
`apixaban or one fewer with warfarin) could make this result nominally statistically insignificant.
`This critical number 1 is dwarfed by 317, the estimated—but still likely optimistic—number of
`patients with missing vital status in the apixaban arm. Even the total difference in deaths
`
`
`
`
`
`Reference ID: 3232518
`
`5
`
`
`
`between the two arms (66) is a small fraction of number of apixaban patients with missing vital
`status.
`
`This numeric fragility may not be surprising but, in addition, the quality of the documentation of
`the deaths is fragile or suspect. I examined deaths around the time of the cutoff date (January 30,
`2011) for efficacy outcomes. Three warfarin patients (and no apixaban patients) died
`(ms)
`One patient in Hungary discontinued treatment on 4jan11 for bleeding with the clinical
`event assessment shown in Figure 1 dated 10jan1