Minimally Invasive Surgery With Thrombolysis for Intracerebral Hemorrhage Evacuation

Abstract

Background and Objectives

Bayesian analysis of randomized controlled trials (RCTs) can extend the value of trial data beyond interpretations based on conventional p value–based binary cutoffs. We conducted an exploratory post hoc Bayesian reanalysis of the minimally invasive surgery with thrombolysis for intracerebral hemorrhage (ICH) evacuation (MISTIE-3) trial and derived probabilities of potential intervention effect on functional and survival outcomes.

Methods

MISTIE-3 was a multicenter phase 3 RCT designed to evaluate the efficacy and safety of the MISTIE intervention. Five hundred and six adults (18 years or older) with spontaneous, nontraumatic, supratentorial ICH of ≥30 mL were randomized to receive either the MISTIE intervention (n = 255) or standard medical care (n = 251). We provide Bayesian-derived estimates of the effect of the MISTIE intervention on achieving a good 365-day modified Rankin Scale score (mRS score 0–3) as relative risk (RR) and absolute risk difference (ARD), and the probabilities that these treatment effects are greater than prespecified thresholds. We used 2 sets of prior distributions: (1) reference priors, including minimally informative, enthusiastic, and skeptical priors, and (2) data-derived prior distribution, using a hierarchical random effects model. We additionally evaluated the potential effects of the MISTIE intervention on 180-day and 30-day mRS and 365-, 180-, and 30-day mortality using data-derived priors.

Results

The Bayesian-derived probability that MISTIE intervention has any beneficial effect (RR >1) on achieving a good 365-day mRS score was 70% using minimally informative prior, 87% with enthusiastic prior, 68% with skeptical prior, and 73% with data-derived prior. However, these probabilities were ≤55% for RR >1.10 and 0% for RR >1.52 across a range of priors. The probabilities of achieving RR >1 for 180- and 30-day mRS scores are 65% and 80%, respectively. Furthermore, the probabilities of achieving RR <1 for 365-, 180-, and 30-day mortality are 93%, 98%, and 99%, respectively.

Discussion

Our exploratory analyses indicate that across a range of priors, the Bayesian-derived probability of MISTIE intervention having any beneficial effect on 365-day mRS for patients with ICH is between 68% and 87%. These analyses do not change the frequentist-based interpretation of the trial. However, unlike the frequentist p values, which indirectly evaluate treatment effects and only provide an arbitrary binary cutoff (such as 0.05), the Bayesian framework directly estimates the probabilities of potential treatment effects.

Trial Registration Information

ClinicalTrials.gov/ct2/show/NCT01827046.

Classification of Evidence

This study provides Class II evidence that minimally invasive surgery (MIS) + recombinant tissue plasminogen activator (rt-PA) does not significantly improve functional outcome in patients with ICH. However, this study lacks the precision to exclude a potential benefit of MIS + rt-PA.

Introduction

Primary intracerebral hemorrhage (ICH) is associated with high mortality rate and poor functional outcomes.^1,2 Moreover, ICH lacks effective treatment options to improve patient outcomes,^3,4 with large clinical trials failing to demonstrate that conventional surgical modalities significantly improve functional outcomes or mortality.^5-7 The minimally invasive surgery with thrombolysis in ICH evacuation-3 (MISTIE-3) was a phase 3 randomized controlled trial (RCT) designed to evaluate the effectiveness of combining minimally invasive surgery with thrombolysis to improve functional outcomes.⁸ This trial was designed after experiences and preliminary evidence of safety and efficacy obtained from MISTIE-2 trial.^9,10 Specifically, the MISTIE-3 trial was 88% powered to detect an increase in the proportion of patients achieving a 365-day modified Rankin Scale (mRS) score¹¹ of 0–3, from 25% in the control group to 38% in the intervention group (absolute risk difference [ARD] of 13% and relative risk [RR] of 1.52). However, MISTIE-3 did not demonstrate a significant improvement in mRS outcome in the intervention group (ARD = 4%; p value = 0.33).

Of note, MISTIE-3 used conventional frequentist statistical paradigm to draw inferences about the effectiveness of the MISTIE intervention. The frequentist approach does not directly estimate the probability of observing hypothesized treatment effects.¹² Instead, treatment effects are indirectly evaluated by estimating the probability that the true treatment effect is the same as or larger than what the observed treatment effect would be if the null hypothesis was true. Hence, generating and interpreting a p value with an arbitrary cutoff (usually less than 0.05) to make a binary decision of whether there is sufficient evidence to reject the null hypothesis. By contrast, the Bayesian paradigm directly evaluates the probability of a hypothesized treatment effect given the observed data.^13-16 Moreover, unlike the frequentist approach, Bayesian analysis combines information from previous studies regarding the plausible values of the treatment effect (prior distribution) with the likelihood of observing the data, given the hypothesized treatment effect (likelihood function), to estimate an updated distribution of the probabilities of observing hypothesized treatment effects (posterior distribution). These direct estimates of treatment benefit probabilities may be relevant to clinicians and patients' families, in guiding them toward shared decision-making, particularly for a devastating condition such as ICH, which has no clinically proven treatment modalities.¹⁷ Therefore, using a range of priors, we conducted a post hoc Bayesian analysis of the MISTIE-3 trial to estimate the posterior probabilities of the effect of the MISTIE intervention on functional and survival outcomes.

Methods

MISTIE-3 was an open-label, blinded end point, phase 3 RCT designed to evaluate the efficacy and safety of image-guided, catheter-based removal of 30 mL or more of ICH plus alteplase administration (study protocol and registration at ClinicalTrials.gov/ct2/show/NCT01827046). The trial was conducted at 78 hospitals across the USA, Canada, Europe, Asia, and Australia (patient selection chart in eFigure 1, links.lww.com/WNL/D72). Eligible participants included 506 adult patients (18 years or older) with a spontaneous, nontraumatic, supratentorial ICH of ≥30 mL. Patients with life-threatening mass effect requiring surgery were excluded (full inclusion/exclusion criteria reported elsewhere).⁸ Combining block randomization (block size 4–6) and covariate-adaptive randomization (with age, presentation Glasgow Coma Scale score, and ICH size), trial staff and statisticians used computer-generated number sequence to randomly assign patients to either receive MISTIE intervention (n = 255) or standard medical care (n = 251). An adjudication committee, which was blinded to treatment allocation, determined study outcomes.

MISTIE-3 was designed to detect an increase in the proportion of patients attaining a 365-day mRS score¹¹ of 0–3 (henceforth described as good mRS score), from 25% in the control group to 38% in the intervention group (RR = 1.52; ARD = 13%). Data from MISTIE-2 trial^9,10 provided the basis for sample size and power calculations. Other secondary outcomes included 30- and 180-day mRS scores as well as 30-, 180-, and 365-day mortality (main results of original trial in eTables 1A and 1B, links.lww.com/WNL/D72). Further details of the trial, including reasons for protocol amendments, have been reported in Hanley et al.⁸

Post hoc Bayesian reanalysis has been previously used to augment interpretations of clinical trial results.^18,19 Using a similar approach, we sought to estimate the probabilities (posterior) of the potential effect of MISTIE intervention on achieving a good 365-day mRS score at prespecified thresholds of RR (RR >1; RR >1.02; RR >1.10; and RR >1.52) and ARD (ARD ≥1%; ARD ≥2%; ARD ≥3%; ARD ≥5%; ARD ≥11%; and ARD ≥13%). Assuming an annual ICH incidence of 70,000 in the United States,²⁰ an ARD of 1% would be equivalent to 700 more ICH survivors attaining good functional recovery or 700 lives saved annually.

Bayesian analysis derives eminence from its ability to incorporate prior beliefs about the plausible values of parameter estimates (Appendix Methods, links.lww.com/WNL/D72). To choose suitable prior distributions, we employed 2 commonly used approaches.^16,18,21,22 First, we chose a set of reference priors that reflect varying degree of enthusiasm or skepticism regarding the effect of the MISTIE intervention on functional outcome^23-26: (1) minimally informative prior, (2) enthusiastic prior, and (3) skeptical prior. Each of these reference prior distributions assume specific values of median RR and variance of the RR, thereby reflecting the degree of uncertainty regarding the effectiveness of the MISTIE intervention.^23-26 The minimally informative prior assumes that all values of log RR are equally likely, producing a posterior distribution that primarily relies on the likelihood function derived from MISTIE-3 trial data. The enthusiastic prior distribution assumes that the probability of obtaining RR of 1.52, the RR used to power the MISTIE-3 trial, is 50%. The skeptical prior assumes, with substantial certainty, that MISTIE intervention is equally likely to lead to benefit or harm (probability of obtaining RR <1 is 50%). In the context of this study, the minimally informative prior reflects a scenario in which we have no prior information or estimate regarding the plausible values of the treatment effect. The enthusiastic prior reflects RR used to power the MISTIE-3 trial, and the skeptical prior reflects the fact that prior clinical trials have failed to demonstrate a clear benefit of the MISTIE intervention on functional outcomes. Table 1 presents the computed sample sizes required for a hypothetical clinical trial to detect the median RR specified for each prior distribution. These sample sizes were calculated by multiplying the sample size of MISTIE-3 trial by the ratio of the variance of the log-RR of MISTIE-3 trial to the variance of the respective reference prior (eFigure 2a).

Table 1.

Representation of Prior Beliefs (Prior Probability Distributions) Regarding the Effect of MISTIE on 365-Day Functional Outcome (mRS Score)

Prior belief	Assumed median RR	Assumed SD of logarithm of RR	Prior evidence equivalent	Probability that RR >prespecified threshold, %
				RR >1	RR >1.02	RR >1.10	RR >1.52
Minimally informative	1.0	10	Similar to having no prior belief	50	50	50	48
Enthusiastic	1.52	0.26	Equivalent to a prior RCT with 83 enrolled participants finding 52% increase in RR	95	94	90	50
Skeptical	1.0	0.26	Equivalent to a prior RCT with 83 enrolled participants finding 0% increase in RR	50	47	35	5

Abbreviations: MISTIE = minimally invasive surgery with thrombolysis in intracerebral hemorrhage evacuation; mRS = modified Rankin Scale; RCT = randomized controlled trial; RR = relative risk.

In the second approach, we computed a data-derived prior distribution using a Bayesian hierarchical random effects model, as recommended in a previous methodological review, as well as the US Food and Drug Administration (FDA) guidance on Bayesian analysis.^16,27 In this model, the MISTIE-2 trial¹⁰ data provided the prior for what the effect of the MISTIE intervention would be in the MISTIE-3 trial. This prior was then combined with MISTIE-3 trial data to update the posterior distribution of the effect of MISTIE intervention on achieving a good 365-day mRS score after the MISTIE-3 trial. A minimally informative prior was placed on the parameter that captured between-study variability. To further account for potential differences in MISTIE-2 and MISTIE-3 trial settings, we inflated the variance of MISTIE-2 data to varying degrees (from 0% to 100%), thereby “downweighing” or reducing the influence of MISTIE-2 patients on the pooled treatment effect of the MISTIE intervention (eFigure 2B, links.lww.com/WNL/D72).

Similarly, we used data-derived prior distributions to evaluate the posterior probabilities of the effect of the MISTIE intervention on achieving good 30- and 180-day mRS scores as well as 30-, 180-, and 365-day mortality. In these secondary analyses, we estimated the posterior probabilities that the treatment effect of the MISTIE intervention is greater than prespecified thresholds of mRS outcomes (RR >1; RR >1.02; RR >1.10; and RR >1.52 and ARD ≥1%; ARD ≥2%; ARD ≥3%; ARD ≥5%; ARD ≥11%; and ARD ≥13%) and mortality outcomes (RR <1; RR <0.9; RR <0.8; and RR <0.7 and ARD ≥1%; ARD ≥3%; ARD ≥5%; ARD ≥7%; ARD ≥9%; and ARD ≥13%).

We fit separate Bayesian models for each prior distribution to estimate the effect of the MISTIE intervention on the log-RR of achieving good mRS and mortality outcomes. For each model, the number of participants with the specified outcome in the intervention and control groups was treated as independent samples of binomial distributions. Furthermore, we assigned a uniform prior distribution to the probability of observing the specified outcome in the control group. Consequently, the probability of observing the specified outcome among the MISTIE intervention group was calculated as RR multiplied by the probability of the outcome in the control group. We used Markov chain Monte Carlo simulation, with 3 chains, 60,000 saved iterations, and 20,000 burn-in iterations per chain, to report the mean log-RR of achieving good 365-day mRS score and 95% credible intervals (CrIs) (i.e., the 2.5th and 97.5th percentiles of the posterior probability distribution). Of note, the Bayesian 95% CrIs provide direct estimates of the interval within which the true (but unknown) parameter lies with 95% probability, given the observed data. By contrast, the frequentist's 95% CI is an indirect estimate, which stipulates, based on repeated hypothetical experiments, that the true parameter estimate would lie between the upper and lower limits of the interval. Furthermore, we evaluated the posterior probability that the RR of achieving a good mRS score among the MISTIE group is greater than 1, 1.02, 1.10, and 1.52. In addition, we computed the ARD for treatment effects using the estimated RR and the baseline risk, which is the proportion of individuals among the control group with the specified outcome. We used the Gelman-Rubin statistic²⁸ to assess model convergence (Appendix Methods, links.lww.com/WNL/D72).

All Bayesian analyses were conducted as per the trial's original modified intention-to-treat (mITT) paradigm. However, for 365-day mRS, we additionally conducted a per protocol analysis to estimate the probability of treatment benefit among patients in the MISTIE group who achieved the surgical aim of the trial (postprocedural clot size ≤15 mL). We conducted all analyses in R version 4.0.2²⁹ using R2jags.³⁰

Standard Protocol Approvals, Registrations, and Patient Consents

Each participating institution in the MISTIE-3 trial obtained ethical approval from its local institutional review board committee. A written informed consent was obtained from each participant or their authorized representative. Study protocol, registration, statistical analysis plan (Bayesian addendum in eAppendix, links.lww.com/WNL/D72), and informed consent are available at ClinicalTrials.gov/ct2/show/NCT01827046.

Data Availability

Deidentified data not published in this study are available to qualified investigators on request and signing of a data use agreement at the National Institute of Neurological Disorders and Stroke data archive (ninds.nih.gov/) or the Virtual International Stroke Trials Archive (virtualtrialsarchives.org/vista/).

Results

Between December 30, 2013, and August 15, 2017, 506 eligible participants were randomized (MISTIE = 255; Medical = 251), of whom 499 patients (MISTIE = 250; Medical = 249) participated in the trial and constituted the mITT population set (median age 62, interquartile range 52–71 years; 61% male) (eTable 2, links.lww.com/WNL/D72).

Effect of MISTIE Intervention on 365-Day mRS Outcome (mITT Analyses)

Using the minimally informative prior, the mean RR (CrI) of achieving good 365-day mRS score in the MISTIE group was 1.06 (0.86–1.30), and the posterior probability that the MISTIE intervention has any beneficial effect on mRS score (RR >1) was 70% (Table 2 and Figure 1). In addition, the probabilities of observing RR >1.02 and RR >1.10 in the MISTIE group was 63% and 34%, respectively. The posterior probabilities of observing RR >1, RR >1.02, and RR >1.10, respectively, was 87%, 82%, and 55% with enthusiastic prior and 68%, 60%, and 30% with a skeptical prior.

Table 2.

Bayesian Estimation of the Posterior Probabilities for the Effect of MISTIE on 365-Day, 180-Day, and 30-Day mRS Score of 0–3 of Disaggregated by Varying Degrees of Prior Beliefs About the Treatment Effect

Prior belief	Posterior mean RR (95% credible interval)	Probability that the true RR is ≥prespecified threshold, %
		RR >1	RR >1.02	RR >1.10	RR >1.52
365-d mRS outcome
Reference priors
Minimally informative	1.06 (0.86–1.30)	70	63	34	0
Enthusiastic	1.12 (0.92–1.35)	87	82	55	0
Skeptical	1.05 (0.87–1.26)	68	60	30	0
Data-derived prior
No (0%) downweighing	1.07 (0.87–1.30)	73	66	38	0
50% downweighing	1.07 (0.87–1.30)	72	65	37	0
75% downweighing	1.06 (0.86–1.30)	70	64	35	0
180-d mRS outcome
No (0%) downweighing	1.05 (0.84–1.30)	65	59	32	0
50% downweighing	1.05 (0.83–1.29)	63	56	30	0
75% downweighing	1.04 (0.83–1.29)	62	55	29	0
30-d mRS outcome
No (0%) downweighing	1.26 (0.77–1.97)	80	78	67	18
50% downweighing	1.25 (0.76–1.97)	79	76	66	18
75% downweighing	1.25 (0.75–1.97)	78	76	65	18

Abbreviations: MISTIE = minimally invasive surgery with thrombolysis in intracerebral hemorrhage evacuation; mRS = modified Rankin Scale; RR = relative risk.

Note that Table 2 presents the effect of the MISTIE intervention on a positive outcome, that is, attaining mRS score of 0–3.

Figure 1. Probability Distributions for the RR of Attaining 365-Day mRS Score of 0–3 Among Patients Receiving MISTIE Intervention, by Reference and Data-Derived Priors.

This graph demonstrates the influence of the choice of prior distribution on the MISTIE-3 trial interpretation. The red lines represent the individual prior distributions. The posterior distribution is represented by the blue line, and the likelihood function is represented by the shaded area. Situated below each distribution set is a blue point representing the median RR and a line representing the 95% credible interval. Note that the likelihood function is similar across all prior distributions. Therefore, differences in the posterior distributions, displayed in each subgraph, are driven by the choice of prior distribution. MISTIE = minimally invasive surgery with thrombolysis in intracerebral hemorrhage evacuation; mRS = modified Rankin Scale; RR = relative risk.

With data-derived prior, the mean RR (CrI) of achieving good mRS score in the MISTIE group was 1.07 (0.87–1.30). The posterior probabilities of RR >1, RR >1.02, and RR >1.10, respectively, were 73%, 66%, and 38% (Table 2 and Figure 1). Both the treatment effect and posterior probabilities remained relatively similar across the spectrum of downweighing (Table 2, Figure 1, and eFigure 3, links.lww.com/WNL/D72). Of note, the probability of observing RR >1.52 was 0% with both reference and data-derived priors. Assuming a baseline rate of 42%, which is the proportion of patients in the control group who achieved a good 365-day mRS score, the probabilities of observing an ARD >1% with the minimally informative, enthusiastic, skeptical, and data-derived (without downweighing) priors were 61%, 81%, 59%, and 65%, respectively. Further details of the probabilities of observing specific thresholds of ARD are available in eTable 3 and eFigure 4.

Effect of MISTIE Intervention on 365-Day mRS Outcome (per Protocol Analyses)

In per protocol analysis, with minimally informative prior, the probabilities of observing RR >1, RR >1.02, and RR >1.10 for good 365-day mRS were 74%, 68%, and 40%, respectively. These probabilities for other priors were 89%, 85%, and 60% (enthusiastic prior); 74%, 66%, and 36% (skeptical prior); and 78%, 72%, and 44% (data-derived prior without downweighing) (eTables 4 and 5, links.lww.com/WNL/D72).

Other Secondary Analyses

The posterior probabilities that MISTIE intervention has any beneficial effect (RR >1) on achieving good 180-day and 30-day mRS scores, respectively, were 65% and 80% (Table 2). Furthermore, the posterior probabilities of observing RR >1.02 and RR >1.10, respectively, were 59% and 32% for 180-day mRS and 78% and 67% for 30-day mRS. However, the posterior probability of observing RR >1.52 for good 180-day mRS was 0%, while the corresponding probability for good 30-day mRS was 18% (see eTable 3, links.lww.com/WNL/D72, for probabilities of observing specific thresholds of ARD for both outcomes).

The posterior probabilities of observing any beneficial effect of MISTIE intervention on 365-, 180-, and 30-day mortality (RR <1), respectively, were 93%, 98%, and 99% (Table 3). The posterior probabilities that the treatment effect (RR) of MISTIE intervention is <0.9 for 365-, 180-, and 30-day mortality outcomes, respectively, were 79%, 93%, and 96%. Furthermore, the posterior probability of observing RR <0.7 for 30-day mortality was 76% (see eTable 6, links.lww.com/WNL/D72, for probabilities of observing specific thresholds of ARD for mortality outcomes).

Table 3.

Bayesian Estimation of the Posterior Probabilities for the Effect of MISTIE on Mortality Outcomes Using Priors Derived From MISTIE 2 Trial

Prior belief	Posterior mean RR (95% credible interval)	Probability that the true RR is ≥prespecified threshold, %
		RR <1	RR <0.9	RR <0.8	RR <0.7
365-d mortality
No downweighing	0.80 (0.57–1.08)	93	79	54	24
50% downweighing	0.79 (0.56–1.07)	94	81	57	26
75% downweighing	0.78 (0.55–1.07)	94	82	58	27
180-d mortality
No downweighing	0.70 (0.48–0.98)	98	93	80	53
50% downweighing	0.69 (0.47–0.97)	98	94	81	56
75% downweighing	0.68 (0.47–0.96)	99	94	83	58
30-d mortality
No downweighing	0.61 (0.37–0.94)	99	96	90	76
50% downweighing	0.62 (0.37–0.96)	98	95	88	73
75% downweighing	0.63 (0.38–0.98)	98	95	86	71

Abbreviations: MISTIE = minimally invasive surgery with thrombolysis in intracerebral hemorrhage evacuation; RR = relative risk.

Note that Table 3 presents the effect of the MISTIE intervention on a negative outcome, that is, mortality.

Of note, both treatment effects and posterior probabilities did not substantially change across the spectrum of downweighing (from 0% to 75% downweighing) for all secondary outcomes.

Discussion

In this study, we used the Bayesian approach for reanalysis of the MISTIE-3 trial data to assess the probabilities of treatment effect attributable to the MISTIE intervention on functional and survival outcomes. The MISTIE-3 trial did not demonstrate a statistically significantly improvement in functional outcome (365-day mRS score of 0–3) among ICH patients who received the MISTIE intervention. Of note, our analysis does not provide results that deviate from this original conclusion. However, we believe that the reliance on a frequentist-derived binary p value cutoff of 0.05 may preclude valuable clinical information. Specifically, reliance on p values may limit the interpretational value of efficacy data obtained from clinical trials, particularly for patient populations with limited treatment options, such as ICH. In our Bayesian reanalysis of the MISTIE-3 trial, we demonstrate that even with a skeptical prior—which assumes no treatment effect—the probability that the MISTIE intervention has any beneficial effect (RR >1) on 365-day mRS is 68%. We would like to highlight that the idea of such unplanned post hoc analyses is not to generate alternative conclusions but to explore application of Bayesian methods to clinical trial data. Therefore, we do not seek to claim that Bayesian analysis of the MISTIE trial would have resulted in a positive trial. However, we do believe that Bayesian interpretation may provide an enhanced ability to appreciate the potential of treatment effect both for the providers and patients and/or their families. This is specifically important for ICH, which is associated with very high mortality and morbidity and does not have any proven definitive treatment. Given the safety profile of the MISTIE-3 intervention, it becomes relevant to explore effectiveness using Bayesian approaches and provide a framework for future trials. Of note, Bayesian interpretation of effects for other clinical scenarios will require condition-specific targeted considerations.

A certain degree of controversy exists in the literature regarding Bayesian reanalysis of clinical trials. On one hand, Bayesian analyses provide a systematic and mathematical framework to incorporate information (prior beliefs and data) to yield an updated summary of knowledge without having the requirement that the conclusion is binary. On the other hand, this seemingly intuitive methodology of generating knowledge relies heavily on prior information, which may be subjective at times. A systematic review and meta-analysis evaluated the discordance between results generated from frequentist and Bayesian methods.³¹ Trials were considered to be “improbably beneficial” if the posterior probability of attaining an ARD greater than or equal to a minimum clinically important difference (MCID) is ≤50%.³¹ On the other hand, trials were considered “potentially beneficial” if the posterior probability of attaining ARD ≥ MCID is >50%. Using these criteria, 15% (12 of 78 studies) shifted from “improbable benefit” to “potential benefit” based on the choice of skeptical vs an enthusiastic prior. We want to note that we have not defined a MCID for our analyses. However, the posterior probability for a >10% benefit (RR >1.10) in 365-day mRS was between 34% and 38% with all priors other than the enthusiastic prior. Although our enthusiastic prior (which assumed that the median RR for MISTIE benefit is 1.52) yielded a 55% probability of observing RR >1.10, we do not consider our analyses to contradict the overall nonsignificant results of the MISTIE-3 trial, particularly given that the probability of observing RR >1.52 is 0 for all priors. Several additional considerations need to be incorporated in the application of such results, including but not limited to consensus around MCID and potential safety profile of the intervention.

The Bayesian framework treats parameter estimates, such as RR, as random variables with specific probability distributions, thereby making it possible to calculate probabilities of observing specific values of treatment effects, such as RR >1, RR >1.02, RR >1.1, and RR >1.52. In addition, distribution-based summary measures, such as mean and median RR, can be estimated using the Bayesian framework. By contrast, the frequentist's framework considers parameter estimates as fixed values arising from the observed data. Therefore, direct estimation of the probabilities of observing specific values of treatment effects is not possible under the frequentist paradigm.

Although Bayesian methods were first described in the 18th century,¹⁶ recent advances in computational capacity have led to the resurgence of the Bayesian approach as an alternative method for evaluating medical hypotheses. Despite its seemingly attractive application for interpretation of clinical data, a major limitation of Bayesian analyses continues to be the selection of an appropriate prior. In 2010, the FDA issued the “Guidance for the Use of Bayesian Statistics in Medical Device Clinical Trials.”²⁷ In this document, the FDA outlined the value of using multiple of priors. We used this approach and included uninformative (minimally informative), informative, and data-derived priors obtained through Bayesian hierarchical models. Since the publication of this guidance document, various studies have used Bayesian methods for post hoc analysis of clinical trials.^18,32-34 In 2019, for example, the FDA used a novel post hoc Bayesian analysis, using a combination of a skeptical prior and a data-derived prior borrowed from the adult population, to evaluate the efficacy of intravenous belimumab among children (aged 5–17 years) with systemic lupus erythematosus.³¹ Furthermore, the FDA recommended that studies used to construct priors should be similar to the current study with respect to their target population, study protocol, study end points, data collection time frame, investigators, and study sites.^8,33 Therefore, by using a variety of empirical priors and a data-derived prior (using MISITE-2 data) for Bayesian reanalyzes of the MISTIE-3 trial data, our work aligns well with the recommendations.

Our analyses indicate that the mean RR (CrI) of achieving a good 365-mRS score in the MISTIE group was 1.06 (0.86–1.30) with minimally informative prior, 1.12 (0.92–1.35) with enthusiastic prior, and 1.05 (0.87–1.26) with a skeptical prior. Using the mITT analysis, the probability of observing an ARD >1%, which is equivalent to 700 more ICH survivors attaining good functional recovery, was 81% using an enthusiastic prior, 61% using a minimally informative prior, 59% using a skeptical prior, and 65% using a data-derived prior. As expected, the probability estimates obtained with per protocol analysis were relatively higher than the corresponding estimates of mITT analysis. This finding confirms that procedural performance and achievement of the surgical aim of MISTIE intervention (postprocedural clot size ≤15 mL) may influence the effectiveness of MISTIE intervention. In addition, we observed that the posterior probabilities of treatment effects drop-off substantially at higher values of RR, indicating that the overall effect of the MISTIE intervention on mRS score may be, at best, modest. Furthermore, with the data-derived prior (without downweighing), the probability of any beneficial effect (RR >1) of MISTIE intervention on 30-day and 365-day mRS score was 80% and 73%, respectively. The longer-term dilution of the potential effectiveness of the MISTIE intervention is not surprising, given the complex nature of ICH, the intensity of ICU therapy, and the role of associated comorbidities and complications.^8,35

Although exploratory, the posterior probabilities of the effect of MISTIE intervention on mRS outcomes reported in this study are nontrivial. This is specifically relevant, given the severity of ICH, the lack of evidence-based treatments, and safety profile of the MISTIE intervention. Nevertheless, the exploratory nature of our work and the observed drop-off at higher effectiveness levels do not provide compelling support to ubiquitous adoption of the MISTIE intervention. An argument can conversely be made that given no other clinically proven treatments for ICH and MISTIE intervention being safe, these nontrivial probabilities may be “enough” to convince many. Previous studies indicate that surgeons are more likely to recommend craniotomy procedures if they believe that the prognosis of survival and functional outcomes is better.^36,37 Therefore, these findings may provide a reference that individual clinicians can use to apply their clinical judgment and make informed decisions regarding adoption of MISTIE intervention for patients with ICH, particularly in cases where a minimally invasive procedure is feasible and optimal hematoma evaluation can be achieved.^17,38

Much like the Decompressive Surgery for the Treatment of Malignant Infarction of the Middle Cerebral Artery trial,³⁹ which demonstrated a significant effect of decompressive craniotomy on mortality, but with no significant effect on functional outcomes, the MISTIE intervention showed a 93% probability of reducing 365-day mortality and a 99% probability of reducing 30-day mortality (RR <1). Moreover, the probability of attaining ARD >1% was 83% for 365-day mortality and 95% for 30-day mortality. It is pertinent to note that the MISTIE-3 trial was not powered to evaluate a mortality benefit. However, given the severity of ICH as well as the lack of treatment options conferring unequivocal clinical benefit, treatment options with substantial probability of benefit and negligible probability of harm may be worthy of consideration by clinicians and caregivers. Among patients with ICH, a survival benefit has been previously demonstrated to be a major determinant of patient-centered health utility.⁴⁰ The findings of this study may therefore provide additional context to the clinician-caregivers decision-making process regarding adoption and acceptance of MISTIE intervention for ICH management. In addition, the posterior probabilities estimated in this study may inform whether there is significant uncertainty about the effect of MISTIE intervention (equipoise) to warrant the development of new clinical trials investigating the effect of MISTIE on functional and survival outcomes.

A major underpinning and limitation of Bayesian analyses remains the selection of appropriate priors, as the posterior probabilities are largely influenced by such analytical decisions. We, therefore, used a broad range of priors, including data-driven ones, to assess the robustness of our estimates. Relatively stable and consistent posterior probabilities across these priors provide credence to our findings. In addition, MISTIE-3 was an open-labeled multisite trial, and the results of our analyses need to be interpreted in the light of heterogeneity observed in the original trial. Although MISTIE-3 enrolled a large number of patients for an ICH surgical intervention trial, the trial was conducted at selected sites. These sites may have had higher interventional capabilities along with the availability of enthusiastic and skilled interventionalists. Although optimal site selection allowed for successful implementation of the MISITE intervention in the trial, the wider generalizability of MISTIE intervention continues to be evaluated. Finally, MISTIE-3 was not powered to assess a mortality benefit, and hence, our post hoc analyses for mortality outcomes are hypothesis generating.

Across a range of Bayesian priors, the probability of MISTIE intervention having any beneficial effect on 365-day mRS for patients with ICH is between 68% and 87%. Bayesian-derived interpretations of clinical trial data may guide research priorities and provide a contextual framework for shared decision-making related to adoption of interventions for management of diseases, such as ICH, which carry high risk of poor outcomes and have no proven treatment modalities.

Glossary

ARD

absolute risk difference

CrI

credible interval

FDA

US Food and Drug Administration

ICH

intracerebral hemorrhage

MCID

minimum clinically important difference

MISTIE-3

minimally invasive surgery with thrombolysis in ICH evacuation-3

mITT

modified intention-to-treat

mRS

modified Rankin scale

RCT

randomized controlled trial

RR

relative risk

Appendix. Authors

Name	Location	Contribution
Abdulaziz T. Bako, MBBS, MPH, PhD	Department of Neurosurgery, Houston Methodist, TX	Drafting/revision of the manuscript for content, including medical writing for content; analysis or interpretation of data; study concept or design
Thomas Potter, PhD	Department of Neurosurgery, Houston Methodist, TX	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design
Alan P. Pan, MS	Center for Health Data Science and Analytics, Houston Methodist, TX	Drafting/revision of the manuscript for content, including medical writing for content; study concept or design
Jonika Tannous, PhD	Department of Neurosurgery, Houston Methodist, TX	Study concept or design; drafting/revision of the manuscript for content, including medical writing for content
Gavin Britz, MD, MPH, MBA	Center for Health Data Science and Analytics, Houston Methodist, TX; Weill Cornell Medical College, New York, NY; Houston Methodist Academic Institute, TX	Major role in the acquisition of data; drafting/revision of the manuscript for content, including medical writing for content
Wendy C. Ziai, MD, MPH	Division of Brain Injury Outcomes, Johns Hopkins University, Baltimore, MD	Major role in the acquisition of data; drafting/revision of the manuscript for content, including medical writing for content; study concept or design
Issam Awad, MD	Department of Neurological Surgery, University of Chicago Medicine and Biological Sciences, Chicago, IL	Drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design
Daniel Hanley, MD	Division of Brain Injury Outcomes, Johns Hopkins University, Baltimore, MD	Study concept or design; drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data
Farhaan S. Vahidy, MBBS, MPH, PhD	Department of Neurosurgery, and Center for Health Data Science and Analytics, Houston Methodist, TX; Weill Cornell Medical College, New York, NY; Houston Methodist Academic Institute, TX	Study concept or design; analysis or interpretation of data; major role in the acquisition of data; drafting/revision of the manuscript for content, including medical writing for content

Study Funding

MISTIE III trial was funded by a NIH National Institute of Neurological Disorders and Stroke grant (Grant No. 1U01NS080824-01A1; Trial Innovation Center Grant No. 1U24TR001609-01) and received material support from Genentech.

Disclosure

The authors report no disclosures relevant to the manuscript. Go to Neurology.org/N for full disclosures.