Bayesian Outcome Prediction After Resuscitation From Cardiac Arrest

Abstract

Background and Objectives

Postarrest prognostication research does not typically account for the sequential nature of real-life data acquisition and interpretation and reports nonintuitive estimates of uncertainty. Bayesian approaches offer advantages well suited to prognostication. We used Bayesian regression to explore the usefulness of sequential prognostic indicators in the context of prior knowledge and compared this with a guideline-concordant algorithm.

Methods

We included patients hospitalized at a single center after cardiac arrest. We extracted prospective data and assumed these data accrued over time as in routine practice. We considered predictors demographic and arrest characteristics, initial and daily neurologic examination, laboratory results, therapeutic interventions, brain imaging, and EEG. We fit Bayesian hierarchical generalized linear multivariate models predicting discharge Cerebral Performance Category (CPC) 4 or 5 (poor outcomes) vs 1–3 including sequential clinical and prognostic data. We explored outcome posterior probability distributions (PPDs) for individual patients and overall. As a comparator, we applied the 2021 European Resuscitation Council and European Society of Intensive Care Medicine (ERC/ESICM) guidelines.

Results

We included 2,692 patients of whom 864 (35%) were discharged with a CPC 1–3. Patients' outcome PPDs became narrow and shifted toward 0 or 1 as sequentially acquired information was added to models. These changes were largest after arrest characteristics and initial neurologic examination were included. Using information typically available at or before intensive care unit admission, sensitivity predicting poor outcome was 51% with a 0.6% false-positive rate. In our most comprehensive model, sensitivity for poor outcome prediction was 76% with 0.6% false-positive rate (FPR). The ERC/ESICM algorithm applied to 547 of 2,692 patients and yielded 36% sensitivity with 0% FPR.

Discussion

Bayesian models offer advantages well suited to prognostication research. On balance, our findings support the view that in expert hands, accurate neurologic prognostication is possible in many cases before 72 hours postarrest. Although we caution against early withdrawal of life-sustaining therapies, rapid outcome prediction can inform clinical decision making and future clinical trials.

Guidelines recommend that clinical providers use multiple diagnostic modalities to inform neurologic prognostication after cardiac arrest.¹ Prognostication research typically describes the association of single prognostic indicators or limited combinations of indicators with patient outcome, assumes that prognostic information is known simultaneously, and reports performance metrics such as sensitivity and false-positive rate (FPR).^2-8 These approaches ignore the sequential nature of real-life data acquisition and assume that test acquisition and results are not influenced by prior clinical knowledge. Estimates of uncertainty (e.g., CI) diverge from clinical intuition and rely on approximations that may not hold.⁹ Realistic descriptions of how precision or uncertainty of prognostic testing changes over time are important to inform providers for real-world clinical decision making.

Bayesian analyses offer advantages particularly well suited to describing prognostication. These include more rigorous treatment and intuitive interpretations of uncertainty, ability to incorporate prior knowledge, and flexibility to handle hierarchical models and missing data.⁹ We leveraged a large data set and used Bayesian regression to explore the usefulness of sequential prognostic indicators in the context of prior clinical knowledge, with particular focus on quantitative EEG trajectories that past research suggests improve outcome prediction.^3,10,11 We compared this approach with a standard guideline-concordant prognostication algorithm.¹²

Methods

Standard Protocol Approvals, Registrations, and Patient Consents

The University of Pittsburgh Human Research Protection Office approved this research with a waiver of informed consent (STUDY19020205).

Patients

We included consecutive patients hospitalized at a single academic medical center after in-hospital or out-of-hospital cardiac arrest from January 1, 2010, to December 31, 2019. We excluded patients who were arrested due to primary neurologic or traumatic etiologies, those who were transferred more than 24 hours postarrest, and those for whom time of arrest or hospital arrival was unknown.

Predictors

We extracted data both from our prospective registry and, as an automated retrieval, from the electronic health record. As predictors, we included variables with established clinical relevance or association with outcomes from prior prognostication literature and that at our hospital are acquired as standard of care in most patients. These were selected a priori and included patient age, sex, arrest location (in-hospital or out-of-hospital and treating emergency department arrests as in-hospital¹³), presenting arrest rhythm (ventricular tachycardia/fibrillation, pulseless electrical activity, asystole, or unknown), interfacility transfer status, witnessed collapse, layperson cardiopulmonary resuscitation (CPR), prehospital return of spontaneous circulation, shocks delivered from an automated external defibrillator (yes or no), CPR duration in minutes, number of intra-arrest epinephrine boluses administered, and Charlson Comorbidity Index. We also considered elements of the neurologic examination on presentation, day 1, and day 2, including motor examination summarized by the Full Outline of UnResponsiveness score (follows commands, localizing to pain, flexion/withdrawal, extension, no motor response, or myoclonus), pupillary light reflex (present in 1 or both eyes or bilaterally absent), corneal reflex (present in 1 or both eyes or bilaterally absent), gag reflex, and cough reflex. In addition, on presentation, day 1, and day 2, we considered peak serum or whole blood lactate level and Sequential Organ Failure Assessment cardiovascular, renal, and respiratory subscales. We considered initial therapeutic interventions including the provision of percutaneous coronary angiography and targeted temperature management (TTM) strategy (36C, 33C, no TTM, or other TTM strategy). Finally, we considered early prognostic test results including gray matter-to-white matter attenuation ratio (GWR) measured on brain computed tomography (CT), which we calculated at the level of the basal ganglia as previously described¹⁴; initial EEG background (continuous or near-continuous, discontinuous or burst suppression, or generalized suppression); epileptiform activity on initial EEG (at least some cortical background activity [i.e., not generalized suppression] without epileptiform findings or only nonperiodic epileptiform discharges, any other epileptiform activity including burst suppression with identical bursts, or generalized suppression), defined as previously described and consistent with consensus guidelines^2,15; and initial EEG reactivity during a standardized stimulation protocol. In patients with coma, the EEG technologist performs auditory and tactile stimulation on EEG initiation and once daily thereafter. Finally, we fit a 4-group trajectory model of hourly median suppression ratio from EEG initiation to hour 48 using quadratic polynomials and a beta distribution, as previously described,^3,10,11 from which we extracted patient-specific posterior probabilities of group membership (PPGMs).

We assumed clinical and prognostic data were acquired over time in an order that resembles routine clinical practice (Table 1). When a data element for a patient was obtained >6 hours later than the estimated time frame, we treated it as missing in that time point's model.

Table 1.

Predictors Entered Into Hierarchical Modeling

Outcomes

Our primary outcome of interest was discharge Cerebral Performance Category, (CPC) which we dichotomized as 1 to 3 (i.e., awake and alive) vs 4 or 5 (i.e., unconscious or deceased). In survivors, CPC evolves over time after cardiac arrest. In our system of care, recovery to a CPC of 3 is generally viewed as sufficient to allow discharge to a postacute care setting (e.g., inpatient rehabilitation). Many of these patients are long-term survivors and demonstrated improvement to 1 or 2 at 6 and 12 months,¹⁶ a finding that has been replicated by others.¹⁷

Modeling and Statistical Methods

We used R (R Foundation for Statistical Computing, Vienna, Austria) for all analyses. We used the mice package¹⁸ to perform multiple imputation with chained equations to handle missing data and created 100 imputed data sets. We used the brms package¹⁹ to fit outcome models at each time point. To inform our initial prior distribution for the outcome, we examined 602 cases treated at our center from 2005 to 2009 of whom 26.3% experienced a positive outcome and so used a normal distribution in the logit scale with mean logit (0.263) and standard error for the logit (p) estimate of 0.0926. We used default priors for model coefficients. We fit models using the Markov chain Monte Carlo algorithms with 800,000 to 2.1 million chains depending on model complexity. We assessed convergence by verifying that the R-hat values for each imputed data set were close to 1.

Our primary goal was to explore the usefulness of sequential prognostic indicators in the context of prior clinical knowledge. To this end, we considered several clinically relevant metrics of model performance. First, we examined the mode of each patient's outcome posterior probability distribution (PPD) as a point estimate. We created histograms of these estimates from each model to describe the population-level effect of sequentially acquired information. We also calculated sensitivity and FPR for each model when predicting poor outcomes at mode outcome probability thresholds <0.05 and <0.01. To calculate these metrics, we assumed that predicted poor prognosis might prompt the withdrawal of life-sustaining therapies (WLSTs) and so classified patients as predicted to have poor outcome if any previous model's posterior was below threshold. As a secondary analysis, we considered metrics for each model in isolation. For each model, we also identified patients who newly crossed below these thresholds (i.e., those for whom the prior probability distribution had a mode ≥0.05 and the PPD had a mode <0.05) and explored the prognostic indicator(s) that drove this shift. In our main analysis, we did not include patients who had experienced outcome (e.g., died) antecedent to a particular model's estimated time point (Table 1). In a secondary analysis, we included these patients describing their outcome PPD as a spike function at 0 or 1. We specifically explored the incremental value of adding PPGMs from our qEEG group–based trajectory model. We compared the modes of each patient's outcome PPD with and without this information, the width of the 95% credible intervals of these PPDs, and sensitivity and FPR at mode outcome probability thresholds <0.05 and <0.01. We selected case examples and plotted priors and PPDs to graphically depict the effect of GBTMs on patient-level predictions. Finally, as a comparator, we applied the 2021 European Resuscitation Council and European Society of Intensive Care Medicine (ERC/ESICM) guidelines for postarrest prognostication to our cohort and report results.

Data Availability

Anonymized data not published within this article will be made available upon request from qualified investigators for purposes of replicating procedures and results.

Results

Overall, 2,692 patients presented during the study period after resuscitation from cardiac arrest, of whom 73 arrived through interfacility transfer >24 hours or uncertain duration postarrest, 121 arrested from traumatic or neurologic causes, and 19 arrived with CPR ongoing due to prehospital rearrest and never regained pulses. Of 2,479 patients included, 864 (35%) were discharged with a CPC 1–3.

As expected, modes of patients' outcome PPDs were initially in an intermediate range and gradually shifted toward 0 or 1 as additional prognostic data became available (Figures 1 and 2). This shift was most prominent after the initial examination data were ascertained. In parallel, sensitivity predicting poor outcome at set outcome PPD thresholds increased over time. In our most comprehensive model including data available through intensive care unit (ICU) day 2, sensitivity predicting poor outcome was 0.76 at an FPR of 0.006 (Table 2).

Figure 1. Patients' Outcome Posterior Probability Distribution Modes, Stratified by Eventual Outcome, From Sequential Models Based on Data Available: (A) Before Arrival, (B) After Initial History, (C) After Initial Examination, (D) After Emergent Diagnostics and Management, (E) On ICU Arrival, and (F) On ICU Arrival With PPGMs.

Patients who experienced an outcome before each time point are not included in these plots. All plots through ICU day 2 are available in eFigure 2 (links.lww.com/WNL/C159). ICU = intensive care unit.

Figure 2. Alluvial-Style Plot Showing the Shift in (A) the Mode of Each Patient's Outcome Posterior Probability Distribution and (B) the Mode of Nonsurvivors’ Outcome Posterior Probability Distributions in Sequential Models.

Vertical boxes depict categories of outcome probability and are scaled to reflect the size of the cohort within each bin at each model time point. (A) Data from patients with favorable outcome are depicted in blue, and data from those with poor outcome are depicted in green. ICU = intensive care unit.

Table 2.

Cumulative Model Performance Metrics Predicting Poor Outcome (CPC 4 or 5)

Addition of PPGMs to other available prognostic information had little effect on sensitivity and FPR (Table 2), the mode of patient outcome PPDs (Figure 3), the shape of these PPDs (Figure 4), or the width of the 95% credible intervals (eFigure 1, links.lww.com/WNL/C159). In a small number of patients, addition of PPGMs substantially shifted their outcome posterior probability distributions toward zero (Figures 3 and 4).

Figure 3. Shift in the Mode of Each Patient Outcome Posterior Probability Distribution From Addition of Trajectory Model Output (A) on ICU Admission, (B) On ICU Day 1, and (C) on ICU Day 2.

.ICU = intensive care unit.

Figure 4. Specific Examples of Patients' Outcome Posterior Probability Distributions Over Time.

Patient 1: On ICU arrival, patient 1 had low outcome probability distribution driven by presenting pulseless electrical activity rhythm, 54 minutes of CPR and absent motor examination, and brainstem reflexes on arrival. EEG showed a continuous background with reactivity and significantly improved outcome predictions. Patient 2: On ICU arrival, patient 2 had an intermediate outcome probability distribution. qEEG showed high suppression ratio and substantially lowered the estimated recovery probability. Patient 3: Patient 3 had an extremely poor predicted outcome driven by diffuse cerebral edema on head CT. No other prognostic data meaningfully altered the posteriors.

Overall, 311 patients had a mode outcome PPD of <0.05 after initial history alone. These were almost exclusively resuscitated from prolonged (median 45 minutes [interquartile range 38–60 minutes]) out-of-hospital cardiac arrests with either initial asystole or multiple unfavorable arrest characteristics (e.g., unwitnessed collapse without layperson CPR). An additional 407 patients had a mode outcome PPD <0.05 after the initial examination, which was driven by a combination of either myoclonus or absent motor response to pain and multiple absent brainstem reflexes. After emergent diagnostics and management, only 80 additional patients had a mode outcome PPD <0.05, which was driven by severe diffuse cerebral edema on CT imaging. Overall, 273 patients had severe cerebral edema on CT imaging, most of whom already had a mode outcome PPD <0.05 based on other factors. After initial ICU diagnostics, an additional 180 patients had a mode outcome PPD <0.05, of whom 130 had burst suppression with epileptiform bursts and 50 had generalized suppression. Overall, 319 patients had burst suppression with epileptiform bursts and 372 patients had generalized suppression, most of whom already had a mode outcome PPD <0.05 based on other factors.

Of 2,479 included patients, 912 (37%) were awake on arrival or awakened <72 hours postarrest. Of those these, 783 (81%) were discharged with CPC 1–3. An additional 912 (37%) patients died <72 hours post-arrest. For these groups, the ERC/ESICM prognostic algorithm does not apply. Of the remaining 655 patients who remained alive and comatose 72 hours postarrest, 78 (12%) could not have an accurate neurologic assessment because they required continuous sedation or neuromuscular blockade and 30 (5%) had Glasgow Motor Score >3, leaving 547 eligible for algorithmic prognostication (22% of the overall cohort). Of these, 79 (14%) were discharged with CPC 1–3, sensitivity of the ERC/ESICM algorithm predicting that the poor outcome was 36% (95% CI 32%–40%), and there were no false positives.

Discussion

We applied a Bayesian approach to explore the influence of sequentially acquired information on prognostication in a framework that parallels clinical practice. The approach offers conceptually intuitive and easy-to-interpret results at both population and individual levels. It is unsurprising that patients with established predictors of poor outcome such as severe cerebral edema or burst suppression with epileptiform bursts had recovery estimates near zero.^1,2,14,20 Many of these patients already had near-zero priors because of previously available prognostic data, limiting the additional information gleaned from these tests. This is expected given different clinical measures of the same underlying brain injury are strongly correlated after cardiac arrest²¹ but speaks to the importance rigorously evaluating the incremental prognostic information gained from a novel tool in the context of other available knowledge. Sensitivity of our model was substantially higher than guideline-recommended prognostic algorithms, while maintaining extremely low false-positive rates.⁸

Most (51%) nonsurvivors had recovery probabilities <1% based on information available at or before ICU admission, far sooner that current guidelines recommend predictions be made. This likely reflects several factors. First, we included both a greater number and complexity of predictors than accounted for in guidelines,¹ which are crafted in part to optimize simplicity and ease of use. Second, we incorporated prior knowledge of expected postarrest outcomes. Finally, patients received somewhat regimented care in this single-center cohort.

Our work has relevance to both clinical care and research. Clinically, our findings support the view that in expert hands, sequential multimodality testing often allows accurate neurologic prognostication before 72 hours postarrest.^22,23 In many cases, baseline demographics, history, and presenting examination alone differentiate patients with recoverable postarrest illness from those who will not recover. Although we caution against WLSTs at this early time point, early outcome prediction can nevertheless inform clinical decision making. For example, resource-intensive interventions such as mechanical circulatory support or transfer to specialty care could be targeted to patients most likely to derive benefit. From the perspective of clinical research, past neutral trials of postarrest interventions such as temperature control or coronary angiography have typically selected patients based on arrest location, presence of coma on arrival, or initial arrest rhythm.^24-26 Our work suggests that a more nuanced approach incorporating presenting neurologic examination and measures of cardiopulmonary failure may be a useful approach to enrich trials for likely treatment responders.

In contrast to prior work by our group,^3,10,11 addition of patients' PPGMs from a group-based trajectory model of suppression ratio did not meaningfully alter outcome prediction. In this study, we considered many additional data elements, notably including expert EEG interpretations. Although suppression ratio trajectories are strongly associated with outcome, this reduced the incremental information they contributed. Trajectory models based on other qEEG features may be more useful in this setting, an area of active investigation by our group.

Our study has several important limitations. As noted above, the single-center design reduces generalizability of our findings. Because we used multiple imputation to handle missing data and examined pooled estimates, our analyses may overestimate uncertainty and thus failed to detect a true effect of adding qEEG on patient outcome estimates. Our goals in this analysis were descriptive, and so we quantified overall model performance rather than performing training and test set splits. This might have increased performance compared with what would have been observed using an out-of-sample test set. Although covariates we included in models all have established relationships with outcome after resuscitation from cardiac arrest, not all are included in consensus guidelines. In some cases, this is because, although associated with outcome, they lack highly specific cutoffs (e.g., age and arrest duration). In other cases, this is because of lack of broad accessibility (e.g., qEEG) or validation in multicenter cohorts (e.g., GWR).

As in most studies of postarrest outcomes, there is risk we included patients who would have recovered in the absence of WLST. Insofar because WLST was consistently applied based on observed clinical data, this would also increase the apparent sensitivity of our models. This introduces potential that use of our results to guide future WLST may perpetuate self-fulfilling prophecies whereby clinical findings that triggered WLST in the past could lead to overly pessimistic outcome predictions and avoidable deaths. CPC is an imperfect outcome scale and overlooks many clinically important domains of postarrest recovery, and assessment at hospital discharge does not account for continued recovery over time.^16,27 To address the latter limitation, because long-term outcome assessments were unavailable, we considered CPC 3 patients to have favorable recoveries because many will have a CPC of 1 or 2 after 6 months.¹⁶ This means our findings do not directly map to studies that have considered CPC 3 to be unfavorable. Finally, although Bayesian methods have advantages compared with the frequentist approach, they are computationally intensive to fit (for example, requiring several days to estimate per model). This cost is falling over time with technological advances but still limits the practicability of our approach.

In conclusion, Bayesian regression models offer advantages that are particularly well suited to prognostication research. On balance, our findings support the view that in expert hands accurate neurologic prognostication is possible in many cases before 72 hours postarrest. Early outcome prediction may help guide provision of resource-intensive interventions and enrich clinical trials for patients likely to be treatment responsive.

Glossary

CPR

cardiopulmonary resuscitation

ERC

European Resuscitation Council

ESICM

European Society of Intensive Care Medicine

FPR

false-positive rate

ICU

intensive care unit

PPD

posterior probability distribution

PPGM

posterior probabilities of group membership

qEEG

quantitative electroencephalography

TTM

targeted temperature management

WLST

withdrawal of life-sustaining therapies

Appendix 1. Authors

Appendix 2. Coinvestigators

Study Funding

No targeted funding reported.

Disclosure

J. Elmer's research time is supported by the NIH through grant 5K23NS097629. The other authors report no relevant disclosures. Go to Neurology.org/N for full disclosures.