Highlights
-
Neuroprognostication after cardiac arrest is challenging.
-
A multimodal approach may improve the false positive rate in predicting poor neurological outcomes of post-cardiac arrest patients.
-
The ideal number and combination of modalities predicting poor neurological outcomes remains unknown.
Introduction
Cardiac arrest is a major health problem worldwide and is associated with substantial mortality and morbidity. As few as 10% of patients survive to hospital discharge.Reference Yan, Gan and Jiang1 Most deaths in post-cardiac arrest patients occur following withdrawal of life-sustaining measures (WLSM) because of a predicted poor neurological prognosis.Reference Sandroni, Cronberg and Sekhon2 However, advances in management have increased overall survival and rate of discharge with a favorable neurological outcome.Reference Rossetti, Rabinstein and Oddo3 As such, consistent, objective and evidence-based neuroprognostication is crucial to avoid inappropriate or premature WLSM. Accurate neuroprognostication can also circumvent prolonged invasive, potentially harmful and costly therapies that could perpetuate patient and family suffering when there is no realistic chance of a favorable recovery.
The major determinant of prognosis after cardiac arrest remains brain injury related to global ischemia-reperfusion.Reference Rossetti, Rabinstein and Oddo3 Assessing the extent of injury and therefore informing prognosis has been a focus of many studies that have investigated the utility of physical examination, neurophysiological testing, serum biomarkers and neuroimaging.Reference Rossetti, Rabinstein and Oddo3 However, no single modality has been able to predict a poor outcome for patients with perfect specificity. In addition, all modalities have lacked sensitivity for predicting poor prognosis.Reference Elmer, Torres and Aufderheide4–Reference Geocadin, Callaway and Fink7
To address these concerns, contemporary guidelinesReference Fordyce, Kramer and Ainsworth8–Reference Rajajee, Muehlschlegel and Wartenberg11 outline an approach to neuroprognostication that ensures confounders are excluded and sufficient time has passed, while emphasizing a multimodal approach. However, it is still unknown which modalities are best to combine and whether this approach is superior to unimodal assessments. The main objective of this systematic review was to compare the diagnostic accuracy of multimodal approaches versus unimodal tests in the prediction of poor neurological outcomes at hospital discharge and beyond for adult patients who remain comatose post-cardiac arrest.
Methods
This systematic review was conducted according to established methodological standards and reported in accordance with the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.Reference Page, Moher and Bossuyt12 The protocol was registered in PROSPERO (CRD42022331283) on October 31, 2022.
Literature search
In consultation with a medical librarian, a search strategy was developed (Appendix 1). The search was executed on January 18, 2024, and included three databases (MEDLINE, EMBASE and CINAHL). There were no date or language restrictions. Subject headings (controlled vocabulary) and various full and truncated keywords were combined using Boolean operators and included heart arrest, out-of-hospital cardiac arrest, post-cardiac arrest, prognostication, neuroprognostic, multimodality, multimodal, modalities, pupillary reflex, corneal reflex, electroencephalography (EEG), somatosensory evoked potentials (SSEP), brain CT, brain MRI, neuron-specific enolase (NSE), modified Rankin Scale (mRS), Glasgow Outcome Scale (GOS), treatment outcome and neurological outcomes adult, among others. Reference lists of all papers eligible for full-text review were manually searched to identify additional studies. References were exported and managed using Covidence (Melbourne, Australia).13
Eligibility criteria
Patient population
The population of interest was adult (≥18 years old) post-cardiac arrest (in or out-of-hospital) patients. We excluded studies involving only pediatric patients, but those with over 50% adult patients, in addition to pediatric patients aged 16 years or older, were judged acceptable for inclusion. Studies were included regardless of whether patients underwent targeted temperature management (TTM).
Modalities of interest and defined thresholds predictive of a poor prognosis
We utilized contemporary guidelinesReference Fordyce, Kramer and Ainsworth8–Reference Rajajee, Muehlschlegel and Wartenberg11 to define modalities of interest and thresholds predictive of a poor prognosis (Appendix 2). Specific criteria for malignant patterns on EEG were defined according to the American Clinical Neurophysiology Society Critical Care EEG TerminologyReference Hirsch, Fong and Leitinger14 (Table S1). Most studies reported on modalities at multiple time points, which were tracked, even if the timing post-return of spontaneous circulation (ROSC) was not recommended in contemporary guidelines.Reference Fordyce, Kramer and Ainsworth8–Reference Rajajee, Muehlschlegel and Wartenberg11
Multimodal definition
Studies included had to present the prognostic accuracy of a multimodal combination, defined as the use of two or more modalities recommended for unimodal testing in at least one contemporary guidelineReference Fordyce, Kramer and Ainsworth8–Reference Rajajee, Muehlschlegel and Wartenberg11 (Table S2). To strengthen our multimodal definition, the tests included had to assess different anatomic or physiologic parameters. Therefore, we excluded studies that only combined CT and MRI. We considered EEG and SSEP as assessing different anatomic and physiologic parameters, as well as myoclonus, pupillary light reflex, corneal reflex and the motor component of the Glasgow Coma Scale (GCS-M). Finally, to be considered multimodal, studies had to combine the modalities, meaning that patients must have been assessed with both (not either) modalities. Thus, studies not meeting this requirement in their multimodal definitionReference Glimmerveen, Ruijter, Keijzer, Tjepkema-Cloostermans, van Putten and Hofmeijer15–Reference Lee, Jeung, Lee, Jung and Lee17 as well as studies using a Classification and Regression Tree (CART) analysisReference Kim, You and Lee18,Reference Peluso, Attanasio and Annoni19 were excluded.
Unimodal definition
We planned to compare multimodal prognostic accuracy to the prognostic accuracy of each individual modality from the same study. Most studies reported on individual modalities at various time points post-arrest. For comparison purposes, we selected the time point for individual modalities with the lowest false positive rate (FPR). If a different threshold for EEG Reference Bongiovanni, Romagnosi and Barbella20,Reference Moseby-Knappe, Westhall and Backman21 or NSEReference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22 was used for a modality when used alone versus in combination, the threshold with the higher specificity or sensitivity was used in the unimodal analysis. The same approach was used for multimodal tests within each study. GCS-M was not included in the unimodal data analysis as it is no longer approved in recent guidelines to be used in isolation,Reference Fordyce, Kramer and Ainsworth8,Reference Nolan, Bttiger B.W. and Cariou9,Reference Nolan, Cariou and Cronberg23 unless it was used in the multimodal comparator within the same study.Reference Kim, Park and Chung24–Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 However, if GCS-M was part of the entry point of the ERC-ESCIM algorithm, GCS-M was not used in the unimodal analysis as it was not considered part of the multimodal assessment.Reference Moseby-Knappe, Westhall and Backman21,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28
Outcomes
We included studies assessing unfavorable neurological outcomes at hospital discharge or beyond. In studies that reported outcomes at multiple time points, the longest time post-ROSC was chosen for data extraction. The Cerebral Performance Category (CPC) score is the most used in cardiac arrest literature (Table S3).Reference Group29 For the purposes of our review, we only included studies that defined poor outcome as CPC 3–5 and favorable outcome as CPC 1–2.Reference Geocadin, Callaway and Fink7 If other scores were used, studies had to meet the predefined classification of poor outcomes to be included: mRS 4–6, GOS 1–3 and Glasgow Outcome Scale Extended 1–4.Reference Geocadin, Callaway and Fink7 We excluded studies that did not use these dichotomizations, unless we were able to obtain the necessary raw data from the authors to reclassify patient outcomes.
Our primary outcome was the FPR with confidence intervals of individual and combinations of tests, in addition to specificity, sensitivity and positive/negative predictive values. For inclusion, studies had to at least report the FPRs or provide sufficient data to allow for these to be calculated. Prior to study exclusion for insufficient data, authors were contacted.
Study selection process
All titles and abstracts were independently screened in duplicate by four reviewers (AB, RH, CB and JK) to identify potentially relevant studies. Abstracts identified by a single reviewer as meeting inclusion criteria were moved to full-text review. Full-text articles were subsequently reviewed in duplicate by three reviewers (AB, RH and CB). Disagreements for eligibility were resolved by the involvement of a third reviewer (JK or AK). In addition to the above outlined inclusion and exclusion criteria, studies were also excluded if they were not original research or when the reference found was a conference abstract only with no corresponding peer-reviewed publication.
Bias assessment
Two blinded reviewers (AB and CB) independently assessed the quality of included studies using the QUIPS tool for systematic reviews of prognostic studies.Reference Grooten, Tseli and Äng Börn30 Each of the six criteria (study participation, study attrition, prognostic factor measurement, outcome measurement, study confounding and statistical analysis and reporting) has been rated as low, moderate or high risk.Reference Hayden, van der Windt, Cartwright, Côté and Bombardier31,Reference HAYDEN, Côté and Bombardier32 Disagreements were resolved by the involvement of a third reviewer (JK).
Data extraction and synthesis
All data from included studies were independently extracted and agreed upon in duplicate by two reviewers (AB and CB) using a standardized Microsoft ExcelReference Corporation33 data form created by the study team. Disagreements were resolved by the involvement of a third reviewer (JK or AK). Appendix 3 provides a list of all extracted data.
A meta-analysis was originally planned to use a primary meta-regression to compare specificity between unimodal and multimodal neuroprognostication analyses. A secondary analysis using a similar approach was attempted for sensitivity. A sensitivity analysis was attempted to exclude GCS-M from the primary and secondary analyses as it is no longer a recommended modality in current neuroprognostication guidelines.Reference Fordyce, Kramer and Ainsworth8–Reference Rajajee, Muehlschlegel and Wartenberg11 Heterogeneity was assessed using the I 2 statistic and publication bias using funnel plots, Begg’s testReference Begg and Mazumdar34 and Egger’s test.Reference Egger, Smith, Schneider and Minder35
Results
Search results and study selection
A total of 791 studies were identified, 93 of which were duplicates. A total of 698 abstracts were screened and, 362 were excluded. The remaining 336 studies underwent full-text review, with 324 subsequently excluded for reasons outlined in Figure 1. When two or more studies involved patients from the same population and assessed similar multimodal and unimodal approaches, only the study with the larger number of patients was included. After applying inclusion and exclusion criteria, 12 studies remained.
Flowchart of study selection.
Study characteristics
The characteristics of the 12 included studies are presented in Table 1. There were 4124 patients in total, of which 72% were male,Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24–Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36–Reference Son, Lee and Park39 and the mean age was 61.4 (SD 4.0) years.Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Roger, Palmier and Louart25,Reference Pouplet, Martin and Colin27,Reference Scarpino, Lolli and Lanzo37–Reference Son, Lee and Park39 Three studies reported median ages with interquartile ranges for patients with good and poor outcomes,Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36 and when this dichotomy was used, the median age was 60 (IQR 58.5–61.3) years for patients with good outcomes and 64 (IQR 62.5–64.5) with poor outcomes. Most patients (95%) suffered an out-of-hospital cardiac arrest. Some studies did not report the cause of arrest or the rhythm during cardiac arrest.Reference Moseby-Knappe, Westhall and Backman21,Reference Oddo, Sandroni and Citerio36 When this information was provided, most patients had a primary cardiac etiology. Average time to ROSC varied between studies from 15 to 30 min, with only a few studies reporting no flow and low flow times. All but two studies used TTM (temperature goal ranging from 32°C to 36°C) in 100% of the patients;Reference Bongiovanni, Romagnosi and Barbella20,Reference Moseby-Knappe, Westhall and Backman21,Reference Kim, Park and Chung24–Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36,Reference Ben-Hamouda, Ltaief and Kirsch38,Reference Son, Lee and Park39 TTM was used in 45%Reference Scarpino, Lolli and Lanzo37 and 60%Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22 in those studies. CPC was the outcome score used in 11 of 12 studies, reported at either 320,24,27,38,39 or 6 monthsReference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Roger, Palmier and Louart25,Reference Youn, Park and Kim28,Reference Scarpino, Lolli and Lanzo37 post-discharge, whereas GOS at 3 months was used in the remaining study.Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 Some studies did not provide any details about causes of death.Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26,Reference Oddo, Sandroni and Citerio36,Reference Son, Lee and Park39 WLSM due to a perceived poor prognosis was the main cause of death reported in all but two of the remaining studies; one study excluded patients with WLSM,Reference Youn, Park and Kim28 whereas another did not permit WLSM in their protocol.Reference Scarpino, Lolli and Lanzo37 Details about when decisions regarding WLSM were made were available for two studies.Reference Bongiovanni, Romagnosi and Barbella20,Reference Pouplet, Martin and Colin27 However, when actual numbers were given, between 24% and 52%Reference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Roger, Palmier and Louart25 of all included patients underwent WLSM. Details about the causes of death, and other management data are provided in the supplementary data (Table S4).
Study characteristics
Note: GO = good outcome; PO = poor outcome; OHCA = out-of-hospital cardiac arrest; ROSC = return of spontaneous circulation; TTM = targeted temperature management; CPC score = Cerebral Performance Category score; GOS = Glasgow outcome score; NR = not reported.
Study quality assessment
Risk of bias for study participation was rated as low in nine studiesReference Bongiovanni, Romagnosi and Barbella20,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36–Reference Son, Lee and Park39 and moderate in the remaining three.Reference Moseby-Knappe, Westhall and Backman21,Reference Roger, Palmier and Louart25,Reference Pouplet, Martin and Colin27 Study attrition bias was lowReference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Roger, Palmier and Louart25,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28,Reference Scarpino, Lolli and Lanzo37–Reference Son, Lee and Park39 in eight studies and moderate in the remaining four.Reference Bongiovanni, Romagnosi and Barbella20,Reference Moseby-Knappe, Westhall and Backman21,Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26,Reference Oddo, Sandroni and Citerio36 Prognostic factor measurement bias was low in four,Reference Bongiovanni, Romagnosi and Barbella20,Reference Youn, Park and Kim28,Reference Scarpino, Lolli and Lanzo37,Reference Son, Lee and Park39 moderate in sevenReference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Roger, Palmier and Louart25,Reference Pouplet, Martin and Colin27,Reference Oddo, Sandroni and Citerio36,Reference Ben-Hamouda, Ltaief and Kirsch38 and highReference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 in one study. Most studies had a low risk of bias pertaining to outcome measurements,Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Roger, Palmier and Louart25–Reference Pouplet, Martin and Colin27,Reference Oddo, Sandroni and Citerio36,Reference Scarpino, Lolli and Lanzo37,Reference Son, Lee and Park39 although three studies had a moderate risk of bias.Reference Kim, Park and Chung24,Reference Youn, Park and Kim28,Reference Ben-Hamouda, Ltaief and Kirsch38 The highest rating in bias assessment was for study confounders, which was high in eightReference Bongiovanni, Romagnosi and Barbella20,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24–Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36,Reference Ben-Hamouda, Ltaief and Kirsch38 and moderate in the remaining four studies.Reference Moseby-Knappe, Westhall and Backman21,Reference Pouplet, Martin and Colin27,Reference Scarpino, Lolli and Lanzo37,Reference Son, Lee and Park39 All studies had a low risk of bias for statistical analysis and reporting.Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24–Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36–Reference Son, Lee and Park39 Details of the assessment for each study are presented in Table 2.
Bias assessment according to QUIPS tool
Finally, there may have been publication biasReference Lin and Chu40 as outlined with Begg’s testReference Begg and Mazumdar34 and Egger’s test.Reference Egger, Smith, Schneider and Minder35 Funnel plots (Figures S1 and S2) were also suggestive of the presence of publication bias for the estimate of a pooled specificity (primary analysis) with a unimodal and multimodal approach, both with and without excluding GCS-M (sensitivity analysis). The results of the Egger’s test (t = 4.55 p < 0.001, bias estimate = 1.55 [SE 0.34]) suggested some publication bias of the unimodal specificity for primary analysis and sensitivity analysis (t = 4.35, p < 0.001, bias estimate = 1.28 [SE 0.30]). Publication bias was also suspected with the multimodal approach specificity for primary and sensitivity analysis using the Begg’s test (respectively, z = −2.86, p = 0.004, bias estimate = −69.00 [SE 24.15]; z = −4.01, p < 0.001, bias estimate = −58.00 [SE 14.45]).Reference Lin and Chu40
Diagnostic accuracy of unimodal tests
Table 3 summarizes the results of the unimodal tests reported in each included study. In studies utilizing clinical examination, GCS-M ≤2 either at 72 h post-rewarming or 72 h post-ROSC had an FPR ranging from 5% to 31%,Reference Kim, Park and Chung24–Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 whereas bilaterally absent PLR and/or CR at 72 h or 108 h post-ROSC had an FPR ranging from 0 to 6%Reference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Roger, Palmier and Louart25,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36–Reference Ben-Hamouda, Ltaief and Kirsch38 in predicting poor neurological outcomes. The study by Oddo et al. was the only one using pupillometry and demonstrated that NPi ≤2 at 72 h had an FPR of 0% (95% CI 0.0–3.5) in predicting poor outcomes, while standard PLR was not as specific (94%, FPR 6% [95% CI 2.1–14.3]).Reference Oddo, Sandroni and Citerio36 Status myoclonus had an FPR ranging from 0% to 11% depending on timing post-ROSC within 48–72 hReference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Pouplet, Martin and Colin27,Reference Scarpino, Lolli and Lanzo37,Reference Ben-Hamouda, Ltaief and Kirsch38 or 7 days post-ROSC.Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 NSE greater than 33 mcg/L within 6–72 h post-ROSC resulted in an FPR between 0% and 5%.Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28,Reference Ben-Hamouda, Ltaief and Kirsch38,Reference Son, Lee and Park39 For neuroimaging, CT head within 10–72 h post-ROSC showing either gray-to-white matter ratio <1.07–1.21 or generalized edema with a reduced differentiation between gray and white matter had an FPR between 0% and 9%.Reference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Youn, Park and Kim28,Reference Scarpino, Lolli and Lanzo37,Reference Son, Lee and Park39 Brain MRI showing signs of significant anoxic brain injury between 6 h and 2 weeks post-ROSC had an FPR ranging from 0% to 12%.Reference Moseby-Knappe, Westhall and Backman21,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Youn, Park and Kim28,Reference Son, Lee and Park39 Using either CT head or brain MRI showing diffuse anoxic brain injury at 72 h post-ROSC resulted in an FPR of 0% (95% CI 0.0–60.2) in predicting poor outcomes in one study.Reference Pouplet, Martin and Colin27 Highly malignant patterns on EEG predicted poor outcomes accurately, with an FPR of 0%–3% when EEG was performed within 24–72 hReference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28,Reference Scarpino, Lolli and Lanzo37,Reference Ben-Hamouda, Ltaief and Kirsch38 and as high as 25% when EEG was performed within 2 weeks post-ROSC.Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 Finally, SSEP showing bilaterally absent N20 wave or absence on one side and a pathological N20 potential on the other side at least 24 h post-ROSC demonstrated an FPR of 0–1%.Reference Moseby-Knappe, Westhall and Backman21,Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36–Reference Ben-Hamouda, Ltaief and Kirsch38
Diagnostic accuracy of unimodal data utilized in analysis
Note: SM = status myoclonus; PLR = pupillary light reflexes; CR = corneal reflexes; GCS-M = Glasgow Coma Motor Score; EEG = electroencephalography; SSEP = somatosensory evoked potentials; NSE = neuron-specific enolase; ADC = apparent diffusion coefficient; GWR = gray-to-white matter ratio; FP = false positive; FN = false negative; TP = true positive; TN = true negative; FPR = false positive rate; Sens = sensitivity; Spec = specificity; PPV = positive predictive value; NPV = negative predictive value; CI = confidence interval; EMG = Electromyography.
* All timing post-ROSC unless mentioned otherwise.
Modalities found in individual studies to have an FPR of 0% while also having a sensitivity of ≥50% were NSE >78.9 mcg/L 48–72 h post ROSC,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22 NSE >60 mcg/L 72 h post ROSC,Reference Pouplet, Martin and Colin27 bilaterally absent PLR 72 h post-ROSC,Reference Roger, Palmier and Louart25 SSEP with absent N20 wave on one side with pathological N20 wave on the other side at 72 h post-ROSC,Reference Scarpino, Lolli and Lanzo37 bilateral absent CR and PLR 72 h post-ROSC,Reference Youn, Park and Kim28 bilaterally absent N20 on SSEP at 24 h post-ROSC,Reference Youn, Park and Kim28 highly malignant patterns on EEG at 24 h post-ROSCReference Youn, Park and Kim28 and brain MRI with generalized edema 48–168 h post-ROSC.Reference Youn, Park and Kim28 However, sensitivities and specificities for the same modalities using similar thresholds and timing of assessment were quite variable between studies.
Diagnostic accuracy of multimodal combination of tests
A summary of the multimodal combinations used for each study, with their diagnostic accuracy, is presented in Table 4. All multimodal combinations had an FPR ≤5% for predicting poor outcomes in comatose post-arrest patients. The highest FPRs were seen with the combinations of GSC-M and PLR at 72 h post-ROSCReference Roger, Palmier and Louart25 (FPR 4.5% [95% CI 0.1–22.8]), GCS-M combined with CR at 72 h post-rewarming (FPR 2.1% [95% CI 0.6–5.2])Reference Kim, Park and Chung24 and any combination of two modalities including PLR, SSEP, EEG, NSE or myoclonus between 24 h and 72 h post-ROSC (FPR 0.6% [95% CI 0.0–3.6]).Reference Ben-Hamouda, Ltaief and Kirsch38 Ten studies used combinations resulting in an FPR of 0%.Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26–Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36,Reference Scarpino, Lolli and Lanzo37,Reference Son, Lee and Park39 Many studies used the 2015 ERC-ESICM algorithmReference Nolan, Cariou and Cronberg23 or a slightly modified version excluding GCS-M as a multimodal predictor,Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Scarpino, Lolli and Lanzo37 whereas other studiesReference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28 used the 2021 ERC-ESICM algorithm.Reference Nolan, Bttiger B.W. and Cariou9 Combinations of clinical examination using GCS-M, CR and/or PLR were used in three studies.Reference Kim, Park and Chung24–Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 Oddo et al. found that the combination of NPi with SSEP at 48–72 h led to an FPR of 0% (95% CI 0.0–5.6).Reference Oddo, Sandroni and Citerio36 NSE >54.8 ng/mL was combined with either brain CT or brain MRI within 6 h post-cardiac arrest and had similar results.Reference Son, Lee and Park39
Diagnostic accuracy of multimodal combinations reported in each study
Note: SM = status myoclonus; PLR= pupillary light reflexes; CR = corneal reflexes; GCS-M = Glasgow Coma Motor Score; EEG = electroencephalography; SSEP = somatosensory evoked potentials; NSE = neuron-specific enolase; ADC = apparent diffusion coefficient; GWR = gray-to-white matter ratio; FP = false positive; FN = false negative; TP = true positive; TN = true negative; FPR = false positive rate; Sens = sensitivity; Spec = specificity; PPV = positive predictive value; NPV = negative predictive value; CI = confidence interval.
* All timing post-ROSC unless mentioned otherwise.
The combinations leading to an FPR of 0% while also having a sensitivity ≥50% were GCS-M ≤2 and bilaterally absent PLR at 72 h post-rewarming,Reference Kim, Park and Chung24 bilaterally absent PLR and CR at 72 h post-rewarming,Reference Kim, Park and Chung24 GCS-M ≤2 with bilaterally absent PLR and CR at 72 h post-rewarming,Reference Kim, Park and Chung24 2015 ERC-ESCIM without GCS-M with modalities assessed between 24 and 72 h post-ROSC,Reference Bongiovanni, Romagnosi and Barbella20,Reference Scarpino, Lolli and Lanzo37 2021 ERC-ESCIM with modalities assessed 24 h to 7 days post-ROSC,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28 NPi ≤2 with bilateral absent N20 on SSEP at 48–72 h post-ROSCReference Oddo, Sandroni and Citerio36 and NSE >54.8 ng/mL and either abnormal brain CT or brain MRI within 6 h.Reference Son, Lee and Park39 All combinations excluding GCS-M, unless part of the ERC-ESCIM algorithm, had specificity greater than 99% (FPR <1%),Reference Bongiovanni, Romagnosi and Barbella20–Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36–Reference Son, Lee and Park39 while most of them had sensitivity >50%.Reference Bongiovanni, Romagnosi and Barbella20,Reference Kim, Park and Chung24,Reference Pouplet, Martin and Colin27,Reference Youn, Park and Kim28,Reference Oddo, Sandroni and Citerio36,Reference Scarpino, Lolli and Lanzo37,Reference Son, Lee and Park39
Comparison of the diagnostic accuracy of multimodal versus unimodal testing
Figures 2 and 3 illustrate the comparison of the FPR and sensitivity between multimodal and unimodal data for all included studies. A meta-analysis was not possible for several reasons. We noted significant publication bias in our funnel plots, Egger’sReference Egger, Smith, Schneider and Minder35 and Begg’s testReference Begg and Mazumdar34 (Figures S1 and S2). More importantly, each multimodal combination of tests used different combinations of single tests assessed at various time points after ROSC, with different thresholds for each single test. Moreover, the neurological outcomes were not assessed at the same time point in all the included studies. For these reasons, a summary estimate of specificity and sensitivity for unimodal and multimodal data could not be calculated. In addition, a key assumption of meta-analyses is that estimates are mutually independent,Reference Schmid, Stijnen and White41 which was not met with our data. Lastly, there was a significant amount of heterogeneity in the collected data that varied between 0% and 96% when measured.
Comparison of the false positive rate of multimodal versus unimodal data from included studies.
Comparison of the sensitivity of multimodal versus unimodal data from included studies.
Discussion
This systematic review summarizes the accuracy of both multimodal and unimodal approaches in predicting poor neurological outcomes post-cardiac arrest. While a meta-analysis was not possible, review of the reported data suggests that a multimodal approach, regardless of the combination of modalities, may increase specificity and sensitivity when predicting poor outcomes compared to a unimodal approach. To our knowledge, this is the first systematic review comparing multimodal and unimodal approaches in neuroprognostication of post-cardiac arrest patients. Previous studies assessed multiple tests used as individual modalities but did not compare diagnostic accuracy between multimodal and unimodal approaches as a systematic review.Reference Wang, Chang and Su42–Reference Golan, Barrett and Alali45
Accurate neuroprognostication is paramount to avoid inappropriate WLSM but also to circumvent prolonged, invasive and costly therapies that could perpetuate patient and family suffering when there is no realistic chance of favorable recovery.Reference Girotra, van Diepen and Nallamothu46–48 Despite the profound impact of WLSM decisions, there is a lack of consensus regarding an acceptable FPR in the determination of prognosis post-cardiac arrest. Most guidelines include modalities with an FPR (or the upper level of the confidence interval) ≤5%.Reference Fordyce, Kramer and Ainsworth8–Reference Rajajee, Muehlschlegel and Wartenberg11 However, one recently published national survey suggests that providers prefer an FPR <1% when predicting a poor prognosis and recommending WLSM.Reference Bencsik, Kramer, Couillard, MacKay and Kromm49 An international survey suggested that many even prefer an FPR <0.1%.Reference Steinberg, Callaway and Arnold50 While several factors are integrated into shared decision-making about patients’ goals of care, prognostic certainty forms the foundation for these discussions. Studies suggest that uncertainty, especially when suppressed or ignored, can have a negative impact on families and healthcare providers.Reference Etkind and Koffman51 Up to 70% of physicians report feeling some level of distress when determining post-arrest prognosis, much of which stems from managing uncertainty.Reference Bencsik, Kramer, Couillard, MacKay and Kromm49
While many studies reported FPRs of 0% for individual modalities recommended in recent guidelines, others reported FPRs as high as 6% for bilaterally absent PLR or CR,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Kim, Park and Chung24 4% for status myoclonusReference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Pouplet, Martin and Colin27 and up to 31% for GCS-M ≤2.Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 The highest FPR reported for EEG, SSEP, CT, MRI and NSE used in isolation was 25%,Reference Bisschops, van Alfen, Bons, van der Hoeven and Hoedemaekers26 3%,Reference Moseby-Knappe, Westhall and Backman21 9%,Reference Youn, Park and Kim28 12%Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22 and 5%,Reference Youn, Park and Kim28 respectively. No single test is perfect. The variability of FPRs likely results from several factors, including those inherent to observational studies, such as the inability to control for confounding factors (variably reported in studies), inter- and intra-observer variability when reporting test results and characteristics of each modality, such as spatial and temporal resolution. Providers should also be mindful that positive verification bias (“self-fulfilling prophecy”) that is inherent in many studies may falsely lower the reported FPRs of individual modalities, though findings in our systematic review were concordant in studies where WLSM was not pursued.
The European Resuscitation Council (ERC) and the European Society of Intensive Care Medicine (ESICM) published guidelines for neuroprognostication after cardiac arrest in 2014 Reference Sandroni, Cavallaro and Cronberg52 and 2015Reference Nolan, Cariou and Cronberg23 comprised of a four-step algorithm. This multimodal approach has been tested retrospectively using data from the TTM trialReference Moseby-Knappe, Westhall and Backman21 and predicted poor neurological outcome (CPC 3–5) with a specificity of 100% but failed to identify approximately 60% of patients with poor neurological outcomes.Reference Moseby-Knappe, Westhall and Backman21 The more contemporary ERC/ESICM,Reference Nolan, Bttiger B.W. and Cariou9 Neurocritical Care Society,Reference Rajajee, Muehlschlegel and Wartenberg11 American Heart AssociationReference Panchal, Bartos and Cabañas10 guidelines and Canadian position statementReference Fordyce, Kramer and Ainsworth8 all recommend a multimodal approach, albeit based on low-quality evidence. Recommendations are based on the assumption that a multimodal approach leads to improved FPR and sensitivity, as the chances of discovering findings indicative of a poor prognosis are increased when more tests are performed. Several questions remain, however, including the ideal number and combination of modalities required in a multimodal approach and whether that approach improves specificity and/or sensitivity in predicting poor outcomes.
One limitation of this review includes the publication bias of included studies as demonstrated by our funnel plots, Egger’s and Begg’s tests. Many of the included studies also had a moderate to high degree of bias. This was unfortunately unavoidable, since most studies were retrospective and observational in nature, with inconsistency in reporting confounders, such as the effects of sedation, opioids and profound physiologic or metabolic disturbance, all of which are important considerations for neuroprognostication. Lastly, heterogeneityReference Higgins and Thompson53 between studies was high and was an important reason why conducting a meta-analysis was not possible. The included studies used similar but distinct combinations of tests, with variable diagnostic thresholds and timing, as well as variation in study size, proportion of patients with shockable rhythms, and different cardiac arrest settings, thereby precluding calculation of summary estimates. TTM was utilized in all studies, although not in all patients,Reference Zhou, Maciel, Ormseth, Beekman, Gilmore and Greer22,Reference Scarpino, Lolli and Lanzo37 with variable temperature goals and durations (Table S4). TTM is known to affect the ideal timing and accuracy of modalities used in neuroprognostication.Reference Sandroni, Cavallaro and Callaway54,Reference Sandroni, Cavallaro and Callaway55 While the TTM literature continues to evolve, care providers should be mindful of the possible effects TTM may have on diagnostic tests.
Despite these limitations, this systematic review utilized established methodology and a pre-registered protocol. While unanswered questions remain, these rigorous methods and selection process are emulating contemporary recommendations in neuroprognostication, which in turn increases the strength of our conclusion.
Conclusions
A multimodal approach to neuroprognostication aimed at identifying concordant findings in two or more modalities recommended in contemporary guidelines to predict a poor prognosis is feasible and may improve the FPR and sensitivity compared to an approach that utilizes the findings of individual modalities. More research is required to establish the ideal number and combination of modalities, as well as whether modalities not yet recommended in guidelines may also be of benefit in a multimodal approach.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/cjn.2025.10112.
Acknowledgments
We would like to thank Rachael Houlton for her contribution to the abstract and full-text review process.
Author contributions
AB: conceptualization, methodology, investigation, visualization, supervision, writing – original draft, writing – review and editing.
CB: conceptualization, methodology, investigation.
AS: conceptualization, software, formal analysis.
AK: conceptualization, methodology, writing – review and editing.
JK: conceptualization, methodology, investigation, supervision, writing – review and editing.
Funding statement
None.
Competing interests
Committee member for Neuroprognostication in Post Cardiac Arrest Patient: A Canadian Cardiovascular Society Position Statement and committee member for Canadian Cardiovascular Society/Canadian Cardiovascular Critical Care Society/Canadian Association of Interventional Cardiology Clinical Practice Update on Optimal Post Cardiac Arrest and Refractory Cardiac Arrest Patient Care. The authors confirmed that patient consent is not applicable to this article. This is a review article using already published literature; therefore, our manuscript did not require consent from the patient.
Open access
