Impact of a Specialized Care Pathway for ICU Patients With Disorders of Consciousness: Observational Insights From an Expert Center

ABSTRACT

Background

Management of lasting disorders of consciousness (DoC) is complex and sometimes requires referral of patients to an expert team. We aimed to evaluate the impact of such specialized expert assessment through the description of a single‐center experience.

Method

Monocentric study including all patients with DoC referred to a tertiary neuroICU for multimodal clinical, electrophysiological, and MRI assessment allowing diagnostic and therapeutic proposals and multidisciplinary prognostic prediction. The primary endpoint was the patients' actual outcome, assessed by 3‐month functional disability and 1‐year mortality, according to the predicted prognosis.

Results

Between July 2017 to June 2024, 124 patients (median age 61 [49–70] years, 67% of men) were referred from 43 departments of 37 primary hospitals. Most frequent etiologies were post‐anoxic encephalopathy (62%) and secondary systemic brain injuries (15%). At a median of 23 [15–33] days after DoC onset, predicted prognosis was good in 20 (16%) patients, intermediate in 41 (33%), and poor in 63 (51%). Three‐month GOS‐E was 3 [1–3], 2 [1–3], and 1 [1] (p < 0.001) and 1‐year mortality was 30%, 41%, and 82% (p < 0.001), respectively. Withdrawal of life‐sustaining therapies increased with worsening predicted prognosis (38%, 41%, and 65%, respectively, p < 0.001), but functional disability and mortality remained higher in the poor prognosis group (p < 0.003 and p < 0.039, respectively) even after excluding patients with those decisions.

Conclusion

Specialized consciousness assessment impacts the ethical decisions made by primary centers and helps refine the prediction of recovery but fails to achieve highly accurate individual prediction in a population with overall poor survival and functional prognosis.

In a cohort of 124 patients from 37 ICUs referred to an expert center for lasting disorders of consciousness, specialized multimodal consciousness assessment impacted the ethical decisions made by primary centers and helped to refine the prediction of recovery but without highly accurate individual prediction in a population with overall poor survival and functional prognosis.

Abbreviations

AEP

auditory evoked potentials

BAEP

brainstem auditory evoked potentials

CRS‐r

coma recovery scale—revised

DoC

disorder of consciousness

DTI

diffusion tensor imaging

EEG

electroencephalogram

EMCS

emergence from minimally conscious state

FA

fractional anisotropy

GCS

Glasgow coma scale

GOS‐E

Glasgow outcome scale—extended

HIBI

hypoxic–ischemic brain injury

HR

hazard ratio

ICU

intensive care unit

IQR

interquartile range

LLST

limitation of life‐sustaining therapy

MCS

minimally conscious state

MLAEP

middle latency auditory evoked potentials

MMN

mismatch negativity

MRI

magnetic resonance imaging

OR

odds‐ratio

SAPS‐II

simplified acute physiology score 2

SSEP

somatosensory evoked potentials

TBI

traumatic brain injury

VS/UWS

vegetative state/unresponsive wakefulness syndrome

WLST

withdrawal of life‐sustaining therapy

1. Introduction

Disorders of consciousness (DoC), either due to a primary brain injury or following secondary systemic insults are common in the intensive care unit (ICU) [1, 2, 3] and associated with significant morbi‐mortality [4, 5] and challenges [6]. Clinical decisions must often be made while patients cannot express their wishes. Approximately two‐third of patients with primary brain injuries die after withdrawal (WLST) or limitation (LLST) of life‐sustaining therapies [7, 8, 9] while consciousness can be underestimated [10] and its recovery delayed [11, 12]. In turn unlimited intensity of care can foster surviving with disabilities not aligned with the patient's life goals while mobilizing inappropriate human and material resources. Neuroprognostication, i.e., the prediction of recovery from DoC after a severe brain injury, is a complex process, requiring the integration of many parameters, some of which linked to the patient and others linked to both the primary brain injury and potential secondary insults occurring during the ICU stay [13, 14].

To help the clinicians in this process, many clinical, neurophysiological and brain‐imaging techniques were developed [14, 15] and guidelines were issued, either focused on a specific etiology such as hypoxic–ischemic brain injury (HIBI) following cardiac arrest [16] or more general [17, 18, 19]. They all promote multimodality, i.e., the integration of several tests to increase prediction confidence. However, many of the recommended tests, for instance somatosensory evoked potentials (SSEP), are not available in primary care ICU [20, 21]. Even when tests are available, the prognosis may remain uncertain, as it is the case for approximately 50% of patients with HIBI when applying the ERC/ESCIM algorithm [22]. In other etiologies, such as traumatic brain injury (TBI), stroke, or in secondary brain insults, prognostic tests are overall less reliable and despite existing guidelines [23, 24], no widely accepted, standardized algorithm exists to guide the integration of multimodal prognostication tools. Given this complexity, patients are sometimes referred from primary care units to centers with a specific expertise in neuroprognostication. While recent data suggested a benefit from this expert assessment [25], evidence is still needed to support the organization of these specialized care pathways.

To precise the place such care pathways could have in the healthcare system, we described the characteristics of patients referred to a specialized center for assessments of persisting DoC and assessed the impact of the evaluation in terms of diagnosis, prognosis, and therapeutic changes on the actual outcome of patients.

2. Material and Methods

2.1. Design & Population

This was a retrospective monocentric study including patients referred from primary care centers to the neuroICU of GHU Paris Sainte‐Anne university hospital for the assessment of a lasting DoC. All adult patients suffering from a DoC following an established or suspected primary or secondary acute brain injury, for whom specialized diagnostic, therapeutic, and/or prognostic evaluations were deemed necessary by the primary care team, were included from July 2017 (initiation of the care pathway) to June 2024.

2.2. Ethics

The study was approved by the Ethics Committee of the French Intensive Care Society (22‐049, 10/20/2022). Informed consent was waived, but patients were informed about the use of their data for research purposes.

2.3. Data From Primary Care Team

Data from the primary care team were collected from the hospitalization reports sent with the patients when referred to the neuro‐ICU expert center. The following data were systematically gathered: demographic data (age, sex); dates and etiologies of brain injury (primary brain injury with HIBI, stroke, TBI, other primary or secondary brain injury, including all injuries following metabolic, septic and/or hypoxemic events in the ICU without any other clear explanation) and related prognostic factors upon ICU admission; organ support, sedation and complications during the ICU stay; neurological status (using the terminology used in the reports and standardized scores and categories with eyes opening, eye contact, command following); and neurological complementary examinations performed (electroencephalography, brain imaging, biomarkers, …).

2.4. Specialized Expert Assessment

All patients underwent the same evaluation, including:

• –Consciousness assessment using repeated clinical examination with the Glasgow coma scale (GCS) and Coma Recovery Scale—revised (CRS‐r) [26] (current gold standard for conscious state diagnosis, defining emergence from minimally conscious state–EMCS, minimally conscious state–MCS, vegetative state/unresponsive wakefulness syndrome–VS/UWS and coma).
• –Electrophysiological assessment with 20‐min video‐electroencephalogram (EEG) with auditive and nociceptive stimulations [27], interpreted according to the standardized ACNS terminology [28] and Westhall classification [29]. SSEPs and auditory evoked potentials (AEPs), with short‐ (brainstem, BAEP), middle‐ (MLAEP), and long latencies (N100, mismatch negativity—MMN—and P300 to subjects' own names) [30].
• –Brain magnetic resonance imaging (MRI) conducted on a 3‐Tesla unit, with standard sequences unless contraindicated (e.g., ferromagnetic implants). The presence and location of any anoxic, ischemic, hemorrhagic, metabolic and inflammatory lesion was recorded.
• –In a subset of patients, long‐lasting (12 h to 24 h) continuous video‐EEG recordings to study sleep cycles and patterns and diffusion tensor imaging (DTI) MRI to quantify fractional anisotropy (FA) were performed.

Following multimodal evaluation, a multidisciplinary team, including intensive care medicine, neurophysiology, neuroradiology, and rehabilitation medicine physicians, reviewed all test results to emit a prognostic prediction among three pre‐defined categories (primary exposure variable): good (compatible with neurological improvement and recovery of consciousness), poor (neurological improvement and recovery of consciousness deemed improbable), and intermediate/indeterminate (prognosis unclear due to confounding factors or to inconclusive/discordant test results).

Additionally, modification of the primary diagnosis or identification of associated diagnoses and treatment propositions or modifications related to DoC were collected.

2.5. Outcomes

The primary objective was to determine actual patients' outcomes according to predicted prognoses. Primary endpoints were 3‐month post‐evaluation functional disability using the Glasgow Outcome Scale‐Extended (GOS‐E) [31] and 3‐month and 1‐year mortality assessed via telephone interview with the family or the referring physician. In case of missing data, vital status was retrieved from the Institut National de la Statistique et des Etudes Economiques open death national registry (https://deces.matchid.io/search).

Secondary objectives included the description of neurological status and neuroprognostic tests performed before and during the expert evaluation, and the description of WLST (the removal of life‐sustaining therapies) and LLST (no therapeutic escalation) decisions taken by the primary care teams after the evaluation (without involvement of the expert team).

2.6. Statistics

Analyses were performed using R version 4.3.3. Quantitative variables were described with medians [interquartile ranges] and compared using Wilcoxon rank‐sum or Kruskal‐Wallis tests. Qualitative variables were described as counts (percentages) and compared using Chi‐square or Fisher's exact tests.

The association between 3‐month functional disability and predicted prognosis was assessed using a shift analysis through ordinal logistic regression with proportional odds of ordinal GOS‐E levels (function polr from the MASS package). This analysis yields a common odds‐ratio (OR) for all cut‐offs of GOS‐E, with OR > 1 indicative of higher GOS‐E probability (better prognosis). To complement the ordinal shift analysis, we assessed discrimination using concordance‐based measures. Somers' D was estimated between the three‐level categories of emitted prognosis (bad<intermediate<good) and the 3‐month ordinal GOS‐E score as a non‐parametric measure of ordinal discrimination, and the corresponding C‐index was derived. Kendall's τ was also computed as a rank‐based measure of association. To assess the impact of a bad predicted prognosis, we also computed concordance metrics for the dichotomized emitted prognosis (bad<good and intermediate). Non‐parametric bootstrap resampling with 2000 iterations were performed to obtain 95% confidence intervals. The association with 1‐year mortality was assessed using survival analysis with Cox proportional hazards models to compute raw hazard ratios (HR) of cumulative mortality incidence within the year following expert assessment according to predicted prognosis. Kaplan–Meier curves were compared using the Log‐rank test (ggsurvplot and coxph functions from the survival package). To mitigate potential self‐fulfilling prophecy effects, sensitivity analyses were performed in a subgroup excluding patients with WLST decisions. Concordance metrics between emitted prognosis and 1‐year mortality were also computed following the same method as for the GOSE. Additionally, subgroup analyses were also performed according to brain injury etiology, stratified as anoxic or non‐anoxic. Adjusted OR and HR were computed in multivariable ordinal logistic regression and Cox analysis respectively including the predicted prognosis, age, sex, etiology (HIBI vs. non‐HIBI), time since brain injury, state of consciousness and WLST decisions. All tests were two‐tailed, with a p‐value < 0.05, with no imputation of missing data. Data reporting followed the STROBE guidelines (see STROBE checklist in Table S1).

3. Result

3.1. Population

From July 2017 to June 2024, 124 patients were referred for specialized evaluation from 43 departments of 37 hospitals. All patients were initially hospitalized in ICU, and all but two (98%) were directly transferred from the ICU. Patients were mostly male (67%), with a median age of 61 [49–70] years and SAPS‐II score at primary ICU admission of 59 [47–73]. Etiologies of brain injuries were HIBI in 62.1%, secondary brain injuries in 15.3%, stroke in 8.9%, TBI in 3.2%, other primary brain injury in 8.9%, and unknown in 1.6%. All patients received sedations in the primary ICU for a median duration of 3 [1–12] days (Table 1).

TABLE 1.

Population demographic and primary care unit characteristics.

Characteristics	N = 124
Age (years)	61 [49–70]
Sex (male/female)	83 (67)/41 (33)
Charlson comorbidity index	3 [1–4]
Primary care unit
Number of different units	37
Number of patients/units	2 [1–4]
Unit from university hospital	50 (40)
Intensive care unit	122 (98)
Etiologies of brain injury
Cardiac arrest	77 (62.1)
Stroke	11 (8.9)
Traumatic brain injury	4 (3.2)
Other primary brain injury	11 (8.9)
Secondary brain injury	19 (15.3)
Unknown	2 (1.6)
SAPS‐II, n = 63 (51%)	59 [47–73]
Organ support
Vasopressors	88 (71)
Invasive mechanical ventilation	124 (100)
Renal replacement therapy	35 (28)
Extra corporeal membrane oxygenation	3 (2.4)
Complications during the stay in the primary unit
In‐hospital cardiac arrest	4 (3.2)
Acute respiratory distress syndrome	27 (22)
Infections	108 (87)
Stroke	17 (14)
Seizure	47 (38)
Sedation	124 (100)
Sedation duration (days), n = 122 (98%)	3 [1–12]

Note: Quantitative variable are expressed as median [IQR] and qualitative variable as number (%). Data was complete except specified otherwise (n).

3.2. Neurological Assessment Performed in the Primary Care Unit

In the hospitalization reports from primary care units, neurological examination was mostly descriptive and missing in more than 10% of cases. Most patients had eyes open (76%) but without visual contact (87%) nor command following (90%). The most used terminology to describe neurological status was delayed, absent or pathological awakening (55.4%). GCS score was reported in only 66 (53%) patients and CRS‐r score in 3 (2%) patients (Table S2). Nearly all patients underwent EEG (96%) and brain imaging (85%), including MRI in 73%, while neuron‐specific enolase and SSEP were performed in 28% and 9.7%.

3.3. Multimodal Expert Assessment

The motives to request expert evaluation were diagnosis in 24%, treatment in 6%, and prognosis in 99% of cases with more than one motive in 25%. The assessment was conducted over 2 days [1–3] with a delay of 23 days [15–33] after the initial brain injury and 17 days [9–28] after sedation weaning.

Following repeated CRS‐r assessments, patients were in a VS/UWS (62%), MCS (28.2%), coma (9.6%), and EMCS (0.8%). Qualitative video‐EEG, performed in 98% of patients, revealed unreactive patterns in 28% and electrical seizures in 14%. According to the Westhall classification, EEG was benign in 54%, malignant in 42%, and highly malignant in 4% of patients. Long‐duration video‐EEG, performed in 65% of patients, revealed sleep–wake cycles in 35% of cases. SSEPs and AEPs were performed in 96% of patients, with bilateral abolition of N20 waves in 20% of cases and preservation of MMN and P300 in 45% and 50%, respectively. Brain MRI was conducted in 89% of patients, with DTI sequences in 50%, revealing reduced fractional anisotropy in 68% (Table S3).

An explanation for the DoC was retained for the 30 patients referred for diagnosis, with the identification of an added cause of brain injury in 47% of cases. Therapeutic modifications were implemented in 26% of cases, involving antiepileptic (48%), immunomodulatory (21%) and dopaminergic drugs (21%), zolpidem tests (17%), and other therapeutic adjustments (14%).

Following multimodal assessment, emitted prognosis was good for 20 patients (16%), intermediate for 41 (33%), and poor for 63 (51%). Results of expert assessment (behavioral examination, EEG, SSEPs, AEPs, MRI and treatment modification) according to these three predicted categories are presented in Table S3.

3.4. Patients' Outcomes

Life‐sustaining therapies decisions were made by the primary care team in 75% of cases (WLST in 56% of cases and LLST in 44%) at a median delay of 6 [3–13] days following expert assessment. Frequency and delay of these decisions varied significantly across the good (45%), intermediate (63%), and poor (93%) prognosis groups, respectively (p < 0.001, Table 2; Figure 1). Individual patients' trajectories are presented in Figure S1.

TABLE 2.

Patients' actual primary and secondary outcomes according to predicted prognosis.

Characteristic	Overall N = 124	Good N = 20 (16%)	Intermediate N = 41 (33%)	Bad N = 63 (51%)	p
Limitations & withdrawal of life‐sustaining therapy decisions
Life‐sustaining therapy decisions, n = 121 (98%)	91 (75)	9 (45)	26 (63)	56 (93)	< 0.001
Type, n = 90/91 (99%)					0.055
Withdrawal (WLST)	50 (56)	3 (33)	11 (42)	36 (65)
Limitation (LLST)	40 (44)	6 (67)	15 (58)	19 (35)
Delay from expert assessment (days), n = 70/91 (77%)	6 [3–13]	57 [12–78]	6 [4–11]	6 [2–13]	0.015
3 & 12 months outcomes
3‐month mortality, n = 121 (98%)	73 (60)	6 (30)	17 (41)	50 (82)	< 0.001
Median GOS‐E at 3 months, n = 122 (98%)	1 [1–3]	3 [1–3]	2 [1–3]	1 [1–1]	< 0.001
Dichotomized GOS‐E at 3 months, n = 122 (98%)					0.019
GOS‐E 1–3	109 (89)	16 (80)	34 (83)	59 (97)
GOS‐E 4–8	13 (12)	4 (20)	7 (17)	2 (3)
1‐year mortality, n = 122 (98%)	81 (66)	11 (55)	18 (45)	52 (84)	< 0.001

Note: Quantitative variables are expressed as median [IQR] and qualitative variable as number (%). Data was complete except specified otherwise (n). p‐value < 0.05 are bolded.

FIGURE 1.

Withdrawal and limitations of life‐sustaining therapies decisions. (A) Number of patients with withdrawal and limitations of life‐sustaining therapies (WLLST) and (B) delay of these decisions according to the emitted prognosis.

At 3 months, the median GOS‐E score was 1 [1–3], with a significant association with the predicted prognosis (Table 2). Specifically, compared to poor prognosis, patients with good and intermediate prognosis showed significantly higher odds for better GOS‐E: OR 16.7 (95% CI [5; 50], p = 5 × 10⁻⁷) and 7.56 (95% CI [3.1; 18.6], p = 2 × 10⁻⁶), respectively. No difference was found between good and intermediate prognoses (OR 1.89, 95% CI [0.68; 5.26], p = 0.211, Figure 2A). In complementary concordance analyses, emitted prognosis demonstrated good ordinal discrimination of 3‐month GOS‐E. Somers' D was 0.46 (95% CI [0.23; 0.62]), corresponding to a C‐index of 0.73 (95% CI [0.62; 0.81]). Kendall's τ was 0.45 (95% CI [0.22; 0.61]), consistent with a moderate‐to‐strong ordinal association between predicted and observed outcomes (Table 3).

FIGURE 2.

GOS‐E at 3 months according to emitted prognosis. Proportion of each Glasgow Outcome Scale‐extended (GOS‐E) categories in the three groups of emitted prognoses (bad, intermediate and good) (A) in the whole population (B) and in the subset of patients without withdrawal of life‐sustaining therapies (WLST). P‐values correspond to ordinal logistic regressions with proportional odds (shift analyses).

TABLE 3.

Concordance metrics between predicted prognosis and actual outcome.

Populations	Emitted prognosis: good>intermediate>bad			Emitted prognosis: good & intermediate>bad
	Concordance metrics (95% CI)			Concordance metrics (95% CI)
	Kendall's τ	Dxy	C index	Kendall's τ	Dxy	C index
3‐month GOSE
All	0.45 [0.32; 0.58]	0.46 [0.33; 0.59]	0.73 [0.66; 0.8]	0.46 [0.33; 0.59]	0.43 [0.31; 0.55]	0.71 [0.65; 0.78]
No WLST	0.33 [0.13; 0.51]	0.31 [0.12; 0.48]	0.65 [0.56; 0.74]	0.36 [0.17; 0.54]	0.28 [0.13; 0.43]	0.64 [0.56; 0.71]
Anoxic	0.42 [0.22; 0.61]	0.43 [0.23; 0.62]	0.72 [0.62; 0.81]	0.43 [0.23; 0.61]	0.42 [0.23; 0.59]	0.71 [0.61; 0.8]
Non‐anoxic	0.4 [0.17; 0.61]	0.4 [0.16; 0.62]	0.7 [0.58; 0.81]	0.43 [0.19; 0.63]	0.35 [0.14; 0.54]	0.67 [0.57; 0.77]
1‐year survival
All	−0.32 [−0.48; −0.16]	0.37 [0.19; 0.55]	0.69 [0.6; 0.78]	−0.38 [−0.53; −0.21]	0.4 [0.22; 0.56]	0.7 [0.61; 0.78]
No WLST	−0.13 [−0.36; 0.1]	0.15 [0; 0.41]	0.57 [0.5; 0.7]	−0.23 [−0.46; 0.01]	0.21 [0.02; 0.44]	0.61 [0.51; 0.72]
Anoxic	−0.37 [−0.57; −0.15]	0.43 [0.17; 0.66]	0.71 [0.59; 0.83]	−0.4 [−0.6; −0.17]	0.44 [0.19; 0.67]	0.72 [0.6; 0.83]
Non‐anoxic	−0.18 [−0.45; 0.09]	0.21 [0.01; 0.51]	0.6 [0.51; 0.76]	−0.27 [−0.54; 0]	0.26 [0.03; 0.53]	0.63 [0.51; 0.76]

At 1 year, the overall mortality rate was 66%, with significant differences across the three groups: 55%, 45%, and 84% for the good, intermediate, and poor prognosis groups, respectively (p < 0.001, Table 2). Cumulative one‐year mortality incidence also differed significantly (Log‐rank test p = 6 × 10⁻⁶). The hazard ratio (HR) for mortality was 3.52 (95% CI [1.65; 7.51], p = 0.001) for patients with a poor prognosis and 1.15 (95% CI [0.50; 2.65], p = 0.7) for patients with an intermediate prognosis compared to those with a good prognosis (Figure 3A,B). Concordance analyses confirmed moderate‐to‐strong ordinal association between predicted prognosis and observed mortality (Table 3) with a Somers' D of 0.37 (95% CI [0.19; 0.55]), corresponding to a C‐index of 0.69 (95% CI [0.60; 0.78]) and a Kendall's τ of −0.32 (95% CI [−0.48; −0.16]) (Table 3).

FIGURE 3.

One‐year survival according to emitted prognosis. (A) Kaplan–Meier curves of cumulative incidence of one‐year survival according to the three groups of emitted prognoses in the whole cohort. (B) One‐year survival according to dichotomized emitted prognosis (bad vs. good or intermediate) in the whole cohort and (C) in the subset of patients without withdrawal of life‐sustaining therapies (WLST). P‐values correspond to the log‐rank test for Cox proportional hazards models.

To minimize the self‐fulfilling prophecy bias, sensitivity analyses were performed. First, subgroup analyses excluding patients who had WLST. Good and intermediate prognosis remained significantly associated with better 3‐month functional outcomes compared to poor prognosis (OR 7.69, 95% CI [1.89; 33.33], p = 0.0017, and OR 4.83, 95% CI [1.59; 14.7], p = 0.003) (Figure 2B). Similarly, one‐year survival was significantly higher for patients with a good or intermediate prognosis compared to poor prognosis (Log‐rank p = 0.039, Figure 3C). Second, multivariable analyses. Predicted poor prognosis remained independently associated with lower odds of better GOS‐E (adjusted OR 0.06 95% CI [0.01; 0.28]) and increased one‐year mortality (adjusted HR 5.17 95% CI [2.05; 13], p < 0.001 as compared to good prognosis). WLST was also independently associated with lower odds of better functional outcome and lower survival (Figure 4; Table 3; Tables S4 and S5).

FIGURE 4.

Multivariable analyses. Forest plot of multivariable (A) proportional odds ordinal logistic regression model for 3‐month Glasgow Outcome Scale—Extended (GOS‐E) and (B) Cox proportional hazards model for 1‐year mortality, adjusted for selected baseline covariates: Age, sex, etiology and delay of acute brain injury (ABI), state of consciousness and emitted prognosis according to expert assessment and withdrawal of life‐sustaining therapy decisions (WLST) following assessment. No imputation of missing data was performed; therefore, analyses were restricted to complete cases.

To assess the influence of brain injury etiology on the association between predicted prognosis and actual outcome, we performed subgroup analyses stratified on anoxic and non‐anoxic brain injuries. Good and intermediate predicted prognosis remained significantly associated with a higher likelihood of a better GOS‐E and with a higher cumulative incidence of survival at 1 year in both subgroups as compared with bad emitted prognosis, although performances were higher in the anoxic than in the non‐anoxic population (Table 3; Figure S2). Additionally, etiology was neither independently associated with 3‐month functional outcome nor with 1‐year survival in the multivariable analyses (Figure 4; Tables S4 and S5).

4. Discussion

Overall, ICU patients with lasting DoC referred to a tertiary center for multimodal expertise had a high mortality and poor functional prognosis. Yet, the multimodal expertise significantly predicted neurological outcomes and seemed to impact decisions of care intensity following expert assessment.

The first insight from our study is that expert assessment allowed to complete the neurological assessment of patients with persistent DoC. Over a short period of time (2 days in median), almost all patients had access to a large set of validated EEG, auditory evoked potentials and MRI tools [14, 19]. In primary care units, unexpectedly most patients already had an EEG and brain imaging, while evoked potentials were rarely available (SSEPs in only 9.7% and no AEPs), in line with previous studies [21]. While evoked potentials' interest in the prediction of recovery is increasingly recognized, it is unlikely that this was the only reason to refer patient for prognostic assessment [32]. Expert assessment also allowed for a precise clinical diagnosis of conscious state. Despite being widely available, this assessment was surprisingly limited in the primary care unit. Description of the neurological status of patients in the hospitalization reports was dominated by the use of vague terminology with unclear definition [33]. CRS‐r was reported in only 2% of patients despite being the gold‐standard for consciousness assessment [17] and the recommendation for its early use in the ICU [18]. Although CRS‐r scoring requires training and expertise [34], conscious state is consistently associated with the likelihood of recovery, in our study as in many others [13, 35, 36] and physical examination is considered tier zero for predicting the recovery of consciousness [37, 38]. Overall, it seems that patients were mainly referred for the specific expertise required to assess, interpret and synthesize the complex situations of patients with acute and subacute DoC, rather than because of a lack of available tools in the primary care center. Further supporting this, almost a third of patients were also referred for diagnostic purposes, broadening the spectrum of interest of these specialized care pathways.

Regarding neuroprognostication, we found significant associations between emitted prognosis after multimodal expert assessment and both actual functional status and mortality, especially for bad and good predicted outcomes. These data confirm the ones from a recent study in which emitted prognosis increased accuracy and reduced uncertainty about the patient's evolution [25]. Our study extends these findings to a more acute and severe population (one‐third of patients with HIBI, delay of 33 [23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53] days since brain injury, less than half of patients in coma or VS/UWS in Rohaut et al. [25]). It also adds to the growing body of evidence supporting the multimodality in consciousness assessment [39, 40, 41, 42] which is promoted in recent guidelines [16, 17, 18, 19]. Our findings also support the collegiality which is essential in complex ethical processes such as neuroprognostication [43]. Since our cohort was dominated by patients with anoxic brain injuries, our conclusions primarily apply to this population. However, emitted prognosis remained associated with actual outcomes in patients with non‐anoxic brain injury and its association with 1‐year survival was independent of the type of brain injury, suggesting that specialized assessment is also valuable in non‐anoxic patients, although predictive performance was higher in patients with anoxic injury than in those without, likely reflecting the greater homogeneity of the anoxic population and the more robust performance of prognostic tools in this context. Interestingly, the intermediate prognosis category, corresponding to either diverging results or non‐interpretable tests due to possible confounding factors [44, 45], had a prognosis closer to the good category. This illustrates that confounding factors should be sought thoroughly and that inconsistencies should primarily prompt physicians to wait and reassess the patient later. In any case, doubt should benefit the patient and decisions should be made very carefully [46].

Yet overall functional and vital prognosis were surprisingly poor. While this is clearly anticipated in patients with bad outcomes, it is more surprising in patients with predicted good and intermediate outcomes. First, referred patients were especially severe, both in terms of consciousness impairment (most patients were in a VS/UWS with median CRS‐r at 6 persisting for approximately 3 weeks) and in terms of other organ dysfunction (more than two‐thirds received vasopressors and one‐third underwent renal replacement therapy in the primary ICU). Our results might not generalize to other settings with less severe patients. Second, this might be due to the high rate of limitations of care, 75% in our cohort (including 56% of active WSLT). Moreover, limitations were less frequent and occurred later in patients with good anticipated outcomes, supporting a clear impact of the expertise on primary centers' decisions following assessment. While this obviously raises the possibility of self‐fulfilling prophecy [47, 48], these rates are in line with previous studies [8, 49] and the significant association between predicted and actual outcomes persisted after adjustment on WLST in the overall cohort and in the subset of patients without WLST. Third, despite relying on a large set of validated neuroprognostic tools, the lack of cognitive‐motor dissociation testing might have lowered the accuracy of our prediction [50]. Implementation of active paradigms could have enriched our evaluation but is complex in the ICU. Recent simpler EEG paradigms [51] will probably increase its accessibility together with a better prediction of recovery [10]. Fourth, we might have underestimated patients' outcomes, as 3 months might be too early to assess functional outcomes [11, 12], especially in patients without anoxic brain injury. Yet, one‐year mortality rates seem to confirm this early poor outcome. An alternative explanation for such severe outcomes could be that the overall patients' prognosis and associated decisions of care are not determined only by consciousness improvement but by other organ dysfunctions and systemic complications, which can impact functional outcomes and survival, and which are not accounted in the current prognostication tools. Overall, even though specialized assessment precised the recovery potential of patients supporting the value of a specialized neurological expertise, multimodal assessment did not allow for a highly accurate individual‐level prediction, in line with previous literature [52, 53]. Our study did not intend to develop a prediction model but to assess a real‐world practice relying on the heuristic qualitative integration of multimodal data, which in addition to moderate prognostic performances, limits transparency, reproducibility and external validation as underlying weighting and integration rules for individual tests are not explicitly operationalized. Future efforts should thus be made to develop robust risk‐based models integrating multiple predictors, including multi‐organ parameters, into quantitative, patient‐level probability estimates to reach an optimal prediction of the vital and functional status and guide decision‐making [54].

Given the complexity and heterogeneity of the clinical situations of patients with persisting DoC in the ICU, and of the neuroprognostication process, as well as the considerable ethical stakes, a hierarchical organization within the healthcare system with primary, secondary, and tertiary centers appears necessary, as recently proposed [55]. This organization would facilitate access to multimodal prognostic tools and the development of multivariable prediction models within a risk‐based decision‐making framework [6].

5. Conclusion

Patients with persisting DoC referred for specialized assessment have an overall poor prognosis. Multimodal expertise allows to refine the estimation of their neurological outcome, even when many of them had previously undergone brain imaging, but it fails to achieve highly accurate individual outcome prediction. These findings nevertheless support the value of specialized care pathways for severe DoC and advocate for their development and structured integration at a national scale.

Author Contributions

Study design, development, and study setup were performed by C.C., E.P.‐R., M.G., T.S., and B.H. Site setup, subject enrolment, data collection, and research governance were performed by all authors. B.H. performed the statistical analysis. Initial draft of the manuscript was by C.C. and B.H. All authors reviewed and commented on this and subsequent versions of the manuscript. All authors vouch for the accuracy and completeness of the data and analyses. All authors read and approved the final manuscript.

Funding

The authors have nothing to report.

Ethics Statement

The study received approval from the Ethics Committee of the Société de Réanimation de Langue Française (SRLF) (reference number 22‐049, dated 10/20/2022). Since this was a retrospective study based on a database, informed consent was waived, but patients and their relatives were informed about the potential use of their data for clinical research purposes. For the same reason, the study was not registered on any clinical trial repository.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: STROBE checklist.

Table S2: Description of the neurological status in the primary care unit.

Table S3: Multimodal expert assessment by emitted prognosis.

Figure S1: Swimmer plot representing individual patient's trajectories following expert assessment.

Figure S2: Functional and survival outcomes according to emitted prognosis stratified on brain injury etiology.

Table S4: 3‐month functional outcome proportional odds ordinal logistic regression multivariable analysis.

Table S5: One‐year survival Cox proportional hazards multivariable analysis.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.