Multimodal assessment improves neuroprognosis performance in clinically unresponsive critical-care patients with brain injury

Prognostic evaluation of unresponsive patients after acute brain injury is challenging and has substantial ethical implications, given that withdrawal of life-sustaining therapies (WLST) is a leading cause of death and depends on predicted outcomes and patient wishes. Disease-specific scores and algorithms exist for common etiologies (for example, anoxia and traumatic brain injury), and recent European and US guidelines recommend multimodal assessments (MMA) that integrate behavioral, neuroimaging and electrophysiological data when initial behavioral assessments are unclear or confounded. However, despite the rationale that more evidence should improve neuroprognosis, this had not been empirically validated in clinical practice, and adding modalities may increase marker discrepancies, potentially leading to decision paralysis or bias. This study aimed to evaluate whether MMA improves the performance of neuroprognostication in clinically unresponsive ICU patients with acute brain injury and whether increasing the number of modalities reduces uncertainty and increases accuracy of long-term functional outcome predictions.

Prior work has provided disease-specific prognostic models (for example, for cardiac arrest and TBI) and guideline recommendations emphasizing MMA for disorders of consciousness. Reviews and guidelines (European Academy of Neurology; AAN/ACRM/BIA) endorse combining behavioral, electrophysiological and neuroimaging markers to improve assessment when behavior is ambiguous. Previous studies showed multimodal approaches improve diagnosis beyond behavior alone in disorders of consciousness and reviewed coma prognostication methods, but empirical evidence that increasing the number of modalities improves prognostic accuracy and reduces uncertainty in routine clinical practice had been lacking. Concerns have also been raised about cognitive biases in neuroprognostication and the self-fulfilling prophecy associated with WLST.

Design and setting: Prospective, monocentric observational cohort (2009–2021) at a tertiary Neuro-ICU (La Pitié-Salpêtrière AP-HP Sorbonne Université, Paris, France). Protocols preregistered (NEURO-DOC/HAO-006/20130409; M-NEURO-DOC, NCT04534777) with ethics approval; informed consent from surrogates; patients who recovered could withdraw. Participants: Consecutive adult patients (18–80 years) with acute brain injury requiring ICU care and disorders of consciousness (coma to minimally conscious state) due to etiologies such as anoxia, trauma, hypoglycemia, stroke, or encephalitis. Inclusion required brain injury evident on CT/MRI. Exclusions: deep sedation (e.g., for intracranial hypertension/status epilepticus), severe known neurodegenerative disease, pregnancy. Non-ICU referrals with subacute/chronic injuries were excluded from this analysis. Multimodal assessment (MMA): Over ~1 week, patients underwent repeated behavioral exams and advanced electrophysiological and imaging studies. A weekly multidisciplinary DoC-team meeting (neurointensivists, neurologists, neurophysiologists, neuroradiologists, neuroscientists) integrated all data qualitatively, considering convergence across modalities and predefined red flags (unusual etiologies, data-quality issues, suspected sensory/motor/language deficits, violations of expected hierarchical ERP patterns). The team issued a consensus prognosis categorized as good, poor or uncertain. Modalities included:

• Behavior: Expert neurological exam; Coma Recovery Scale-revised (CRS-r) daily when feasible; clinical categorization (coma, VS/UWS, MCS−, MCS+, EMCS). FOUR score added in 2011; caregiver collective assessments (DoC-Feeling) and habituation of auditory startle reflex added in 2020.
• Electrophysiology: Standard bedside EEG (12 electrodes) and SSEP when indicated; high-density 256-channel EEG with auditory local–global ERP paradigm (probing N100, MMN, P3a, P3b); quantitative EEG classification (power, complexity, connectivity) implemented in 2015; EEG motor-command cognitive-motor dissociation protocol implemented in 2021.
• Imaging: Structural CT/MRI; diffusion tensor imaging (DTI) white-matter fractional anisotropy (WM-FA) quantitative analysis implemented in 2015; functional imaging (resting-state fMRI in 2013; [18F]FDG-PET metabolic index in 2016) assessing default mode network preservation. Outcomes: Primary outcome was GOS-E at 12 months post-MMA, assessed by blinded structured telephone interview. Favorable outcome defined as GOS-E ≥4 for dichotomized analyses. Decisions regarding life-sustaining therapies (pursuit, withholding, WLST) were recorded; to mitigate self-fulfilling prophecy, patients with WLST after MMA and those with unknown WLST decision were excluded from some analyses. Statistics: Quantitative variables as medians (IQR), Wilcoxon rank-sum tests; categorical variables with Fisher’s exact tests, odds ratios (OR) with 95% CIs. Association between DoC-team prognosis and ordinal GOS-E analyzed by proportional odds logistic regression (ordinal shift analysis). Prognostic performance (sensitivity, specificity, PPV, NPV, accuracy) computed for individual markers and DoC-team prognosis. Trends of uncertainty and accuracy versus number of modalities analyzed by Cochran–Armitage tests. A multivariable classifier (sparse partial least squares discriminant analysis, sPLS-DA; mixOmics) was trained on MMA markers (WLST excluded; n=200) to predict favorable outcome, with internal cross-validation to estimate ROC/AUC; missing data handled by NIPALS. Time-period effects (pre-2016 vs post-2016) and number of modalities (<6 vs ≥6) assessed with multivariable logistic regression and likelihood-ratio tests.

Cohort and baseline: Of 503 included, 349 ICU patients met criteria; 314 had DoC-team reports; 277 had 1-year outcomes. Median age 53.2 (IQR 35.7–63) years; 63.6% male; common etiologies: anoxia 36.4%, TBI 18.9%, stroke 14%. Clinical states at assessment: VS/UWS 42%, MCS 46% (MCS− 23.6%, MCS+ 22.6%), coma 3.2%, EMCS 8.9%. Median delay injury-to-assessment 33 (23–53) days. Favorable outcome (GOS-E ≥4) at 1 year occurred in 16.6%. Association of prognosis with outcome:

• All patients: Good prognosis associated with shift toward higher GOS-E versus poor (common OR 26.76; 95% CI 11.88–64.39; P<0.001) and versus uncertain (OR 3.45; 95% CI 1.92–6.23; P<0.001).
• Excluding WLST (n=80) and unknown decision (n=18): Good prognosis remained strongly associated with better outcomes versus poor (OR 14.57; 95% CI 5.70–40.32; P<0.001) and versus uncertain (OR 2.90; 95% CI 1.56–5.45; P<0.001). Impact of number of modalities on uncertainty and accuracy:
• All patients: Using ≥6 modalities versus <6 reduced uncertain prognoses from 57.5% to 32.1% (OR 0.35; 95% CI 0.21–0.59; P<0.001) and increased accuracy from 66.1% to 84.3% (OR 2.72; 95% CI 1.18–6.47; P=0.011).
• Excluding WLST/unknown: Uncertain decreased from 63.6% to 38.8% (OR 0.36; 95% CI 0.19–0.69; P<0.001); accuracy increased from 50.0% to 73.5% (OR 2.73; 95% CI 1.01–7.61; P=0.040). Effects remained significant after adjustment for time; no interaction between time and modality count. Individual markers vs MMA: Individual markers showed varying sensitivity/specificity (e.g., SSEP or EEG reactivity high sensitivity but lower specificity; ERP global effect opposite). None outperformed the DoC-team MMA prognosis in accuracy. Multivariable classifier: sPLS-DA trained on 200 patients (WLST excluded) achieved accuracy ~60% versus DoC-team 63.5%; training AUC 0.80; cross-validated AUC 0.73 ± 0.01. Performance similar to individual markers and to DoC-team; identified informative features (e.g., age, delay) not evidently driving human decisions. Additional descriptive outcomes: Patients with good prognosis had ~33% favorable outcomes at 1 year (vs ~20% uncertain, 0% poor). Among good prognosis, lower proportion in VS (GOS-E=2) but higher severe disability dependent (GOS-E=3 at 36.4%).

This study provides empirical evidence that integrating multiple behavioral, electrophysiological and neuroimaging modalities into a consensus MMA by an expert DoC team reduces prognostic uncertainty and increases accuracy for long-term functional outcomes in clinically unresponsive ICU patients with acute brain injury. The observed dose–response relationship—more modalities leading to less uncertainty and higher accuracy—supports guideline recommendations advocating multimodal neuroprognostication and argues against reliance on single techniques. These findings are robust to exclusion of patients with WLST to mitigate self-fulfilling prophecy bias and persist after adjusting for temporal trends, suggesting benefits are primarily attributable to the number of modalities rather than growing experience over time. While a data-driven multivariable classifier performed comparably to the expert team, combining objective model outputs with expert synthesis may further optimize prognostication and reduce biases. Clinically, implementing MMA and improving access to electrophysiology and advanced imaging could enhance decision-making and counseling about goals of care in this ethically sensitive context.

A 12-year prospective cohort from a tertiary Neuro-ICU demonstrates that multimodal assessment by an expert team improves neuroprognostication for unresponsive patients with acute brain injury, decreasing uncertainty and increasing accuracy of 1-year functional outcome predictions. Increasing the number of modalities strengthens performance, while individual markers alone are insufficient. A multivariable classifier performs comparably to expert consensus and could be integrated to further systematize MMA. Future work should validate generalizability across centers with varying expertise, refine modality integration (potentially via decision trees or hybrid human–AI systems), and incorporate emerging prognostic tools (e.g., advanced DTI, heart–brain coupling, olfaction, TMS–EEG) to further enhance prediction.

• Generalizability: Single-center, tier-3 French DoC expertise center; complex cases likely overrepresented. Performance may differ in less complex populations and other settings.
• Evolving MMA: Number and type of modalities increased over time, potentially confounding with team experience; although analyses suggest effects are driven by modality count, residual confounding may remain.
• Sample size per modality: Some modalities had small n (e.g., resting-state fMRI, SSEP), limiting power to detect differences in individual-marker performance.
• Underestimation of current MMA: Newer validated tools implemented late or after study period (e.g., motor-command EEG protocol, blink-reflex habituation, updated whole-brain WM-FA) could improve performance if systematically available earlier.
• WLST bias: Despite excluding WLST/unknown decisions in key analyses, self-fulfilling prophecy cannot be fully eliminated; pre-MMA factors related to WLST decisions (e.g., ventilation without tracheostomy/gastrostomy) may bias outcomes toward poor prognosis.
• Qualitative integration: The DoC-team weighting of modalities was qualitative without a standardized decision tree, which may introduce subjectivity and limit reproducibility.