This retrospective study was conducted at the University Hospital Tübingen including all neurological ICU-patients treated with amantadine for DoC between January 2016 and June 2021. Data, including baseline information (age, sex, diagnoses, complications, medications, mechanical ventilation duration, vital parameters, clinical scores, and written documentation of clinical course) were extracted from electronic medical records [IntelliSpace Critical Care and Anesthesia (ICCA) Philips GmbH, Market DACH, Hamburg, Germany]. This study did not include a control group due to the challenges of retrospectively identifying comparable patients who did not receive amantadine. Treatment decisions were based on clinical judgment during routine care, and the absence of standardized protocols for non-treatment made it infeasible to reliably match untreated patients with similar characteristics.

The study was approved by the Ethics Committee of the University Hospital of Tübingen (560/2022BO2) and performed in accordance with the Declaration of Helsinki. Data were accessed for research purposes on December, 13, 2022. The authors had no access to any information that could identify any individual patient, either during or after data collection. Since this was a retrospective study, informed consent was not required per local Ethics Committee’s requirements.

ICU-patients aged ≥18 years treated with amantadine for non-traumatic DoC were included (n = 490), excluding those treated for other indications (Figure 1).

Primary brain disorders included AIS, ICH, SAH, non-traumatic subdural hematoma (SDH), and meningitis/ encephalitis. Secondary brain disorders were related to systemic conditions impacting brain function, such as infection, organ insufficiencies, cardiac arrest/ resuscitation, and metabolic disorders. Combined brain disorders primary brain disorders combined with secondary causes.

The Glasgow Coma Scale (GCS) (11) was the primary measure to assess consciousness. The Richmond Agitation-Sedation Scale (RASS) (12) was used to evaluate sedation levels as part of routine ICU protocols. The Simplified Acute Physiology Score II (SAPS II) was used to evaluate the severity of ICU illness. GCS and RASS were collected during clinical routine, while SAPS-II was calculated automatically in ICCA. The National Institutes of Health Stroke Scale (NIHSS) assessed stroke severity in neurological conditions (13). Level of dependence was rated using the modified Rankin Scale (mRS). Good outcome was defined as mRS 0–2, and mortality as mRS 6. Premorbid functional status was determined using the premorbid mRS (pmRS) (14–17).

According to the hospital’s standard operating procedure (SOP), amantadine was administered intravenously at a daily dosage of 200–400 mg for 3–5 days (100–200 mg twice daily at 08:00 a.m. and 02:00 p.m.). Treatment was initiated only after complete weaning from sedative medications. These in-hospital guidelines were based on manufacturer guidance and prior studies, indicating a response within 3–5 days of treatment initiation (8, 18). Amantadine is approved for the treatment of vigilance reduction in post-comatose states as part of an overall therapeutic concept (18). Despite these SOP recommendations, treatment decisions, including the duration and dosage, were ultimately at the discretion of the treating physician, leading to variability in the administration of amantadine.

Routine EEG-recordings were performed as part of clinical work-up. EEG abnormalities were graded based on the written results (19).

Imaging work-up was performed as part of clinical routine, including computed tomography (CT) and/or magnetic resonance imaging (MRI) with pathological findings categorized as AIS, ICH, cerebral atrophy (diffuse, focal, global), hypoxia (hyperintensity on MRI, loss of gray-white matter differentiation on CT, or sulcal effacement on MRI/CT), microangiopathy, and old lesions (ischemic, hemorrhagic, post-traumatic, and post-surgical lesions).

Adverse effects (AE) associated with amantadine were assessed referring to the prescription form (18). Adverse events were categorized as seizures, neuropsychiatric symptoms, delirium after initiation of amantadine, and any cardiac arrhythmias during the ICU stay.

The primary outcome measure was improved consciousness, defined by an ≥3-point improvement in GCS within 5 days of amantadine treatment. Responders were defined as patients achieving this threshold, with clinical documentation reviewed to verify improvements in alertness. Secondary outcomes included RASS and SAPS II changes, GCS at discharge, good outcome (mRS 0–2) and mortality (mRS 6) at discharge as well as at follow-up. Follow-up assessments were conducted 90 days after the index event, defined as the acute illness or neurological injury through outpatient clinical visits or, in cases where direct follow-up was not feasible, via telephone interviews with patients and/or caregivers.

Metric data were tested for normal distribution using the Kolmogorov–Smirnov test. Depending on the distribution, t-tests or Mann–Whitney U tests were applied for continuous variables, while nominal data were analyzed using chi-square or Fisher’s exact test. Statistical significance was defined as a p-value<0.05. False discovery rate (FDR) adjustments were applied to control for multiple comparisons, and Bonferroni correction was used in the multivariate logistic regression model (20–22). Pearson correlation coefficients were calculated to explore relationships between scores and responder status, with correlations interpreted as weak (0.1–0.2), moderate (0.3–0.4), or strong (≥0.5). Univariate logistic regression assessed the relationship between each independent variable and the response to amantadine treatment or other outcome parameters. Variables with statistical significance in univariate analysis (p < 0.05) or deemed clinically relevant were included in the multivariate logistic regression model. Linearity was assessed using the Box-Tidwell procedure (23). The absence of multicollinearity was confirmed by low correlations among predictor variables (r < 0.6). Outliers were evaluated using case-by-case graphs, leverage, and Cook’s distance (13–15). All statistical analyzed were performed using SPSS (version 28.0.1.1). Safety analysis included all patients, regardless of their inclusion in the primary efficacy analysis.

Data preprocessing included standardization, imputation of missing values using the mean for numerical columns and the most frequent value for categorical columns, and feature engineering. Key variables included age, female sex, pmRS, brain imaging and EEG pathologies, GCS, SAPS II, and etiology of DoC. Interaction terms, polynomial features, and a cumulative pathology score were used. Continuous features, such as GCS, were binned into categories (severe, moderate, mild), and categorical variables, such as sex and primary brain disorders, were one-hot encoded. Additionally, a cumulative pathology score combining imaging and EEG findings was developed to capture the overall pathology burden.

The dataset was imbalanced with respect to the target variable (responder to amantadine treatment). A RandomForestClassifier was chosen for its robustness and ability to handle both numerical and categorical data. The dataset was split (80% training, 20% testing). Numerical features were standardized using StandardScaler to ensure that they were on a comparable scale. Three base models were trained on the engineered features: Gradient Boosting, SVM, and Logistic Regression. These models were evaluated using performance metrics including accuracy, precision, recall, F1-score, confusion matrix, and classification report. To enhance model performance, two ensemble methods were employed: a Voting Classifier and a Stacking Classifier. The Voting Classifier combined the predictions of Gradient Boosting, SVM, and Logistic Regression using soft voting. The Stacking Classifier combined the predictions of the same base models using a Logistic Regression meta-classifier. Machine learning models were developed and analyzed using R (version 4.2.0).

The cohort included 442 patients (mean age 73.2 ± 10.7 years, 41.0% female). Amantadine treatment began on average 5.4 ± 7.2 days after beginning of weaning from mechanical ventilation, with a mean duration of 2.2 ± 2.1 days and a cumulative dosage of 900 mg (IQR 600-1200 mg).

Of the total cohort, 267 patients (60.4%) were classified as responders, and 175 (39.6%) as non-responders. Baseline characteristics, including age, sex, risk factors, brain disorders, NIHSS, GCS on admission to hospital, SAPS II on ICU-admission, were not significantly different between groups. Responders had a significantly lower median pmRS (2 [IQR 1, 2] vs. 2 [1, 3], p < 0.001, Cohen’s d = −0.37). Before treatment, responders had a lower median GCS compared to non-responders (7 [IQR 5–9] vs. 8 [IQR 5–10], p = 0.030, Cohen’s d = −0.18).

Of importance, there were no significant differences in underlying brain disorders comparing responders to non-responders, but a non-significant trend of lower incidence of cardiac arrest/ resuscitation as secondary cause for DoC in responders vs. non-responders (6.7 vs. 12%, OR 0.53, 95% CI: 0.27–1.02, p = 0.057).

Responders exhibited fewer pathological brain imaging findings (45.7 vs. 61.1%, OR 0.56, 95% CI:0.36–0.86, p = 0.008), including significantly fewer hypoxia-related changes (0.4 vs. 5.7%, p < 0.001), and fewer EEG abnormalities (23.6 vs. 40.0%, OR 0.47, 95% CI: 0.28–0.78, p = 0.003) (Table 1).

After treatment initiation, responders demonstrated significantly higher median GCS on day 1 (10 [IQR 7–11] vs. 9 [IQR 5–11], p < 0.001, Cohen’s d = 0.37), higher median RASS (−2 [IQR −3, −1] vs. -3 [IQR –4, −1], p = 0.002, Cohen’s d = 0.21), and lower SAPS II (31.0 ± 11.4 vs. 35.3 ± 12.0, p < 0.001, Cohen’s d = −0.36).

These trends persisted on subsequent days, with responders consistently showing better scores after beginning of amantadine treatment (Figures 2A–C and Table 2). Weaning from mechanical ventilation occurred earlier in responders (4.7 ± 6.6 vs. 6.5 ± 7.9 days, p = 0.014, Cohen’s d = −0.25).

Responders had a significantly lower median mRS at discharge (5 [IQR 4–5] vs., 5 [IQR 5–6], p < 0.001, Cohen’s d = −0.47), with a higher proportion achieving good outcomes at follow-up (4.9 vs. 1.1%, OR 6.14, 95% CI: 1.35–28.01, p = 0.004). Mortality was significantly lower in responders both at discharge (13.5 vs. 27.4%, OR 0.41, 05% CI: 0.25–0.67, p < 0.001) and at follow-up (16.9 vs. 32.0%, OR 0.43, 95% CI: 0.24–0.77, p = 0.002) (Table 3).

No significant differences were observed in the overall complication rate between responders and non-responders (27.7 vs. 31.4%, RR 0.89, 95% CI: 0.66–1.18, p = 0.402). Seizures, categorized as occurring after amantadine initiation, were more frequent in non-responders (10.1 vs. 16.6%, OR 0.48, 95% CI: 0.29–0.81, p = 0.046). Responders experienced a higher incidence of delirium (33.0 vs. 22.3%, OR 1.47, 95% CI: 1.06–2.03, p = 0.015). Any cardiac arrythmias during ICU-stay occurred in 37.8 vs. 33.7% (p = 0.38) (Table 4).

Significant correlations were observed between GCS and SAPS II scores, with higher GCS scores associated with lower SAPS II scores (r = −0.643, p < 0.001). Moderate correlations were found between RASS and both GCS and SAPS II. Strong negative correlations were observed between GCS scores and mRS at discharge, indicating that better neurological function was associated with less disability. Similar negative correlations between RASS and mRS at discharge suggested that less sedation or agitation was linked to better functional outcomes. Positive correlations between SAPS II and mRS indicated that greater illness severity predicted worse functional outcomes.

Additionally, GCS scores were strongly negatively correlated with in-hospital mortality, suggesting that higher GCS scores were associated with a lower likelihood of death. Negative correlations between RASS scores and in-hospital mortality indicated that better RASS scores were linked to reduced mortality risk, while positive correlations between SAPS II and mortality showed that greater illness severity was associated with higher mortality rates.

Multivariate logistic regression revealed that responders had significantly lower odds of in-hospital mortality (OR 0.468, 95% CI: 0.257–0.851, p = 0.013), while experiencing cardiac arrest increased mortality risk (OR 0.334, 95% CI: 0.150–0.745, p = 0.007). Females had significantly lower mortality odds compared to males (OR 0.493. 95% CI: 0.260–0.935, p = 0.030). Other factors, including age, pmRS, GCS at admission, total amantadine dosage, pre-existing dementia, and pathological imaging results, did not significantly affect mortality (Figure 3A).

For good outcomes at follow-up, higher pmRS was associated with lower odds of achieving a good outcome (OR 0.740, 95% CI: 0.630–0.869, p < 0.001), while being a responder to amantadine was significantly increased the odds of good outcome (OR 8.189, 95% CI: 1.130–59.356, p = 0.037) (Figure 3B).

In univariate analysis, higher pmRS (OR 0.740, 95% CI: 0.630–0.869, p < 0.001), and pathological brain imaging (OR 0.558, 95% CI: 0.363–0.860, p = 0.008) were associated with a lower likelihood of response to amantadine (Table 5).

In multivariate analysis, higher pmRS remained a significant predictor of lower response odds (OR 0.719, 95% CI 0.590–0.875, p < 0.001), while pathological brain imaging (OR 0.546, 95% CI 0.342–0.871, p = 0.011), and cardiac arrest (OR 0.542, 95% CI 0.307–0.954, p = 0.034) also reduced the likelihood of response (Figure 3C).

The Gradient Boosting Classifier achieved an accuracy of 58.1%, a precision of 55.2%, a recall of 58.1%, and an F1 score of 53.7%. The SVM Classifier achieved an accuracy of 55.9%, a precision of 54.6%, a recall of 55.9%, and an F1 score of 53.9%. The Logistic Regression model showed slightly better performance with an accuracy of 59.1%, a precision of 56.3%, a recall of 59.1%, and an F1 score of 53.3%.

Ensemble models improved overall performance. The Voting Classifier achieved an accuracy of 58.1%, a precision of 57.1%, a recall of 58.1%, and an F1 score of 55.76%. The confusion matrix for the Voting Classifier was [[16, 26, 0], [11, 38, 0], [0, 2, 0]]. The Stacking Classifier demonstrated the best overall performance, with an accuracy of 64.5%, a precision of 66.6%, a recall of 64.5%, and an F1 score of 61.3%. The confusion matrix for the Stacking Classifier was [[16, 26, 0], [5, 44, 0], [0, 2, 0]]. In summary, the Stacking Classifier outperformed the other models, showing the highest accuracy and F1 score. Ensemble methods, particularly the Stacking Classifier, improved performance over individual models, especially in handling majority classes. However, all models struggled with predicting the minority class (Class 2), indicating the need for further balancing or additional feature engineering. Balancing the dataset significantly improved the model’s accuracy, precision, recall, and F1-score. This was also shown by the higher AUC values. The confusion matrices and classification reports illustrated that the balanced model performed well across all classes, including the previously underrepresented class 2 (Figure 4).