Integrating serum neurotrophin profiles and quantitative EEG for predicting motor outcomes in acute ischemic stroke

Abstract

Background and objectives

Accurate prediction of motor recovery after acute ischemic stroke remains challenging. We investigated whether integrating serum neurotrophin profiles with quantitative electroencephalography (qEEG) parameters enhances predictive accuracy for 6-month motor outcomes beyond conventional clinical predictors.

Methods

This prospective observational study enrolled 108 patients with first-ever acute ischemic stroke admitted within 24 h of symptom onset between December 2023 and December 2024. Serum brain-derived neurotrophic factor (BDNF) and nerve growth factor (NGF) were measured at admission (Day 1) and Day 3. qEEG recording was performed within 72 h, with calculation of delta-alpha ratio (DAR), delta-theta/alpha-beta ratio (DTABR), and brain symmetry index (BSI). The primary outcome was motor function at 6 months assessed by Fugl-Meyer Assessment upper extremity score (FMA-UE), dichotomized as good (FMA-UE ≥ 50) or poor (FMA-UE < 50) outcome. Multivariate logistic regression models were developed and compared using receiver operating characteristic analysis.

Results

At 6-month follow-up, 64 patients (59.3%) achieved good motor outcome. Patients with good outcome demonstrated significantly higher BDNF levels at Day 1 (14.05 ± 5.30 vs. 10.97 ± 4.58 ng/mL, P = 0.002) and lower DAR (median 0.42 [IQR 0.22–0.61] vs. 0.83 [0.56–1.16], P < 0.001) compared with poor outcome group. The clinical-only model (Model 1: comprising age, sex, NIHSS, ASPECTS, reperfusion therapy, and diabetes mellitus) achieved a cross-validated area under the receiver operating characteristic curve (AUC) of 0.884 (95% CI: 0.862–0.907). Addition of neurotrophin biomarkers improved cross-validated AUC to 0.903 (95% CI: 0.877–0.928), while addition of qEEG parameters yielded cross-validated AUC of 0.922 (95% CI: 0.901–0.943). The combined model integrating clinical variables, neurotrophins, and qEEG achieved the highest discriminative performance (cross-validated AUC = 0.930, 95% CI: 0.908–0.952) with sensitivity of 92.2% and specificity of 93.2% at the optimal threshold.

Conclusions

A multimodal prediction model combining serum neurotrophin profiles and qEEG parameters provides incremental improvement in prognostic accuracy for motor outcomes in acute ischemic stroke, though differences in discriminative performance between models did not reach statistical significance in this sample. This integrated approach may facilitate individualized rehabilitation planning and resource allocation in clinical practice.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12883-026-05048-4.

Introduction

Stroke remains a leading cause of long-term disability worldwide, with motor impairment representing the most prevalent and debilitating sequela [1, 2]. Despite advances in acute stroke treatment, including intravenous thrombolysis and mechanical thrombectomy, a substantial proportion of survivors experience persistent motor deficits that profoundly impact functional independence and quality of life [3, 4]. Early and accurate prediction of motor recovery trajectories is essential for optimizing rehabilitation strategies, setting realistic therapeutic goals, and allocating healthcare resources appropriately.

Conventional predictors of motor outcome, including initial stroke severity, lesion characteristics, and demographic factors, provide moderate prognostic accuracy but fail to capture the dynamic biological processes underlying neural recovery [5, 6]. Recent investigations have highlighted the potential of circulating biomarkers to reflect neuroplasticity mechanisms operative during stroke recovery. Brain-derived neurotrophic factor (BDNF), a member of the neurotrophin family, plays a pivotal role in synaptic plasticity, neuronal survival, and motor learning processes critical for post-stroke rehabilitation [7–9]. Accumulating evidence suggests that serum BDNF levels correlate with functional outcomes following ischemic stroke, although findings have been inconsistent across studies [10–12].

Nerve growth factor (NGF), another key neurotrophin, promotes neuronal differentiation and survival while modulating inflammatory responses following cerebral ischemia [13, 14]. Preliminary data indicate that elevated serum NGF concentrations associate with favorable functional recovery, suggesting complementary prognostic value to BDNF assessment [15]. However, the combined predictive utility of multiple neurotrophin biomarkers for motor-specific outcomes remains inadequately characterized.

Quantitative electroencephalography (qEEG) offers a non-invasive window into cortical electrophysiology that may capture neural substrate integrity relevant to motor recovery [16, 17]. Power spectral analysis–derived indices, particularly the delta-alpha ratio (DAR) and delta-theta/alpha-beta ratio (DTABR), have demonstrated robust correlations with stroke severity and functional prognosis [18–20]. Elevated DAR values, reflecting pathological delta activity relative to physiological alpha rhythms, indicate cortical dysfunction predictive of poor outcomes [21]. The brain symmetry index (BSI) quantifies interhemispheric asymmetry and provides additional information regarding focal cortical pathology [22].

While both neurotrophin biomarkers and qEEG parameters have shown individual promise for stroke prognostication, their combined predictive value for motor outcomes specifically has not been systematically evaluated. We hypothesized that integrating serum neurotrophin profiles with qEEG parameters would enhance predictive accuracy beyond either modality alone or conventional clinical variables. The present study aimed to develop and validate a multimodal prediction model for 6-month motor outcomes in acute ischemic stroke incorporating BDNF, NGF, and qEEG-derived indices.

Methods

Study design and participants

This prospective observational cohort study was conducted at Nantong University Affiliated Hospital between December 2023 and December 2024. We consecutively enrolled patients meeting the following inclusion criteria: (1) age 18–85 years; (2) first-ever acute ischemic stroke confirmed by computed tomography or magnetic resonance imaging; (3) admission within 24 h of symptom onset; (4) baseline National Institutes of Health Stroke Scale (NIHSS) score ≥ 4; and (5) ability to provide informed consent or availability of a legal representative. Exclusion criteria comprised: (1) hemorrhagic stroke or hemorrhagic transformation; (2) prior stroke within the preceding 3 months; (3) pre-existing severe cognitive impairment precluding assessment; (4) bilateral hemispheric lesions; (5) terminal illness with life expectancy less than 6 months; and (6) conditions precluding EEG recording or blood sampling; and (7) absence of motor impairment at baseline, defined as NIHSS motor arm score = 0 and NIHSS motor leg score = 0 on admission. All 108 enrolled patients demonstrated at least unilateral upper or lower extremity motor deficit at the time of admission, confirming the clinical relevance of motor outcome prediction in this cohort.

The study protocol was approved by the Institutional Ethics Committee of Nantong University (approval number: 2022-K089) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants or their legal representatives.

Clinical assessment

Baseline demographic and clinical data were extracted from medical records, including age, sex, vascular risk factors (hypertension, diabetes mellitus, atrial fibrillation), stroke etiology classified by Trial of Org 10,172 in Acute Stroke Treatment criteria, lesion location, and acute treatment received (intravenous thrombolysis, mechanical thrombectomy, or conservative management).

Stroke severity was quantified using the National Institutes of Health Stroke Scale (NIHSS), a validated 15-item neurological examination tool with scores ranging from 0 (no deficit) to 42 (severe deficit) [23]. The NIHSS was developed by the National Institute of Neurological Disorders and Stroke and is widely used to assess acute stroke severity [24]. The Alberta Stroke Programme Early CT Score (ASPECTS) was applied as a standardized 10-point quantitative topographic scoring system to evaluate early ischemic changes on non-contrast CT imaging [25]. ASPECTS was developed and validated by Barber et al. to predict functional outcomes and hemorrhagic risk in acute ischemic stroke [26].

Serum neurotrophin measurement

Peripheral venous blood samples were collected at two time points: within 24 h of admission (Day 1) and at 72 h post-admission (Day 3). Samples were centrifuged at 3000 rpm for 15 min, and serum aliquots were stored at -80 °C until batch analysis. Serum BDNF concentrations were measured using a commercially available enzyme-linked immunosorbent assay kit (R&D Systems, Minneapolis, MN, USA) with a detection range of 0.25-32 ng/mL and intra-assay coefficient of variation < 6%. Serum NGF levels were quantified using a high-sensitivity ELISA kit (Abcam, Cambridge, UK) with detection range of 7.8–500 pg/mL and intra-assay coefficient of variation < 8%. All measurements were performed in duplicate by laboratory personnel blinded to clinical outcomes. The change in BDNF from Day 1 to Day 3 (ΔBDNF) was calculated as an indicator of early neurotrophin dynamics.

Quantitative EEG recording and analysis

Continuous EEG was recorded within 72 h of stroke onset using a 19-channel digital EEG system (Nihon Kohden, Tokyo, Japan) with electrodes placed according to the international 10–20 system, which represents the standard clinical EEG montage routinely used in our neurophysiology unit. Recordings were obtained for a minimum of 20 min with the patient in a relaxed, eyes-closed, wakeful state. The sampling rate was 256 Hz with bandpass filtering at 0.5–70 Hz. The ground electrode was placed at Fpz; all recordings were referenced to linked earlobes (A1–A2) and no re-referencing was applied during offline analysis. Artifact-free epochs were defined as continuous segments free of eye-movement, muscle activity, movement, or electrode artifacts, identified by visual inspection of the raw EEG traces by a trained neurophysiologist blinded to clinical data. A minimum of 5 contiguous minutes of artifact-free data were required for inclusion; three patients (of 111 who underwent EEG recording) were excluded from the final analysis due to persistent movement or muscle artefacts precluding ≥ 5 min of usable data (2.7% exclusion rate). Brief artifactual intervals shorter than 30 s were excised, and the flanking artifact-free segments were retained for analysis. To avoid introducing spectral discontinuities at the junctions of rejoined segments, power spectral density was not estimated from a raw concatenated time series; instead, each continuous artifact-free segment was divided into shorter 2-s epochs with 50% overlap, a Hann (Hanning) window was applied to every epoch, and the periodograms were averaged across all artifact-free epochs using Welch’s method to obtain the power spectrum for each channel. Because each epoch was tapered to zero at its edges and no epoch crossed a segment boundary, neither the removal of brief artifactual intervals nor the rejoining of clean segments introduced edge artifacts into the spectral estimates. Power spectral and all qEEG parameters (DAR, DTABR) were derived from all 19 channels and averaged across channels; BSI was computed from the eight standard homologous electrode pairs (Fp1/Fp2, F3/F4, C3/C4, P3/P4, O1/O2, F7/F8, T3/T4, T5/T6).

Power spectral analysis was performed using Welch’s method (averaging the fast Fourier transforms of overlapping 2-s Hann-tapered epochs) with computation of absolute and relative power in standard frequency bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz). The following qEEG indices were calculated: [18, 27]

Brain Symmetry Index (BSI) was computed using the pairwise-derived method (distinct from the original single-value BSI described by van Putten et al.), calculated as the normalized absolute difference in power between homologous electrode pairs across the affected and unaffected hemispheres, averaged across all frequency bands (1–25 Hz) [22]. This pairwise approach provides an electrode-specific measure of interhemispheric asymmetry, as previously described by Agius Anastasi et al. [22]. All qEEG analyses were conducted by a neurophysiologist blinded to clinical and biomarker data.

Outcome assessment

The primary outcome was motor function at 6 months post-stroke, assessed using the Fugl-Meyer Assessment upper extremity subscale (FMA-UE) [28, 29]. The FMA-UE is a well-established, internationally recognized stroke-specific performance-based assessment tool originally developed by Fugl-Meyer et al. in 1975 and has been extensively validated across multiple stroke populations [30]. This instrument demonstrates excellent psychometric properties including inter-rater reliability (ICC > 0.95) and construct validity for measuring upper limb motor recovery post-stroke [31].

The FMA-UE comprises 33 items assessing reflexes, voluntary movement, and coordination, with scores ranging from 0 (complete hemiplegia) to 66 (normal motor function). Assessments were performed by trained physical therapists blinded to baseline biomarker and qEEG results. Motor outcome was dichotomized as good (FMA-UE ≥ 50, indicating mild impairment with functional capacity) or poor (FMA-UE < 50) based on established thresholds correlating with functional independence [32, 33].

Secondary outcomes included the total Fugl-Meyer motor score (FMA-Motor, range 0-100) incorporating lower extremity assessment, and the modified Rankin Scale (mRS) at 6 months. The mRS is a widely used global disability scale ranging from 0 (no symptoms) to 6 (death), originally described by Rankin in 1957 and subsequently modified by van Swieten et al. in 1988 [34]. This scale has become the most commonly applied primary outcome measure in acute stroke trials and is recommended by the American Heart Association/American Stroke Association for assessing functional outcomes [35].

Statistical analysis

Normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed continuous variables were expressed as mean ± standard deviation and compared between outcome groups using independent samples t-test; non-normally distributed variables were expressed as median (interquartile range, IQR) and compared using Mann-Whitney U test. Categorical variables were presented as frequencies and percentages and compared using chi-square test or Fisher exact test as appropriate. Correlations between continuous variables and 6-month FMA-UE scores were assessed using Spearman rank correlation coefficients (Fig. 1).

Fig. 1.

Correlation matrix heatmap showing Spearman correlation coefficients between key study variables. Color intensity represents correlation strength (red: negative correlation; blue: positive correlation). Cross symbols (×) indicate correlations that did not achieve statistical significance (P ≥ 0.05). FMA-UE 6 M: Fugl-Meyer Assessment upper extremity at 6 months; BL: baseline. Notable findings include moderate positive correlation between BDNF and motor outcome, and strong inverse correlations between qEEG parameters (DAR, DTABR, BSI) and functional recovery

Four prediction models were developed using multivariate logistic regression: Model 1 (clinical variables: age, sex, NIHSS, ASPECTS, reperfusion therapy, diabetes mellitus); Model 2 (clinical variables plus neurotrophins: BDNF Day 1, NGF Day 1); Model 3 (clinical variables plus qEEG: DAR, DTABR, BSI); and Model 4 (combined model: clinical variables plus neurotrophins plus qEEG). Continuous predictors were standardized prior to model fitting. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) with comparison of nested models visualized in Fig. 2. Internal validation was performed using 5-fold stratified cross-validation, repeated 100 times. In each of the 100 repetitions, patients were randomly and independently partitioned into 5 stratified folds (with stratification by motor outcome to preserve class balance); a distinct random seed (integer seeds 1–100) was assigned to each repetition to ensure full reproducibility while sampling the space of possible fold assignments. All model fitting was performed exclusively on the training partition (four folds, ~ 80% of patients); performance was assessed only on the held-out validation fold (~ 20% of patients), with no predictions derived from models trained on the full dataset included in the cross-validated metrics. This procedure yielded 500 held-out validation AUC values per model (100 repetitions × 5 folds). The reported cross-validated AUC is the mean of these 500 held-out values; 95% confidence intervals were derived from the 2.5th and 97.5th percentiles of the resulting distribution. AUC comparisons between nested models were conducted using the DeLong test [36].

Fig. 2.

Receiver operating characteristic curves comparing prediction models for 6-month motor outcome. Model 1 (gray line): clinical variables only (AUC = 0.914); Model 2 (orange line): clinical + neurotrophins (AUC = 0.935); Model 3 (green line): clinical + qEEG (AUC = 0.952); Model 4 (blue line): combined model (AUC = 0.967). The combined multimodal model demonstrates superior discriminative performance. Dashed line represents the reference (AUC = 0.5)

As a pre-specified sensitivity analysis, baseline upper-extremity motor function (baseline FMA-UE) was additionally incorporated into all four prediction models, designated Models 1–4 S, using the identical 5-fold stratified cross-validation framework (100 repetitions); the discriminative ability of baseline FMA-UE alone was also evaluated by receiver operating characteristic analysis, with the optimal threshold determined by the Youden index.

The discriminative ability of individual biomarkers was assessed separately using ROC analysis (Fig. 3). Specifically, each biomarker was evaluated via univariable logistic regression, and the area under the ROC curve was derived from the model-predicted probabilities. The same 5-fold stratified cross-validation framework (100 repetitions) was applied to individual biomarker assessments to generate internally validated AUC estimates for each predictor. Model calibration was evaluated using calibration plots comparing predicted probabilities against observed outcomes (Fig. 4). Optimal probability thresholds were determined using the Youden index (sensitivity + specificity − 1). For individual biomarkers, optimal cut-off values for predicting good motor outcome were identified from ROC analysis. Odds ratios (OR) with 95% confidence intervals (CI) were calculated for predictors in the final multivariate model and visualized as a forest plot (Fig. 5). A two-sided P value < 0.05 was considered statistically significant. All analyses were performed using Python 3.10 with scikit-learn 1.3.0 and SciPy 1.11.0, with visualization generated using R 4.5.2.

Fig. 3.

Receiver operating characteristic curves for individual biomarkers predicting good motor outcome at 6 months (univariable logistic regression). Each curve represents the discriminative performance of a single biomarker: BDNF Day 1 (AUC = 0.666, 95% CI: 0.643–0.687), BDNF Day 3 (AUC = 0.712, 95% CI: 0.689–0.733), NGF Day 1 (AUC = 0.581, 95% CI: 0.559–0.606), NGF Day 3 (AUC = 0.670, 95% CI: 0.651–0.693), delta-alpha ratio (DAR, AUC = 0.767, 95% CI: 0.746–0.788), delta-theta/alpha-beta ratio (DTABR, AUC = 0.719, 95% CI: 0.695–0.742), brain symmetry index (BSI, AUC = 0.812, 95% CI: 0.793–0.831), NIHSS at baseline (AUC = 0.781, 95% CI: 0.764–0.798), baseline FMA-UE score (AUC = 0.728, 95% CI: 0.708–0.748), and baseline FMA-Motor score (AUC = 0.816, 95% CI: 0.797–0.835). All AUC values represent cross-validated estimates (mean ± 1.96 × SD across 100 repetitions of 5-fold stratified cross-validation). Optimal thresholds and corresponding sensitivity/specificity are provided in the updated Table 3

Fig. 4.

Calibration plot for the combined prediction model (Model 4). The x-axis represents mean predicted probability of good motor outcome, and the y-axis represents observed frequency of good outcome. Point sizes are proportional to sample size in each probability decile. The dashed diagonal line represents perfect calibration. The calibration slope of 1.01 indicates good agreement between predicted probabilities and observed outcomes, with minimal evidence of overfitting. Shaded area represents 95% confidence interval

Fig. 5.

Forest plot displaying odds ratios and 95% confidence intervals from multivariate logistic regression analysis for the combined prediction model. Square markers indicate statistically significant predictors (P < 0.05); circle markers indicate non-significant variables. Colors represent variable categories: gray for clinical variables, orange for neurotrophin biomarkers, and green for qEEG parameters. Higher ASPECTS scores and BDNF Day 1 levels are associated with good motor outcome, while higher NIHSS scores and qEEG parameters (DAR, DTABR, BSI) predict poor outcome

Results

Study population

Of 142 patients screened, 108 met inclusion criteria and completed 6-month follow-up. The study cohort comprised 65 men (60.2%) with a median age of 65.65 years (IQR 58.98–74.35). Baseline NIHSS scores ranged from 4 to 23 (median 8 [IQR 6–11]), and 52.8% of patients received reperfusion therapy. At 6-month follow-up, 64 patients (59.3%) achieved good motor outcome (FMA-UE ≥ 50) while 44 patients (40.7%) experienced poor motor outcome.

Baseline characteristics by outcome group

Demographic and clinical characteristics stratified by motor outcome are presented in Table 1. Patients with good motor outcome were slightly younger (median 65.20 [IQR 57.65–73.00] vs. 66.40 [60.08–75.53] years, P = 0.299), although this difference did not reach statistical significance. No significant difference in sex distribution was observed between groups (60.9% vs. 59.1% male, P = 1.000). Patients achieving good outcome demonstrated significantly lower baseline NIHSS scores (median 7 [IQR 5–9] vs. 10 [8–12], P < 0.001) and higher ASPECTS scores (median 8 [IQR 8–9] vs. 7 [6–7], P < 0.001), reflecting less severe initial stroke burden.

Table 1.

Baseline Demographic and Clinical Characteristics by Motor Outcome

Variable	Good Outcome (n = 64)	Poor Outcome (n = 44)	P Value
Age, years	65.20 (57.65–73.00)	66.40 (60.08–75.53)	0.299
Male sex, n (%)	39 (60.9)	26 (59.1)	1.000
NIHSS score	7.00 (5.00–9.25)	10.00 (8.00–12.00)	< 0.001
ASPECTS score	8.00 (7.75–9.00)	7.00 (6.00–7.25)	< 0.001
Onset to admission, h	3.23 ± 3.04	3.29 ± 3.12	0.968
Hypertension, n (%)	43 (67.2)	35 (79.5)	0.234
Diabetes mellitus, n (%)	22 (34.4)	21 (47.7)	0.233
Atrial fibrillation, n (%)	22 (34.4)	13 (29.5)	0.751
Reperfusion therapy, n (%)	40 (62.5)	17 (38.6)	0.025
Baseline FMA-UE score	31.33 ± 13.78	20.98 ± 10.62	< 0.001
Baseline FMA-LE score	19.53 ± 5.50	13.43 ± 6.15	< 0.001

Abbreviations: NIHSS National Institutes of Health Stroke Scale, ASPECTS Alberta Stroke Programme Early CT Score, FMA-UE Fugl-Meyer Assessment upper extremity, FMA-LE Fugl-Meyer Assessment lower extremity

Normally distributed continuous variables are presented as mean ± SD (compared by independent-samples t-test); non-normally distributed variables as median (IQR) (compared by Mann-Whitney U test). Categorical variables as n (%)

Regarding vascular risk factors, diabetes mellitus was more prevalent in the poor outcome group, though not statistically significant (47.7% vs. 34.4%, P = 0.233). Patients with good outcome were significantly more likely to have received reperfusion therapy (62.5% vs. 38.6%, P = 0.025). Baseline motor impairment as measured by FMA-UE was markedly less severe in the good outcome group (31.33 ± 13.78 vs. 20.98 ± 10.62, P < 0.001).

Neurotrophin and qEEG parameters

Serum neurotrophin levels and qEEG parameters according to motor outcome are summarized in Table 2, with distributions illustrated in Fig. 6. Patients achieving good motor outcome exhibited significantly higher serum BDNF concentrations at Day 1 (14.05 ± 5.30 vs. 10.97 ± 4.58 ng/mL, P = 0.002) and Day 3 (median 14.25 [IQR 10.84–19.63] vs. 9.87 [6.61–14.45] ng/mL, P < 0.001). Notably, the change in BDNF from Day 1 to Day 3 (ΔBDNF) was significantly greater in the good outcome group (median 0.71 [–0.58 to 2.68] vs. − 0.38 [–1.23 to 0.53] ng/mL, P = 0.003).

Table 2.

Neurotrophin and qEEG Parameters by Motor Outcome

Variable	Good Outcome (n = 64)	Poor Outcome (n = 44)	P Value
Serum Neurotrophins
BDNF Day 1, ng/mL	14.05 ± 5.30	10.97 ± 4.58	0.002
BDNF Day 3, ng/mL	14.25 (10.84–19.63)	9.87 (6.61–14.45)	< 0.001
ΔBDNF, ng/mL	0.71 (-0.58–2.68)	-0.38 (-1.23–0.53)	0.003
NGF Day 1, pg/mL	46.90 ± 15.78	42.34 ± 15.58	0.142
NGF Day 3, pg/mL	51.00 (38.89–61.68)	40.28 (30.08–50.74)	0.003
qEEG Parameters
DAR	0.42 (0.22–0.61)	0.83 (0.56–1.16)	< 0.001
DTABR	0.43 (0.25–0.70)	0.72 (0.46–0.97)	< 0.001
BSI	0.16 (0.11–0.24)	0.31 (0.23–0.45)	< 0.001
Relative alpha power, %	21.10 ± 5.68	15.49 ± 5.95	< 0.001
Relative delta power, %	28.87 ± 8.12	41.46 ± 8.74	< 0.001

Abbreviations: BDNF brain-derived neurotrophic factor, NGF nerve growth factor, qEEG quantitative electroencephalography, DAR delta-alpha ratio, DTABR delta-theta/alpha-beta ratio, BSI brain symmetry index

Normally distributed continuous variables are presented as mean ± SD (compared by independent-samples t-test); non-normally distributed variables as median (IQR) (compared by Mann-Whitney U test). Categorical variables as n (%)

Fig. 6.

Distribution of biomarker levels by 6-month motor outcome. Raincloud plots combining density distributions, box plots, and individual data points comparing patients with good outcome (blue, n = 64) versus poor outcome (red, n = 44). Panel A: BDNF Day 1 concentration; Panel B: NGF Day 1 concentration; Panel C: Delta-alpha ratio (DAR); Panel D: Delta-theta/alpha-beta ratio (DTABR); Panel E: Brain symmetry index (BSI); Panel F: Relative delta power. P values from Mann-Whitney U test are displayed for each comparison

Serum NGF levels at Day 1 showed a trend toward higher values in the good outcome group (46.90 ± 15.78 vs. 42.34 ± 15.58 pg/mL), although this difference did not achieve statistical significance (P = 0.142). However, NGF concentrations at Day 3 were significantly elevated in patients with favorable motor recovery (median 51.00 [IQR 38.89–61.68] vs. 40.28 [30.08–50.74] pg/mL, P = 0.003).

All qEEG parameters demonstrated highly significant differences between outcome groups (Figs. 6 and 7). Patients with poor motor outcome exhibited substantially elevated DAR (median 0.83 [IQR 0.56–1.16] vs. 0.42 [0.22–0.61], P < 0.001) and DTABR (median 0.72 [IQR 0.46–0.97] vs. 0.43 [0.25–0.70], P < 0.001) values, indicating pathological enhancement of slow-wave activity relative to faster rhythms. Similarly, BSI was significantly higher in the poor outcome group (median 0.31 [IQR 0.23–0.45] vs. 0.16 [0.11–0.24], P < 0.001), reflecting greater interhemispheric asymmetry.

Fig. 7.

Faceted violin plots with embedded box plots comparing key clinical and biomarker variables between motor outcome groups. Each panel displays the distribution of a specific variable for good outcome (blue) and poor outcome (red) groups: BDNF concentration, NGF concentration, delta-alpha ratio (DAR), delta-theta/alpha-beta ratio (DTABR), brain symmetry index (BSI), relative alpha power, relative delta power, and baseline NIHSS score. This visualization facilitates comprehensive comparison of variable distributions across outcome groups, highlighting the separation between groups across multiple domains

Correlations with 6-month motor function

Spearman correlation analysis revealed significant associations between biomarkers and 6-month FMA-UE scores (Fig. 1). Among neurotrophins, BDNF Day 1 demonstrated a moderate positive correlation with motor outcome (ρ = 0.308, P = 0.001), while NGF Day 1 showed a weaker but significant association (ρ = 0.194, P = 0.045). The relationship between BDNF and motor outcome is further illustrated in Fig. 8, showing that patients with higher BDNF levels generally achieved higher FMA-UE scores at 6 months, with clear separation between outcome groups.

Fig. 8.

Scatter plot with linear regression showing the relationship between serum BDNF Day 1 concentration and 6-month FMA-UE score. Individual patients are represented as points colored by motor outcome (blue: good outcome, FMA-UE ≥ 50; red: poor outcome, FMA-UE < 50). The blue line represents the linear regression fit with 95% confidence interval (shaded area). Horizontal dashed line indicates the good outcome threshold (FMA-UE = 50). Spearman correlation coefficient (ρ = 0.308, P = 0.001) demonstrates a significant positive association between admission BDNF levels and motor recovery

The qEEG parameters exhibited stronger correlations: BSI demonstrated the strongest inverse association (ρ = -0.504, P < 0.001), followed by DAR (ρ = -0.443, P < 0.001) and DTABR (ρ = -0.307, P = 0.001). As expected, baseline NIHSS correlated inversely with motor recovery (ρ = -0.337, P < 0.001), while baseline FMA-UE showed the strongest positive correlation (ρ = 0.696, P < 0.001).

Prediction model performance

Receiver operating characteristic analysis demonstrated incremental improvement in discriminative ability with addition of biomarker modalities (Fig. 2; Table 3). For each model, two AUC estimates are reported: the full-sample AUC (model trained and evaluated on all 108 patients, an optimistic in-sample estimate) and the cross-validated AUC (mean of 100 repetition-level mean AUCs from 5-fold stratified cross-validation, primary performance metric). 95% confidence intervals for cross-validated AUCs are calculated as mean ± 1.96 × SD across the 100 repetition means. The clinical-only model (Model 1) achieved a full-sample AUC of 0.915 and a cross-validated AUC of 0.884 (95% CI: 0.862–0.907). Addition of neurotrophin biomarkers (Model 2) improved the cross-validated AUC to 0.903 (95% CI: 0.877–0.928; full-sample AUC: 0.935). The clinical plus qEEG model (Model 3) yielded a cross-validated AUC of 0.922 (95% CI: 0.901–0.943; full-sample AUC: 0.952). The combined multimodal model (Model 4) demonstrated the highest discriminative performance with a cross-validated AUC of 0.930 (95% CI: 0.908–0.952; full-sample AUC: 0.968).

Table 3.

Prediction Model Performance and Multivariate Logistic Regression Results A. Model Discrimination Performance

A. Model Discrimination Performance
Model	Full-sample AUC	CV-AUC (95% CI)	ΔAUC
Model 1: Clinical only	0.915	0.884 (0.862–0.907)	Ref
Model 2: Clinical + Neurotrophins	0.935	0.903 (0.877–0.928)	+ 0.019
Model 3: Clinical + qEEG	0.952	0.922 (0.901–0.943)	+ 0.038
Model 4: Combined	0.968	0.930 (0.908–0.952)	+ 0.046

B. Multivariate Analysis - Combined Model (Model 4)
Variable	OR	95% CI	P Value
Age (per SD)	0.39	0.14-1.12	0.081
Male sex	1.78	0.74-4.27	0.195
NIHSS (per SD)	0.15	0.06-0.43	<0.001
ASPECTS (per SD)	8.41	2.63-26.84	<0.001
Reperfusion therapy	0.75	0.31-1.80	0.521
Diabetes mellitus	1.12	0.53-2.37	0.759
BDNF Day 1 (per SD)	3.05	1.30-7.14	0.010
NGF Day 1 (per SD)	1.82	0.81-4.12	0.150
DAR (per SD)	0.31	0.11-0.92	0.036
DTABR (per SD)	0.44	0.21-0.92	0.030
BSI (per SD)	0.38	0.14-1.00	0.050

Abbreviations: AUC area under the receiver operating characteristic curve, CV cross-validated, CI confidence interval, OR odds ratio, SD standard deviation, NIHSS National Institutes of Health Stroke Scale, ASPECTS Alberta Stroke Programme Early CT Score, BDNF brain-derived neurotrophic factor, NGF nerve growth factor, DAR delta-alpha ratio, DTABR delta-theta/alpha-beta ratio, BSI brain symmetry index

Comparison of cross-validated AUCs between nested models using the DeLong test revealed that the combined model numerically outperformed the clinical-only model (ΔCV-AUC = 0.046, P = 0.101), and the Δfull-sample AUC was 0.052 (Table 3A). Although these differences did not reach conventional statistical significance, likely reflecting the limited sample size, the combined model demonstrated numerically superior performance across all 100 cross-validation repetitions, supporting the incremental clinical utility of the multimodal approach.

At the optimal probability threshold of 0.612 determined by Youden index, the combined model achieved sensitivity of 92.2% and specificity of 93.2% for predicting good motor outcome. Calibration assessment (Fig. 4) demonstrated good agreement between predicted probabilities and observed outcomes, with a calibration slope of 1.01, indicating excellent calibration.

Multivariate predictors of motor outcome

In the final multivariate logistic regression model (Table 3; Fig. 5), several predictors emerged as independently associated with good motor outcome. Higher ASPECTS score demonstrated the strongest positive association (OR 8.41, 95% CI 2.63–26.84, P < 0.001), followed by elevated BDNF Day 1 (OR 3.05, 95% CI 1.30–7.14, P = 0.010). Lower baseline NIHSS was strongly predictive of favorable outcome (OR 0.15, 95% CI 0.06–0.43, P < 0.001). Among qEEG parameters, lower DAR (OR 0.31, 95% CI 0.11–0.92, P = 0.036), lower DTABR (OR 0.44, 95% CI 0.21–0.92, P = 0.030), and lower BSI (OR 0.38, 95% CI 0.14-1.00, P = 0.050) were each independently associated with good motor outcome. Age showed a trend toward inverse association (OR 0.39, 95% CI 0.14–1.12, P = 0.081), while NGF Day 1 did not achieve statistical significance in the multivariate model (OR 1.82, 95% CI 0.81–4.12, P = 0.150).

Individual biomarker discriminative performance

The discriminative ability of individual biomarkers was assessed using separate ROC analyses (Fig. 3). For BDNF Day 1, a threshold of 12.96 ng/mL yielded AUC of 0.666 with sensitivity of 64.1% and specificity of 70.5%. Among qEEG parameters, DAR demonstrated superior individual discriminative ability (AUC = 0.767), with an optimal cut-off of 0.54 providing 67.2% sensitivity and 77.3% specificity. DTABR showed moderate discrimination (AUC = 0.719) with optimal threshold of 0.37. BSI achieved AUC of 0.767 with similar performance characteristics. These findings underscore that while individual biomarkers provide useful prognostic information, their combined assessment substantially enhances predictive accuracy compared to any single marker.

Sensitivity analysis incorporating baseline motor function

In a pre-specified sensitivity analysis, baseline FMA-UE score was incorporated into all four prediction models (Models 1–4 S). Assessed alone, baseline FMA-UE demonstrated moderate discriminative ability (cross-validated AUC = 0.728, 95% CI: 0.708–0.748; optimal threshold ≥ 31 points, sensitivity 57.8%, specificity 81.8%), comparable to individual neurotrophin biomarkers but lower than qEEG parameters (BSI: AUC = 0.812; DAR: AUC = 0.767). When baseline FMA-UE was added to each model, the cross-validated AUCs were 0.901 (95% CI: 0.877–0.924) for Model 1 S, 0.912 (95% CI: 0.887–0.937) for Model 2 S, 0.930 (95% CI: 0.910–0.951) for Model 3 S, and 0.935 (95% CI: 0.912–0.958) for Model 4 S (Supplementary Figure S1). The incremental pattern across models was preserved, indicating that serum neurotrophin profiles and qEEG parameters contributed prognostic information beyond baseline motor function.

Discussion

The present study demonstrates that integrating serum neurotrophin profiles with qEEG parameters significantly enhances prediction of 6-month motor outcomes following acute ischemic stroke. Our multimodal approach achieved excellent discriminative performance (cross-validated AUC = 0.930) with sensitivity of 92.2% and specificity of 93.2%, substantially surpassing conventional clinical predictors alone. These findings have important implications for individualized rehabilitation planning and prognostic counseling in clinical practice.

The prognostic value of serum BDNF in ischemic stroke aligns with its established roles in neuroplasticity and recovery [8, 9]. Our observation that higher admission BDNF levels predict favorable motor outcome is consistent with the landmark multicenter analysis by Zhu and colleagues, who reported decreased risks of poor 3-month outcomes in patients with elevated serum BDNF [10]. The moderate correlation between BDNF and motor recovery (ρ = 0.308) in our cohort falls within the range reported in prior investigations and supports its utility as a complementary prognostic marker [11, 12]. Importantly, the early dynamics of BDNF (ΔBDNF) distinguished outcome groups, with patients achieving motor recovery demonstrating a rising trajectory while those with poor outcomes exhibited declining levels. This temporal pattern may reflect differential activation of neuroplasticity mechanisms and warrants further investigation as a dynamic prognostic indicator.

Serum NGF provided additional prognostic information, particularly at Day 3, consistent with previous observations that higher NGF concentrations associate with favorable functional outcomes [15]. However, NGF did not achieve independent significance in multivariate analysis, suggesting that its prognostic contribution may be partially captured by other variables in the model. The differential behavior of BDNF and NGF likely reflects their distinct biological roles: while BDNF primarily mediates activity-dependent synaptic plasticity, NGF predominantly influences neuronal survival and inflammatory modulation [13, 14].

The qEEG parameters demonstrated robust associations with motor outcome, with DAR, DTABR, and BSI all achieving independent significance in multivariate analysis (Fig. 5). These findings corroborate the growing body of evidence supporting qEEG-derived indices as powerful prognostic tools in acute stroke [17–20]. The pathophysiological basis for these associations is well-established: excessive delta activity reflects cortical dysfunction and metabolic disturbance, while preserved alpha rhythms indicate intact thalamocortical circuitry essential for motor function [21]. The BSI captures interhemispheric dynamics relevant to motor recovery, as restoration of bilateral cortical coordination is critical for functional motor performance [22].

A notable strength of our approach lies in the complementary information provided by neurotrophin biomarkers and qEEG parameters. Neurotrophins reflect systemic neurotrophic support and neuroplasticity potential, while qEEG directly assesses cortical electrophysiological integrity. Their combination captures both humoral and neural substrates of recovery, explaining the incremental predictive gain observed. Similar multimodal strategies combining blood biomarkers with neuroimaging or neurophysiological measures have shown promise in recent prediction model studies [37, 38].

Several methodological considerations merit discussion. First, we focused on upper extremity motor outcome using the FMA-UE, a well-validated assessment tool with established psychometric properties and clinical relevance [28, 29]. The dichotomization threshold of 50 points aligns with functional capacity for activities of daily living and prior prognostic studies [32, 33]. The apparent clustering of exactly 30 patients (27.8% of the cohort) at an FMA-UE score of exactly 50 in Fig. 8 reflects the positioning of these patients precisely at the dichotomization boundary. We did not identify a directly comparable concentration at this exact score in previously published cohorts; the clustering may partly reflect a ceiling-related effect, whereby patients with mild-to-moderate initial deficits frequently recover to a just-functional upper-limb status, together with the ordinal nature of the FMA-UE scale. We explicitly acknowledge this clustering at the dichotomization boundary as a limitation (see below). Second, our sample size of 108 patients, while adequate for model development, limits the precision of AUC estimates and precludes robust external validation. The cross-validated AUC of 0.930 provides a more conservative estimate of expected performance in independent cohorts. Third, baseline motor function (FMA-UE Baseline) was deliberately excluded from the primary prediction models to evaluate the incremental value of biomarkers beyond readily available clinical assessments. A pre-specified sensitivity analysis that incorporated baseline FMA-UE into all models (see Methods and Results, Models 1–4 S, and Supplementary Figure S1) confirmed that serum neurotrophin profiles and qEEG parameters provide incremental prognostic value beyond both standard clinical variables and baseline motor function.

The clinical implications of our findings are several-fold. The combined model may inform early rehabilitation intensity decisions, with patients predicted to achieve good outcome potentially benefiting from aggressive early mobilization while those with poor predicted outcome may require modified expectations and longer-term support planning. The non-invasive nature and feasibility of the assessments—requiring only blood sampling and a brief EEG recording—support practical implementation. Furthermore, the identified biomarkers may serve as targets for therapeutic monitoring, with rising BDNF levels during rehabilitation potentially indicating favorable neuroplastic response [8].

Limitations of this study include its single-center design, which may limit generalizability to other populations. The observational nature precludes causal inference regarding biomarker-outcome relationships. We did not assess genetic polymorphisms such as the BDNF Val66Met variant that may modulate neurotrophin signaling and recovery [9]. Finally, while our model demonstrated excellent discrimination, future studies should evaluate decision curve analysis to quantify net clinical benefit and guide implementation thresholds. Furthermore, exactly 30 of the 108 patients (27.8%) scored precisely at the FMA-UE dichotomization threshold of 50 points. Because these patients lie at the boundary between good and poor outcome, modest measurement variation could reclassify them; this concentration at the cut-point may therefore have influenced the outcome classification and the downstream analyses. As we could not attribute the clustering to previously published score distributions, and it may partly reflect characteristics of our single recruiting centre, results in the vicinity of the threshold should be interpreted with appropriate caution.

Future directions should include external validation in independent multicenter cohorts, assessment of model performance across stroke subtypes and severity strata, and exploration of machine learning approaches that may capture non-linear biomarker interactions. Integration with neuroimaging biomarkers such as corticospinal tract integrity and functional connectivity may further refine prognostic precision. Longitudinal biomarker sampling throughout rehabilitation could enable dynamic outcome prediction and treatment response monitoring.

Conclusions

This prospective study demonstrates that integrating serum neurotrophin profiles (BDNF and NGF) with quantitative EEG parameters (DAR, DTABR, and BSI) provides incremental improvement in prediction of 6-month motor outcomes in acute ischemic stroke beyond conventional clinical variables. Although the 95% confidence intervals of model AUCs overlapped and differences between models did not reach conventional statistical significance, the combined model consistently demonstrated numerically superior discriminative performance across all 100 cross-validation repetitions. The multimodal prediction model achieved excellent discriminative performance with cross-validated AUC of 0.930 (95% CI: 0.908–0.952), sensitivity of 92.2%, and specificity of 93.2%. These findings support a precision medicine approach to stroke rehabilitation, wherein multimodal biomarker assessment facilitates individualized prognostication and therapeutic planning. Prospective validation and implementation studies are warranted to translate these findings into clinical practice.

Authors’ contributions

Yuyuan Lu: Conceptualization, Methodology, Formal analysis, Writing – original draft. Yuchen Guo: Investigation, Data curation, Validation, Writing – original draft. Songhua Ma: Conceptualization, Supervision, Project administration, Writing – review & editing.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. This research was funded by Key Research Projects of Higher Education in Jiangsu Province in 2025 (2025JGZD163), Nantong University Clinical Medicine Special Research Fund Project (2025LY037).

Data availability

The datasets generated during and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Institutional Ethics Committee of Nantong University (approval number: 2022-K089) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants or their legal representatives prior to enrollment.

Consent for publication

The authors declare that there are no conflicts of interest regarding this study.

Competing interests

The authors declare no competing interests.