Exploring the possible relationship between skin microbiome and brain cognitive functions: a pilot EEG study

Human microbiota, residing primarily in the gut and on the skin, influences host physiology via metabolites including short chain fatty acids (SCFAs). Gut-derived SCFAs participate in intestinal homeostasis, immune modulation, energy balance, and vitamin synthesis, and have been implicated in gut-brain axis (GBA) signaling that affects cognition and emotion. Dysbiosis of gut microbiota is associated with neurological and psychiatric conditions (e.g., ASD, depression, Parkinson’s disease) and altered SCFAs profiles. Skin microbiota similarly maintains skin homeostasis and produces SCFAs that can modulate local immunity and metabolism. Prior work shows skin bacteria (e.g., Staphylococcus epidermidis, Cutibacterium acnes) generate SCFAs that impact skin processes, yet it remains unknown whether skin-derived metabolites influence brain cognitive functions. This study tests the hypothesis that changes in the skin bacterial population are associated with alterations in EEG-based cognitive markers (ERPs N2 and P3) during an auditory oddball task, and employs machine learning to assess separability of EEG features under different skin microbiome manipulations.

Evidence for gut-brain interactions highlights SCFAs as mediators of bidirectional signaling (Mayer 2011; Dalile et al. 2019). Diet-driven changes in gut microbiota diversity have been linked to memory improvements in animal models (Li et al. 2009), while probiotics can modulate attention and emotional bias in humans. Gut dysbiosis correlates with altered SCFAs and microbiota composition in ASD and Parkinson’s disease. The skin microbiome contributes to skin health and immunity; dysbiosis is implicated in dermatological conditions. Skin commensals like Staphylococcus epidermidis can ferment to produce SCFAs that inhibit pathogens or modulate inflammation via receptors such as FFAR2/HDAC pathways, and Cutibacterium acnes-derived propionic acid can traverse skin and modulate melanogenesis without disrupting microbiome balance. ERP components N200 and P300 in oddball paradigms are robust markers of selective attention, novelty detection, cognitive control, and memory updating, with abnormalities linked to disorders such as ADHD. Despite parallels in SCFA production, the potential influence of skin microbiome on cognitive ERPs has not been reported prior to this pilot study.

Subjects and task: Twenty-four healthy participants aged 20–35 years were recruited; four were excluded for non-compliance, leaving n=20 (14 males, 6 females; mean age 25.5 ± 3.2 years). None had psychiatric/neurological disease or substance abuse. An auditory oddball paradigm presented two tones: odd (target) at 2000 Hz, infrequent (100 ± 10 trials; 20%), and standard at 1000 Hz, frequent (400 ± 40 trials; 80%). Participants mentally counted odd tones while ignoring standard tones. Each participant completed the task three times under different skin bacterial manipulations (alcohol, water, glycerol) in randomized order, with one-week intervals between sessions. Ethical approval: Procedures adhered to the Declaration of Helsinki and were approved by the Research Ethics Committee of China Medical University & Hospital (CMUH107-REC3-152). Written informed consent was obtained. Bacterial manipulation and measurement: Forehead skin was selected due to abundance of Cutibacterium acnes and Staphylococcus epidermidis. Alcohol manipulation: 75% alcohol applied 5 min before task to reduce bacteria. Water manipulation: double distilled water applied, wait 5 min to mimic natural growth. Glycerol manipulation: 65% glycerol applied; forehead covered with plastic wrap for 6 h to promote bacterial growth; pilot established 6 h as optimal. Skin samples from all experiments were incubated in 10 mL rich media (3 g/L TSB) with C12-14 alkyl benzoate, cetyl ethylhexanoate, and cetearyl isononanoate at 37°C, 200 rpm, 2 days, with 0.002% (w/v) phenol red as pH indicator. Color shift from red–orange to yellow indicated increased acidity due to fermentation. Bacterial population quantified via OD562 nm. EEG acquisition and processing: Fourteen-channel EEG (10–20 system: C3, C4, Cz, F3, F4, FC1, FC2, Fz, FP1, FP2, FCz, Pz, T3, T4) recorded at 250 Hz during tasks. Horizontal and vertical EOG recorded near right eye. Preprocessing: band-pass 0.1–30 Hz (5th-order Butterworth), epoch −500 to 1000 ms relative to stimulus onset (SPM12), automated EOG artifact correction via regression-based method (Schlogl et al. 2007). Trials separated by condition (standard vs odd). ERPs computed by averaging, baseline correction (−150 to −100 ms). N2 window 150–400 ms; P3 window 300–700 ms. Extracted ERP parameters: peak amplitudes and latencies for N2 and P3. Statistical analysis: OD and EEG features assessed for normality via Shapiro–Wilk test (MATLAB implementation). Repeated-measures ANOVAs (rmANOVA) on OD (factors: phase [pre/post], manipulation [alcohol/water/glycerol]) and on ERP peak amplitudes/latencies (factors: manipulation [alcohol/water/glycerol], condition [standard/odd]) across channels. Post hoc paired t-tests for pairwise comparisons; significance at p<0.05 with Tukey’s correction. Machine learning: Four classifiers (SVM, Naïve Bayes, decision tree, neural network) used for 2-class classification to separate datasets by manipulation pairs. Features per subject were limited to those showing at least one significant difference between experimental manipulations to reduce overfitting. Ten-fold cross-validation evaluated classification accuracy.

Bacterial population: rmANOVA showed a significant main effect of phase (pre vs post) on OD [F=39.273, p<0.0001]. Post hoc tests indicated significant pre-to-post differences within each manipulation: alcohol (t=9.483, p<0.0001), water (t=4.349, p=0.0003), glycerol (t=4.177, p=0.0005), and significant differences between post-manipulation conditions: glycerol vs alcohol (t=−9.483, p<0.0001), glycerol vs water (t=−2.527, p=0.021), water vs alcohol (t=−11.117, p<0.0001). Phenol red color changes visually confirmed greater acidity after water and glycerol compared to alcohol, consistent with fermentation. Oddball validation (water condition): rmANOVA confirmed a significant main effect of condition (oddball vs standard) on peak amplitude [F=80, p<0.0001]. Paired t-tests showed significantly higher P3 peaks for odd vs standard at Fz, FCz, Cz, and Pz, and higher N2 peaks at Fz and FCz; latencies of N2 and P3 differed significantly between odd and standard at Fz, FCz, Cz, and Pz. Experimental effects on ERPs: rmANOVA indicated significant main effects of manipulation on amplitudes at Cz [F=4.498, p=0.0361] and FCz [F=4.416, p=0.0378]. P3 amplitudes to oddball stimuli were significantly increased after alcohol (bacteria-reduced) compared to water and glycerol (bacteria-present) at fronto-central channels, while latencies did not show corresponding increases. Machine learning classification: EEG features distinguished manipulation pairs with high accuracy. Reported ten-fold CV accuracies (%): Alcohol vs Glycerol—SVM 95, NB 88.2, DT 85.3, NN 91.2; Alcohol vs Water—SVM 97.1, NB 96.5, DT 88.2, NN 97.3; Water vs Glycerol—SVM 96.5, NB 91.2, DT 89.2, NN 97.1. All classifiers achieved >88% accuracy across comparisons.

The classical oddball results under the water condition confirmed effective elicitation of N2 and P3 components associated with attention, validating the task. Removal of skin bacteria via alcohol was associated with enhanced frontal/central P3 amplitudes to oddball targets relative to water and glycerol, suggesting increased attentional engagement or allocation when bacterial load is reduced. The lack of significant P3 differences between water and glycerol despite higher OD and acidity under glycerol implies that increased bacterial quantity does not necessarily degrade cognitive ERP signals; glycerol may selectively promote commensal or probiotic-like bacteria and suppress pathogenic growth, and not all skin bacteria likely influence brain signals. Potential confounding from alcohol’s arousal or olfactory effects was mitigated by a 5-minute wait, and selective effects on odd trials were considered unlikely given standard EEG preparation practices. A mechanistic interpretation posits that dysbiosis-driven changes in local SCFAs may influence neural processes reflected in P3 amplitude, paralleling gut SCFAs’ known neuromodulatory roles in animal and human studies. Colorimetric acidity changes and OD increases support fermentation activity under water/glycerol, aligning with the hypothesis that microbiome state relates to cognitive ERP modulation. Multivariate ML classification further supports that EEG feature patterns differ across manipulations even when certain univariate comparisons (e.g., water vs glycerol P3) are not significant, underscoring the utility of multivariate approaches.

In healthy participants performing an auditory oddball task, reducing the skin bacterial population (alcohol manipulation) was associated with enhanced P3 amplitudes along midline/fronto-central channels compared to conditions supporting bacterial presence (water and glycerol). Machine learning classifiers reliably separated EEG datasets by manipulation with >88% accuracy, indicating that skin microbiome manipulations alter brain signal features related to attention. This pilot study provides initial evidence for a potential skin-brain axis and motivates future research to identify specific bacterial taxa or metabolites (e.g., SCFAs) that modulate cognitive functions, elucidate underlying mechanisms, and explore clinical applications for neurological and psychiatric conditions.

Bacterial taxa were not specifically identified; the study examined overall bacterial load without determining probiotic or pathogenic species. The skin microbiome is highly variable and influenced by external factors (environment, hygiene, individual differences), which may limit generalizability. The sample size (n=20) is relatively small given high inter-subject EEG variability, framing the work as a pilot. Machine learning classification may be susceptible to overfitting given feature count (16) and sample size; although cross-validation was used, larger datasets are needed for validation. Potential confounds from alcohol application (e.g., odor/arousal) cannot be completely excluded despite mitigation steps. The study infers roles of SCFAs from acidity and fermentation indicators but does not directly quantify SCFA concentrations.