Application of machine learning and complex network measures to an EEG dataset from ayahuasca experiments

Ayahuasca is a traditional Amazonian plant blend used ceremonially and medicinally and has been proposed as a candidate treatment for several neuropsychiatric conditions because it modulates brain regions involved in emotion, memory and executive function. Previous human studies reported acute changes in brain activity and blood perfusion in frontal, temporal and occipital regions, and some clinical trials and open‑label studies have suggested antidepressant effects. Complex network (graph theory) approaches and machine learning (ML) have both been applied to EEG data to capture disease‑related network changes and to build automatic classifiers, and recent interpretability tools such as SHapley Additive exPlanations (SHAP) facilitate identification of the most informative features for ML models. L. and colleagues set out to determine whether ML can automatically detect ayahuasca‑related changes in EEG and to compare three data abstraction levels for that task: (A) raw EEG time series, (B) Pearson correlation connectivity matrices between electrodes, and (C) complex network measures derived from those connectivity matrices. They also introduced new community‑based network measures (average path length within the largest community found by several community detection algorithms) and used SHAP values to identify the most important electrodes, connections and network measures for classification. The goal was both to identify which abstraction level best captures ayahuasca effects and to offer interpretable biomarkers that relate to known neurophysiological mechanisms.

The study used an open EEG dataset made available by the Federal University of São Carlos. The extracted text contains an inconsistency about sample size: the Methods first state "Sixteen healthy male and female patients" but subsequently list "eight women" and "12 men", which sum to 20; the extracted text does not clearly report the exact sample size. Participants had prior ayahuasca experience and underwent screening exclusions (under 21 years, personal psychiatric history, current psychiatric medication, cardiovascular disease, and recent neurological disorders or brain damage). Recordings began 25 minutes before ingestion and continued for 200 minutes after ingestion (total duration 225 minutes). Participants were instructed to rest with eyes closed and were monitored by nursing staff. EEG was acquired with 62 electrodes (10–10 system), downsampled to 500 Hz, bandpass filtered 0.5–150 Hz, and movement artifacts removed. The ayahuasca brew chemical composition is listed in the extract. The analysis framed a two‑class classification problem: class 0 (control) comprised the first time window (25 minutes pre‑ingestion to 50 minutes post‑ingestion, when blood DMT concentration is low) and class 1 comprised two subsequent post‑ingestion windows (both treated as influenced by ayahuasca). Splitting the post‑ingestion recording into two windows increased available data points for ML, while independent subject count remained unchanged. Three input data abstraction levels were prepared: (A) raw EEG time series matrices (subjects × time points × electrodes) for each window; (B) Pearson correlation connectivity matrices computed for each time window and participant, flattened to vectors of electrode‑pair correlations; and (C) complex network measures extracted from graphs built on each connectivity matrix. Computed network measures included assortativity, average path length, betweenness centrality, closeness centrality (CC), eigenvector centrality, diameter, hub score, average degree of nearest neighbours, mean degree, second moment degree, degree entropy and several newly developed metrics combining community detection with average path length. Community detection algorithms used were leading eigenvector, label propagation, edge betweenness, spinglass and multilevel; their largest‑community average path lengths were denoted with an "A" prefix (for example, ALC for the leading eigenvector average path length). For classification the authors primarily used a support vector machine (SVM) and compared it (results reported in appendices not included here) with Random Forest, Naive Bayes, multilayer perceptron, stochastic gradient descent, logistic regression and XGBoost; selected deep‑learning variations were also tested and reported separately. Data were split into training and test sets with 25% held out for testing, and performance estimates used 10‑fold cross‑validation and grid search hyperparameter tuning. Evaluation metrics included accuracy, precision, recall (sensitivity), F1 score and area under the receiver operating characteristic curve (AUC), using micro and macro averaging as appropriate. Model interpretability relied on SHAP values to rank electrodes, electrode‑to‑electrode connections and complex network features by importance.

Across the three abstraction levels the classifier achieved above‑chance performance for detecting ayahuasca influence. The top‑line reported accuracies were: connectivity matrices (level B) 92%, raw EEG time series (A) 88%, and complex network measures (C) 83%. For the raw EEG time series input the test‑sample performance reported was mean AUC 0.85, mean precision 0.88, mean F1 score 0.86, mean recall 0.85 and mean accuracy 0.86 (the paper also reports an overall 88% figure for this level elsewhere). SHAP analysis identified a subset of electrodes as most informative, with T7 (temporal region) ranked highest and other electrodes such as FC1, Fp1 and P5 appearing among the next most important; the extracted text truncates the full ranking. For connectivity matrices (Pearson correlation between all electrode pairs) the classifier performed best: mean AUC 0.88, mean accuracy 0.92, mean F1 score 0.90, mean recall 0.88 and mean precision 0.94. SHAP values highlighted the connection between left frontal electrode F3 and right parietal‑occipital electrode PO4 as the most important single connection for classification. For complex network measures the test set performance was lower: mean AUC 0.75, mean accuracy 0.83, mean F1 score 0.78, mean recall 0.75 and mean precision 0.90. SHAP identified closeness centrality (CC) as the most important network measure, followed by assortativity and one of the newly defined community‑based metrics (ALC, the leading eigenvector largest‑community average path length). The authors report that median CC increased with ayahuasca, whereas median assortativity decreased (though the upper confidence interval for assortativity increased). The new community measures ranked among the top twenty features for classification, and ALC in particular was third in feature importance, showing larger values after ayahuasca consistent with larger communities and longer path lengths within those communities. The authors further note that single features (for example CC or assortativity alone) showed limited separation between control and ayahuasca windows, but combining multiple features enabled successful classification, supporting the idea that a multivariate feature set serves better as a biomarker than any lone metric.

The authors interpret their results as evidence that ML can automatically detect acute ayahuasca‑related changes in brain activity from EEG data, and that interregional connectivity changes (captured by Pearson correlation matrices) are more informative for classification than raw time series or summary network measures. They highlight practical advantages of using connectivity matrices: higher accuracy, reduced data storage and faster ML training, which may be relevant for larger datasets and clinical systems. At the electrode level, raw EEG analyses emphasised frontal and temporal lobes as most affected, consistent with prior SPECT and fMRI findings that implicate frontal, temporal and occipital regions in ayahuasca effects and relate to introspection, emotional processing and cognitive functions. At the connectivity level the F3–PO4 connection (left frontal to right parietal‑occipital) emerged as the top feature; the authors link this to previous reports of gamma synchrony between parietal‑occipital and frontal cortices during face recognition tasks and suggest it may reflect cognitive processes akin to face recognition during ayahuasca‑mediated visual hallucinations. Among network metrics, closeness centrality and assortativity were most informative. The authors note parallels with Alzheimer’s disease (AD) literature—CC typically decreases in AD but increased here with ayahuasca, while assortativity has been reported to increase in AD whereas it decreased (median) under ayahuasca in their data—prompting the authors to suggest a potential therapeutic mechanism. They refer to a proposed biochemical mechanism whereby DMT agonises the sigma‑1 receptor (Sig‑1R), modulating endoplasmic reticulum stress and the unfolded protein response, processes implicated in neuropsychiatric and neurodegenerative conditions; the authors present this as a possible link to therapeutic effects but do not claim causal proof. The new community‑based metrics (for example ALC) ranked highly, and their increase after ayahuasca suggests larger communities with longer internal path lengths. The authors describe this as a disruption in the balance between functional segregation and integration—larger communities implying slower information distribution across the functional network. They also caution that single features were insufficient as biomarkers and that the discriminative signal arose from the combination of multiple measures. Limitations and future directions discussed by the authors include that the analysis focused on acute effects; longitudinal or long‑term impacts on functional connectivity remain to be studied. They also note that their workflow could be applied to EEG data from other psychedelics (for example pure DMT) and to in vitro neuronal network recordings, but further work is required to generalise findings. The extracted text contains an inconsistency concerning sample size which the authors do not explicitly resolve in the provided excerpts.

L. and colleagues conclude that machine learning applied to EEG can automatically detect acute ayahuasca‑induced changes in brain activity and that Pearson correlation‑based connectivity matrices constitute the most effective data abstraction level among those tested. The classifier identified frontal and temporal regions as prominently affected, the F3–PO4 connection as the single most important electrode‑to‑electrode link, and closeness centrality, assortativity and a newly defined community‑based metric (ALC) as the most informative network measures. The authors interpret larger community sizes and increased path lengths after ayahuasca as indicating slower information distribution and increased network entropy, while noting that psychedelics also produce strong, long‑range functional connections not present in the normal state. They present the computational workflow as a robust, interpretable tool for acute evaluation of psychedelic effects and propose applying it to other datasets and experimental systems in future work.