Machine-Learning Based EEG Biomarkers for Personalized Interventions

Details

This project aims to develop an artificial intelligence model to predict the response to a neuromodulation treatment (transcranial Direct Current Stimulation, tDCS) for neuropathic pain (NP) following spinal cord injury (SCI), based on electroencephalographic (EEG) signals and clinical assessments. The project consists of two stages: Stage 1 involves an open trial where participants with SCI and NP will receive neuromodulation treatment at our center, with data collected before and after treatment. Pre-Treatment Evaluation: * Clinical assessment through interviews and validated questionnaires targeted at factors associated with neuropathic pain, depression, and other relevant components. * EEG recording using a 64-channel device (Brain Products GmbH, Germany). EEG will be recorded in a soundproof room with participants in a resting state, first with eyes open for 5 minutes and then with eyes closed for another 5 minutes. Participants will be asked to avoid alcohol 12 hours prior and caffeine 3 hours before the recording. Neuromodulation Treatment: * The treatment protocol involves 10 sessions of non-invasive stimulation, each lasting 30 minutes. * tDCS will be administered using a battery-powered DC stimulator (Sooma tDCS, Helsinki, Finland) with 6 cm² saline-saturated circular electrodes. * The anode will be placed over C3 (EEG 10/20 system) to stimulate the primary motor cortex (M1) and the cathode over the contralateral supraorbital area (FP2). * For asymmetric pain, stimulation will be applied to the M1 contralateral to the more painful hemibody. For symmetric pain, the dominant hemisphere (C3) will be stimulated. * Maximum current delivered will be 2 mA (current density: 0.06 mA/cm²). * Sessions will be held once daily for two weeks (Monday to Friday), totaling 10 sessions. All stimulation parameters adhere to general safety guidelines for transcranial electrical stimulation . Post-Treatment Evaluation: • Conducted through interviews and the same validated questionnaires used in the pre-treatment assessment. As part of the intervention, participants will undergo EEG recording to study the brain's bioelectrical activity non-invasively. Active surface electrodes with electrode gel will be used to enhance skin conductivity. EEG recordings will be conducted at rest, with participants looking at a blank wall in a soundproof room, for 5 minutes with eyes open and 5 minutes with eyes closed. Stage 2 involves developing a predictive model to classify patients based on their response to the neuromodulation treatment. The model will use metrics derived from pre-treatment EEG recordings and clinical assessments conducted before and after the treatment, with the goal of predicting which patients will respond favorably to tDCS. EEG preprocessing will be performed by means of the Python programming language, using a custom-made preprocessing pipeline based on the MNE-Python library including: selective outlier channel and segment elimination, frequency filters, supervised auto-labeled independent component analysis for the elimination of muscular and ocular activity, and detection of bridged electrodes. The EEG recordings will be analyzed using metrics derived from the frequency, complexity and connectivity of the EEG signal. These metrics were selected due to their demonstrated potential in related publications, which highlight the capability of these features to capture differences between groups, either between treatment responders and non-responders, or between healthy subjects and those suffering from NP, among others. Based on these EEG features and other features derived from patient questionnaires, a feature selection process based on metric independence and relevance in previous literature will be carried out in order to maximize model generalizability. A machine learning (ML) model, with the main candidate model being a support vector machine (SVM), will be used in order to classify between responders and non-responders. The model will be validated by means of k-fold cross-validation. Given satisfactory results, an undersampling of EEG channels (adhering to typical 10:20 setups) will be used to evaluate whether an EEG with less electrodes can yield similar predictive results, thus reducing the need for EEG systems with a high electrode count.