Integrative and interpretable framework to unveil the neurophysiological fingerprint of Alzheimer’s disease and mild cognitive impairment: A machine learning-SHAP approach

• M/EEG features allowed generating the neurophysiological fingerprints of AD and MCI. • Neurophysiological fingerprints summarise the most relevant alterations of AD and MCI. • Gradient Boosting obtained the highest three-class accuracy. • SHAP provides the interpretation of M/EEG parameters relevance to the ML model. • Our methodology shows a high level of replicability among M/EEG databases. Dementia and mild cognitive impairment (MCI) due to Alzheimer’s disease (AD) are neurological pathologies associated with disruptions in brain electromagnetic activity, typically studied using magnetoencephalography (MEG) and electroencephalography (EEG). To quantify diverse brain properties, different families of parameters can be computed from MEG and EEG (i.e., spectral, non-linear, morphological, functional connectivity, or network structure and organisation). However, studying these characteristics separately overlooks the complex nature of brain activity. Integrative frameworks can be useful to unveil the intricate neurophysiological fingerprint, as well as to characterise pathological conditions comprehensively. To that purpose, data fusion methodologies are crucial, despite their interpretational challenges. In this study, Machine Learning (ML) models were trained to discriminate between groups of severity, whereas the SHapley Additive eXplanations (SHAP) algorithm was afterwards utilised to assess the relevance of the input characteristics into the output classification. Three databases were analysed: MEG (55 healthy controls, HC, 42 MCI patients, and 86 AD patients), EEG1 (51 HC, 52 MCI, and 100 AD), and EEG2 (45 HC, 69 MCI, and 82 AD). The best results for the three-class classification problem were obtained by Gradient Boosting for the MEG database: 3-class Cohen’s kappa coefficient of 0.5452 and accuracy of 72.63 %. Afterwards, using SHAP on Gradient Boosting, it has been shown that spectral features were identified as highly relevant across all databases. Furthermore, morphology measures presented high relevance for the MEG database, whereas EEG1 and EEG2 databases showed functional connectivity and multiplex organisation measures, respectively, as relevant subgroups of parameters. Finally, commonly relevant features across databases were selected using SHAP to generate the neurophysiological fingerprints of AD and MCI. This study highlights the relevance of different MEG and EEG parameters in characterising neurological pathologies. The proposed framework, based on MEG and EEG, can be used to generate interpretable, robust, and accurate neurophysiological fingerprints of AD and MCI.