Automatic bad channel detection in implantable brain-computer interfaces using multimodal features based on local field potentials and spike signals

"Bad channels" in implantable multi-channel recordings bring troubles into the precise quantitative description and analysis of neural signals, especially in the current "big data" era. In this paper, we combine multimodal features based on local field potentials (LFPs) and spike signals to detect bad channels automatically using machine learning. On the basis of 2632 pairs of LFPs and spike recordings acquired from five pigeons, 12 multimodal features are used to quantify each channel's temporal, frequency, phase and firing-rate properties. We implement seven classifiers in the detection tasks, in which the synthetic minority oversampling technique (SMOTE) system and Fisher weighted Euclidean distance sorting (FWEDS) are used to cope with the class imbalance problem. The results of the two-dimensional scatterplots and classifications demonstrate that correlation coefficient, phase locking value, and coherence have good discriminability. For the multimodal features, almost all the classifiers can obtain high accuracy and bad channel detection rate after the SMOTE operation, in which the Random Forests classifier shows relatively better comprehensive performance (accuracy: 0.9092 ± 0.0081, precision: 0.9123 ± 0.0100, and recall: 0.9057 ± 0.0121). The proposed approach can automatically detect bad channels based on multimodal features, and the results provide valuable references for larger datasets.