Towards threshold invariance in defining hippocampal ripples

Abstract

Objective. Hippocampal ripples are transient neuronal features observed in high-frequency oscillatory bands of local field potentials (LFPs), and they occur primarily during periods of behavioral immobility and slow-wave sleep. Ripples have been defined based on mathematically engineered features, such as magnitudes, durations, and cycles per event. However, the ‘ripple’ could vary from laboratory to laboratory because their definition is subject to human bias, including the arbitrary choice of parameters and thresholds. In addition, LFPs are often influenced by myoelectric noise arising from animal movement, making it difficult to distinguish ripples from high-frequency noises. These problems have to be overcome. Approach. We extracted ripple candidates under few constraints and labeled them as binary or stochastic ‘true’ or ‘false’ ripples using Gaussian mixed model clustering and a deep convolutional neural network (CNN) in a weakly supervised fashion. Main results. Our automatic method separated ripples and myoelectric noise and was able to detect ripples even when the animals were moving. Moreover, we confirmed that a CNN detected ripples defined by our method. Leave-one-animal-out cross-validation estimated the accuracy, the area under the precision-recall curve, the receiver operating characteristic curve to be 0.88, 0.99 and 0.96, respectively. Significance. Our automatic ripple detection method will reduce time spent on performing laborious experiments and analyses.