RippleNet: a Recurrent Neural Network for Sharp Wave Ripple (SPW-R) Detection

Abstract

Hippocampal sharp wave ripples (SPW-R) have been identified as key bio-markers of important brain functions such as memory consolidation and decision making. Understanding their underlying mechanisms in healthy and pathological brain function and behaviour rely on accurate SPW-R detection. In this multidisciplinary study, we propose a novel, self-improving artificial intelligence (AI) detection method in the form of deep Recurrent Neural Networks (RNN) with Long Short-Term memory (LSTM) layers that can learn features of SPW-R events from raw, labeled input data. The approach contrasts conventional routines that typically relies on hand-crafted, heuristic feature extraction and often laborious manual curation. The algorithm is trained using supervised learning on hand-curated data sets with SPW-R events obtained under controlled conditions. The input to the algorithm is the local field potential (LFP), the low-frequency part of extracellularly recorded electric potentials from the CA1 region of the hippocampus. Its output predictions can be interpreted as time-varying probabilities of SPW-R events for the duration of the inputs. A simple thresholding applied to the output probabilities is found to identify times of SPW-R events with high precision. The non-causal, or bidirectional variant of the proposed algorithm demonstrates consistently better accuracy compared to the causal, or unidirectional counterpart. Reference implementations of the algorithm, named 'RippleNet', are open source, freely available, and implemented using a common open-source framework for neural networks (tensorflow.keras) and can be easily incorporated into existing data analysis workflows for processing experimental data.

Keywords: Deep learning; Hippocampus CA1; Machine learning; Neuroscience; Recurrent neural networks; Sharp wave ripples (SPW-R).

Fig. 1

Example of a single detected SPW-R event and application of RippleNet to LFP data. a Raw LFP data with an SPW-R event. b Band-pass filtered LFP signal (150–250 Hz). c LFP spectrogram. d Illustration of a bidirectional RippleNet instance and its application to LFP signal X(t) (bottom trace) for predicting time-varying probabilities of SPW-R events $ŷ\left(t\right)$ (top trace). The subscript and superscripts annotating each layer denotes input and output dimensions respectively for an input sequence of length N

Fig. 2

Experimental setup. A set of recordings used for this study were acquired concomitant to two-photon microscopy from head-fixed mice. a Mice were prepared with a single electrode in the the hippocampal CA1 region and a contralateral reference electrode, chronic glass window for two-photon microscopy and a head bar for head fixation. b LFP recordings were recorded concomitant to two-photon microscopy in head-fixed mice

Fig. 3

Snapshots of experimental data. a Samples of raw LFP traces (X^〈1〉(t),X^〈2〉(t),…) with at least one labeled SPW-R event. The diamonds mark the times of the labeled events. Each column corresponds to samples j from the validation dataset. b Band-pass filtered LFP traces ${\varphi }_{\text{BP}}^{\langle j\rangle }\left(t\right)$. c Wavelet spectrograms S^〈j〉(t,f) computed from the LFP traces

Fig. 4

a Training and validation loss J as function of training epoch for for unidirectional RippleNet variants. Each instance M1-3 are instantiated using different random seeds. b same as panel (a), but for the non-causal bidirectional variant. c, d Training and validation MSE as function of training epoch

Fig. 5

Comparison of RippleNet predictions on samples from the validation set. Each column corresponds to different input LFP samples X^〈j〉 shown at the top. a) Input LFP samples X^〈j〉. The diamonds mark the times of labeled SPW-R events. b One-hot encoded label vectors y^〈j〉(t). c Predictions ${ŷ}^{\langle j\rangle }\left(t\right)$ made by the different instances of the unidirectional RippleNet variant. SPW-R events found by the peak-finding algorithm are marked with diamond markers. d Same as panel (c), but for the bidirectional RippleNet variant

Fig. 6

Effect of varying threshold and width parameters for the peak finding algorithm on counts of TP, FP and FN events in the validation dataset and derived metrics for different RippleNet instances. a Each row corresponds to different model instances of the unidirectional RippleNet variant. The columns correspond to different metrics. Colorbars are shared among panels in each column. The cross hatches in the last F₁ column correspond to parameter combinations maximizing the F₁ score as summarized in Table 5. b Same as panels in (a) but for instances of the bidirectional RippleNet variant

Fig. 7

Examples of validation samples j resulting in at least one FP prediction per sample. FN predictions may also occur. Columns show a input sequences with times of labeled SPW-R events denoted by diamond markers, b band-pass filtered LFP, c spectrograms and d predictions with detected events. The diamond and upward/downward pointing triangle markers denote times of TP, FP and FN events, respectively

Fig. 8

Same as Fig. 7, but showing a set of samples with at least one FN prediction per sample

Fig. 9

Application of RippleNet on continuous data. a 10 s excerpt of input LFP signal X(t) = ϕ(t). The diamonds marks the times of manually labeled SPW-R events. b band-pass filtered LFP ϕ_BP(t); c Time-frequency resolved spectrogram S(t,f) of the LFP. d label array y(t); e prediction $ŷ\left(t\right)$. The diamond and triangle markers represents the times of detected TP and FP SPW-R events using the threshold and width parameters that maximize the F₁ score for the model

Fig. 10

a Cumulative counts of predicted SPW-R events as function of labeled SPW-R events. b Cross-correlation coefficients between predicted ripple event times and labeled event times as function of time lag τ (2ms bin size). c Band-pass filtered LFP SPW-R event energy ($E_{{\varphi }_B{P}^{\langle j\rangle }}$) as function of $(1-{ŷ}^{\langle j\rangle })$ of SPW-R events (orange dots). The contour lines show the bivariate kernel density estimate of the kdeplot method in the Seaborn plotting library. The top and bottom panel shows labeled and predicted SPW-R events, respectively. d Averaged spectrograms for labeled (top) and predicted (bottom) SPW-R events

Fig. 11

Labeled events (input LFP, band-pass filtered signal, spectrograms) from the hidden test set with highest and lowest RippleNet confidence. Columns in rows 1–3 show eight events which maximized the RippleNet-predicted probabilities ($ŷ\left(t^{\langle j\rangle }\right)\approx 1$), while rows 4–6 shows labeled events with the lowest predicted event probability

Fig. 12

Same as Fig. 11 but for eight FP SPW-R events detected at or above threshold by the RippleNet algorithm