A Computational Framework for Detecting Rhythmic Spiking Through the Power Spectra of Point Process Model ResidualsCox, Karin (2024) A Computational Framework for Detecting Rhythmic Spiking Through the Power Spectra of Point Process Model Residuals. Doctoral Dissertation, University of Pittsburgh. (Unpublished) This is the latest version of this item.
AbstractNeurons communicate through rapid, all-or-nothing action potential events (“spikes”). Researchers often predict that oscillatory drives will shape spike trains. For example, computational models of Parkinson’s Disease (PD) predict pathological 12-30 Hz spike rhythms. The detection of spike train oscillations presents an algorithmic challenge. We must devise an automated, scalable means of inferring when a noisy point process arose from a rate function that oscillates at a specific frequency, versus one that oscillates at a different frequency, or does not reliably oscillate. To identify oscillations, a naïve algorithm might compute the spike train’s power spectral density (PSD) – the distribution of signal power over frequencies – and detect oscillations as significant PSD peaks, contrasted against an assumed flat baseline. Yet non-oscillatory spike trains can exhibit aperiodic features that render this flat baseline inappropriate. This dissertation investigates whether two common baseline-distorting features can be removed through point process models (PPM), which predict instantaneous spike rates as a function of covariates. I first focus on the “recovery period” (RP): an inevitable, transient post-spike suppression in subsequent spiking. The RP creates global spectral distortion. An established “shuffling” method removes this distortion, but can also reduce the power associated with true spike rhythms. I developed an alternative “residuals” method that accounts for the RP-associated variance in the spike train with a PPM, and generates a corrected PSD from the PPM residuals. In some spike trains, a second, “burst-firing” feature can create further distortion. To accommodate bursts, I developed a “two-state” residuals method. This method infers the timing of burst- and non-burst states, and separately accounts for these states in the PPM. I compared the above methods’ ability to enable accurate oscillation detection with flat baseline-assuming tests. Over synthetic data, the residuals method improved upon the shuffling method’s detection accuracy, and the two-state variant offered further improvement when bursts were simulated. Moreover, in empirical data from a parkinsonian monkey, the residuals PSDs yielded increased incidence of the anticipated pathological oscillations. This work demonstrates that we can use PPMs to remove the distortion that aperiodic features introduce into power spectra, thereby improving the sensitivity and specificity of oscillation detection. Share
Details
Available Versions of this Item
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}