[Paper Review] Spectral Processing of COVID-19 Time-Series Data

Susan L. McGovern|arXiv (Cornell University)|Aug 13, 2020

COVID-19 epidemiological studies15 references4 citations

TL;DR

This paper proposes advanced spectral processing techniques—using IIR filters and Fourier-based frequency-domain methods—to improve smoothing and isolate short-term oscillations in COVID-19 time-series data. It demonstrates that these methods outperform the widely used seven-day moving average by eliminating phase distortion and more effectively suppressing high-frequency noise, revealing persistent 8–21-day oscillations in U.S. mortality data.

ABSTRACT

The presence of oscillations in aggregated COVID-19 data not only raises questions about the data's accuracy, it hinders understanding of the pandemic. Spectral analysis is used to reveal additional properties of the data, and the oscillations are replicated using sinusoidal resynthesis. The precise behavior of the seven-day moving average is also discussed, specifically, the cause of its jaggedness and the phase error it introduces. In comparison, other filtering techniques and Fourier processing produce superior smoothing and have zero phase error. Both of these are presented, and they are extended to isolate several frequency ranges. This extracts some of the same short-term variability that is resynthesized, and it shows that fluctuations with periods between 8 and 21 days are present in U.S. mortality data. These methods have applications that include modeling epidemiological time-series data as well as identifying less obvious properties of the data.

Motivation & Objective

• To address the problematic oscillations in aggregated COVID-19 time-series data, which hinder accurate trend analysis and may stem from reporting inconsistencies.
• To investigate the limitations of the seven-day moving average, particularly its phase distortion and incomplete suppression of high-frequency noise.
• To develop and compare advanced spectral filtering techniques—elliptic IIR filters and Fourier-based frequency-domain processing—for improved data smoothing and oscillation isolation.
• To identify and analyze short-term periodic fluctuations (8–21 days) in U.S. mortality data that are obscured by conventional smoothing methods.
• To demonstrate that spectral resynthesis can accurately reproduce observed oscillations, including double-humped patterns, offering predictive insight into weekly data variability.

Proposed method

• Applied spectral analysis to the first differences of daily case and death counts to enhance visibility of short-period oscillations (e.g., 3.5, 2.33 days) and detect harmonics.
• Used Hanning windowed spectrograms over 193-day windows to visualize frequency content in the 0.1–0.475/day range, normalizing magnitudes for comparison.
• Implemented elliptic IIR filters and Fourier-based frequency-domain filtering with identical transfer functions to compare performance and ensure consistency.
• Designed low-pass, high-pass, and band-pass filters to selectively remove or isolate specific frequency bands, including 8–21-day oscillations.
• Employed sinusoidal resynthesis to reconstruct oscillations from spectral components, validating the method’s ability to reproduce observed data shapes.
• Validated results by comparing filter outputs across methods, confirming near-perfect agreement between IIR and Fourier-based approaches.

Experimental results

Research questions

• RQ1 What causes the prominent oscillations observed in daily COVID-19 case and death counts across multiple countries?
• RQ2 How does the seven-day moving average distort the phase of oscillations and fail to adequately suppress high-frequency noise?
• RQ3 Can spectral filtering techniques such as IIR and Fourier-based methods outperform the seven-day moving average in smoothing and preserving phase integrity?
• RQ4 What periodic fluctuations exist in U.S. mortality data that are not visible with standard smoothing techniques?
• RQ5 To what extent can spectral resynthesis accurately reconstruct the observed oscillatory behavior in time-series data?

Key findings

• The seven-day moving average introduces phase errors and fails to fully suppress high-frequency oscillations, leading to jagged, misleading trends.
• Spectral analysis revealed 7-day, 3.5-day, and 2.33-day oscillations in multiple countries, with the latter two being harmonics of the 7-day cycle.
• Elliptic IIR filters and Fourier-based frequency-domain processing produced nearly identical results, confirming methodological consistency and reliability.
• Band-pass filtering isolated oscillations with periods between 8 and 21 days in U.S. mortality data, indicating persistent short-term variability not removed by standard smoothing.
• Sinusoidal resynthesis successfully recreated the shape, minima, maxima, and double-humped patterns of the original data, validating spectral modeling accuracy.
• The time-series data contain significant aperiodic oscillations in low-case periods (e.g., early data), which are artifacts of spectral leakage but are mathematically recoverable.