Characterization of Clinically-Relevant Increases in Surgical Site Infection Rates Using Time-Frequency Decomposition Analysis

Presenter: QIFAN ZHAI

Research Category: Interdisciplinary Topics, Centers and Institutes
Student Type: Graduate
Additional Authors: Yang Yang, Iulian Ilie?, James Benneyan
PI: Qifan Zhai
Award Winner Category: Interdisciplinary Topics

Background: Surgical site infections (SSIs) are one of the most common and costly hospital-acquired conditions. Traditional epidemiological surveillance strategies have been shown to decrease SSI rates, but can be slow and difficult to interpret, resulting in delayed or incomplete outbreak detection. To address these deficiencies, various statistical quality monitoring methods adapted from other industries, such as statistical process control (SPC) charts, are increasingly used across healthcare. While effective, their performance is limited by the small but important distinction between statistical significance and clinical relevance.

Approach: We examined SSI rates after 13 commonly-performed surgical procedures across 58 community hospitals over 13 years. As part of a previous study, epidemiologists evaluated a set of 3,176 potential signals and rated their clinical relevance from 1 (no concern) to 4 (severe). We explored the feasibility of two signal processing methods, Fourier and wavelet decomposition, in identifying patterns distinguishing between clinically-relevant and -irrelevant signals. Examined variations included different window functions, respectively mother wavelets; analysis intervals of 3 to 24 months; and using raw or normalized SSI rates.

Results: Clinically-relevant signals were associated with increased spectral power density across all frequency bands and window sizes evaluated. Medium-term analysis windows (6-12 months preceding each signal) and low-frequency bands (period ≥ six months) yielded similar sensitivity and specificity levels (0.7-0.8) to optimized SPC charts. Multi-band classification using off-the-shelf algorithms largely increased predictive accuracy, with support vector machines performing best. These findings indicate that clinically-relevant signals are characterized by non-linear time-frequency patterns that could be exploited in infection surveillance.