427. Retrospective Optimization of Statistical Process Control (SPC) Charts to Detect Clinically-Relevant Increases in Surgical Site Infection (SSI) Rates
Session: Poster Abstract Session: HAI: Epidemiologic Methods
Thursday, October 5, 2017
Room: Poster Hall CD
Posters
  • ID Week SPC Methods Poster_final.pdf (838.2 kB)
  • Background: SSIs are common and costly healthcare-associated infections with a wide range of negative patient outcomes. While using SPC charts to monitor SSI rates has become increasingly common, little is known about which specific chart types and design parameters maximize detection of clinically-relevant SSI rate increases with minimal false alarms.

    Methods: We retrospectively analyzed 12 years of data on 13 surgical procedures across 49 hospitals in the Duke Infection Control Outreach Network. Using 50 different SPC charts and designs, we identified potentially important SSI rate increases at individual units. Epidemiologists scored the clinical severity of 2711 representative signals. Results were used to identify the SPC approach that maximized accuracy (sum of sensitivity and specificity) among an extended set of 3600 chart variations and over 32 million combinations (Figure 1). An additional year of data was used to validate results (Figure 2).

    Results: The optimal SPC method (Figure 3) simultaneously employed two moving average charts with different design parameters (rolling baseline windows and lags) and baseline rates (network vs. hospital) to detect different types of important signals. The first chart better identified small sustained increases above network-wide baseline SSI rates, while the second better detected short-term increases above individual hospital baselines. This combination performed well on both original and validation datasets, with high sensitivity (0.89) and negative predictive value (0.93) as well as practical specificity (0.71) and positive predictive value (0.61). More common SPC variations such as Shewhart p-charts with 3 standard deviation (σ) control limits had much poorer sensitivity (< 0.35).

    Conclusion: This study is the largest empirical validation of SPC methods in healthcare to date, including data-driven chart optimization. Tighter control limits (1σ) were optimal given our use of SPC to screen potential true positives for review, rather than to identify only near-certain increases as traditionally done. Our solution currently is being tested in a large-scale multi-site randomized controlled trial to evaluate the performance of optimized SPC methods for prospective SSI surveillance.

     

    Iulian Ilies, PhD1, Deverick Anderson, MD, MPH, FSHEA, FIDSA2, Margo Jacobsen, BA3, Joseph Salem, BE3, Arthur W. Baker, MD, MPH2 and James Benneyan, PhD3, (1)Healthcare Systems Engineering Institute, Northeastern University, Boston, MA, (2)Duke Center for Antimicrobial Stewardship and Infection Prevention, Durham, NC, (3)Northeastern University Healthcare Systems Engineering Institute, Boston, MA

    Disclosures:

    I. Ilies, None

    D. Anderson, None

    M. Jacobsen, None

    J. Salem, None

    A. W. Baker, None

    J. Benneyan, None

    Findings in the abstracts are embargoed until 12:01 a.m. PDT, Wednesday Oct. 4th with the exception of research findings presented at the IDWeek press conferences.