Population-Based Data

Invasive pneumococcal surveillance was conducted from 1996 to 1998 as part of the Active Bacterial Core Surveillance/ Emerging Infections Program Network (ABCs) using previously described methods (1). Briefly, project personnel communicated at least twice each month with contacts in all participating microbiology laboratories serving acute-care hospitals in San Francisco County, California; Connecticut; eight counties in Georgia (Cobb, Clayton, De Kalb, Douglas, Fulton, Gwinnett, Newton, and Rockdale) with 12 additional Atlanta-area counties starting in 1997; six counties in Maryland (Anne Arundel, Baltimore, Baltimore City, Carroll, Harford, and Howard); seven counties in Minnesota (Anoka, Carver, Dakota, Hennepin, Ramsey, Scott, and Washington); seven counties in New York starting in 1997 (Genesee, Livingston, Monroe, Ontario, Orleans, Wayne, and Yates); three counties in Oregon (Clackamas, Multnomah, and Washington); and five counties in Tennessee (Davidson, Hamilton, Knox, Shelby, and Williamson).

A case was defined as the isolation of Streptococcus pneumoniae from a normally sterile site (e.g., blood or cerebrospinal fluid) from a resident of a surveillance area. Periodic audits were conducted in each area. Any cases newly identified by audits were included in the surveillance database.

All isolates were sent to one of two centralized laboratories for susceptibility testing by broth microdilution, with a panel of drugs that included (in 1998) penicillin, amoxicillin, cefotaxime, cefuroxime, meropenem, erythromycin, clindamycin, chloramphenicol, vancomycin, rifampin, levofloxacin, trovafloxacin, and quinupristin-dalfopristin (Synercid7). Nonsusceptibility (resistance and intermediate susceptibility) was determined according to criteria of the National Committee for Clinical Laboratory Standards (13).

Ability of Sentinel Laboratory Groups To Estimate Proportion of Resistant Isolates

In each surveillance area for 1998, we generated all possible simple random samples of three, four, and five laboratories, excluding laboratories with <10 isolates. We limited our selection to up to five laboratories because a central objective of sentinel surveillance is to reduce required resources by reducing the number of facilities participating in the surveillance system. We refer to these simple random samples as sentinel groups of laboratories. We then calculated the percent of penicillin-nonsusceptible (MIC >0.1 µg/mL) pneumococci (%PNSP) among isolates in each of these sentinel groups and compared these percentages to the area’s actual %PNSP, as measured by ABCs. The %PNSP in sentinel groups was considered to be accurate if it was within 5 percentage points of the area’s actual %PNSP. We chose this interval because variation in the %PNSP within this range is unlikely to influence public health decisions (12).

We used a finite population correction based on the total number of isolates in each surveillance area to assess the number of randomly sampled isolates that would be needed to estimate an area’s actual %PNSP within 5 percentage points (14). We compared that number with the number of isolates in sentinel groups in each area.

Ability of Sentinel Groups To Track Changes in Prevalence of Drug-Resistant Pneumococci over Time

In each surveillance area, we subtracted the %PNSP in each possible group of five laboratories in 1996 from that measured for the group of five laboratories in 1998. We included only laboratories with >10 isolates in each of the 2-year periods. We then measured how often the change in %PNSP in sentinel groups was within 5 percentage points of the area’s actual change in %PNSP during the same time periods, based on ABCs data. We performed a similar analysis using the percentage of erythromycin-nonsusceptible (MIC >0.5 µg/mL) isolates as the outcome measure.

Ability of Sentinel Goups To Detect Emerging Fluoroquinolone Resistance

Using data from 1998, we measured the proportion of all possible groups of five sentinel laboratories within each surveillance area that captured any pneumococcal isolates with fluoroquinolone (levofloxacin or trovafloxacin) nonsusceptibility. We then compared that proportion with area-specific data on the presence of pneumococcal fluoroquinolone resistance from ABCs in 1998.

Evaluation of Hospital Predictors of %PNSP

We merged ABCs data from 1997 and 1998 with purchased data on hospital characteristics collected by the American Hospital Association (AHA) as part of the AHA Annual Survey of Registered American Hospitals in 1997. We categorized each hospital that matched between the two datasets into the following PNSP classes: >5 percentage points above the surveillance area proportion PNSP (high PNSP), <5 percentage points above or below the surveillance area PNSP (average PNSP), or >5 percentage points below the surveillance area PNSP (low PNSP). We used logistic regression to perform univariate analyses. We compared hospital characteristics in the high group with those in the average group, separately comparing hospital characteristics in the low group with those in the average group. We categorized continuous variables according to their quartiles or medians based on their distributions. We limited our analysis to hospital characteristics that might plausibly influence a hospital’s %PNSP based on findings of previous studies (15,16).

Population-Based Data

Figure

Figure. . Between-laboratory variation in percent penicillin-nonsusceptible isolates (%PNSP) and number of invasive pneumococcal isolates in selected surveillance areas. A) Connecticut; B) Selected counties of Maryland; C) Selected counties of Tennessee. Solid...

The %PNSP across surveillance areas in 1998 varied from 15 (California and New York) to 35 (Tennessee) (Table 1). The number of laboratories that isolated invasive pneumococci and the total number of invasive pneumococcal isolates also varied by surveillance area (Table 1). Consistent with previous observations (12), each surveillance area had striking variation across laboratories in the %PNSP in invasive pneumococcal isolates (Figure).

Ability of Sentinel Laboratory Groups To Estimate %PNSP

In New York, California, and Oregon (areas with a relatively small number of laboratories with >10 invasive pneumococcal isolates), sentinel groups of three, four, or five laboratories all did well at estimating the area’s actual %PNSP (Table 1). In the remaining areas, increasing the number of laboratories included in sentinel groups from three to five increased the probability that the sentinel %PNSP approached the area’s actual %PNSP. However, in Georgia and Tennessee, the two areas with the highest actual %PNSP, sentinel groups of five laboratories still poorly estimated the area’s actual percentage (Table 1).

In surveillance areas where most sentinel groups had an adequate sample size to estimate %PNSP accurately (i.e., the number of isolates met the sample size requirement), sentinel groups performed well compared with population-based surveillance (Table 2). In contrast, in Georgia and Tennessee, where sentinel groups performed poorly, a smaller proportion of sentinel groups met the minimum sample size requirements. However, in some states that failed to meet sample size requirements (e.g., Connecticut), sentinel groups performed well.

Ability of Sentinel Groups To Detect Trends in Prevalence of Nonsusceptible Pneumococci

The actual change in %PNSP in 1998 compared with that in 1996 varied across areas, ranging from Georgia’s 2% decline to Maryland’s 7% increase (Table 3). Because sentinel groups of five were the most accurate at predicting an area’s actual %PNSP, we focused strictly on groups of five for this analysis. Laboratories participating in ABCs in 1998 were often not the same as those participating in 1996 because of hospital or laboratory mergers, closing or opening of microbiology facilities in the surveillance areas, and expansion of areas under surveillance. Consequently, only a subset of all possible sentinel groups in 1998 matched those in 1996.

Over two thirds of each area’s sentinel groups of five accurately estimated changes in %PNSP, except in Tennessee, where only 45% correctly estimated a <5 percentage point change (Table 3). In the three areas with large changes in %PNSP (>3 percentage points), >90% of sentinel groups in each area predicted the direction of the change (increases in each case).

Trends in the proportion of isolates that were erythromycin nonsusceptible also varied by area, and three areas showed large increases from 1996 to 1998 (Table 3). Similar to trends observed for penicillin nonsusceptibility, sentinel groups had a high probability of detecting these increases in erythromycin nonsusceptibility (Table 3).

Ability of Sentinel Groups To Detect Emerging Fluoroquinolone Resistance

In 1998, seven isolates submitted to ABCs were nonsusceptible to levofloxacin; five of these were also nonsusceptible to trovafloxacin. The isolates came from seven different hospitals, located in five of the eight surveillance areas (California, Connecticut, Maryland, Minnesota, and Oregon). One of these hospitals, the only hospital from Oregon, had only five invasive pneumococcal isolates in 1998 and thus was excluded from our analysis of sentinel groups. Approximately 40% of sentinel groups of five laboratories in these areas (range 37% in Connecticut to 45% in Maryland) included a laboratory with a fluoroquinolone-nonsusceptible isolate, except in California, where there was only one possible sentinel group of five laboratories and this group included the fluoroquinolone-nonsusceptible isolate.

Evaluation of Hospital Predictors of %PNSP

The merged dataset of ABCs and AHA hospitals contained 104 hospitals: 24 (23%) were in the high PNSP category, 52 (50%) were in the average PNSP category; and 28 (27%) were in the low PNSP category. Hospitals that admitted only children (four hospitals that matched between the two datasets) were significantly more likely to be in the high PNSP group than in the average group (all four hospitals fell in the high category; Fisher’s exact test, p=0.008). Larger hospitals (measured by adjusted inpatient days, total beds, or total beds set up and staffed) were more likely to fall in the average category, but this trend was not consistent for all indicators capturing hospital size (Table 4). Additional variables tested by univariate analysis were not predictive of falling in the high or low category (Table 4). When we performed similar analyses using the percent of erythromycin-nonsusceptible isolates or of isolates with resistance to more than one drug class as the primary outcome measure, no additional predictors were identified.

In areas where sentinel surveillance did not accurately estimate the %PSNP (Georgia and Tennessee), can hospital predictors be used to improve performance? When we limited sentinel groups of five to the laboratories with the largest number of isolates, the range in %PNSP narrowed, but accuracy was not guaranteed (range in Georgia 29%-34%; range in Tennessee 36%-44%). Additionally, consistent with the analysis above, hand-picking sentinel hospitals to include those with a high proportion of pediatric isolates was likely to overestimate the actual %PNSP; in Georgia the children’s hospital had a %PNSP of 61%, whereas the area’s true %PNSP was 33% (Table 1).