Abstract
Administrative databases can be used to study outcomes including patients outside of clinical trials and have been used to identify relapse and HSCT in adult and adolescent/young adult leukemia populations. However, there are no published studies using validated billing and diagnostic codes to identify timing of relapse or HSCT in children with ALL. Published approaches are limited to relapses occurring after cessation of therapy, but a substantial proportion of pediatric ALL relapses occur on therapy. We hypothesized HSCT and early and late relapses could be detected accurately in a previously assembled cohort of children with ALL (Fisher 2014 Med Care), using pharmacy billing and ICD-9 diagnosis and procedure codes. We present our methods, validated at two large freestanding children's hospitals, and incidence estimates of relapse or HSCT as first events in a national cohort.
The Pediatric Health Information System (PHIS) cohort included patients aged 0-21 admitted between 1/1/2004 and 12/13/2013, previously identified with de novo ALL. We reviewed daily inpatient pharmacy, diagnosis, and procedure codes for patients in the PHIS ALL cohort from the Children's Hospital of Philadelphia (CHOP; 2004-2013) and Texas Children's Hospital (TCH; 2007-2013). Events were captured until the first of 5 years from diagnosis or last day of PHIS data. Relapses were identified using ICD-9 diagnosis/procedure codes and PHIS pharmacy codes (Figure 1A) correlating with relapse regimens. Manual review of daily PHIS data was performed for second-line chemotherapy at any time, reinduction-style chemotherapy365 days after diagnosis, or a relapsed ALL ICD-9 diagnosis code (204.02). HSCTs were identified using ICD-9 procedure and PHIS pharmacy code patterns consistent with conditioning (Figure 1B). We reviewed electronic medical records (EMR) for patients with do novo ALL from CHOP and TCH for all relapses and HSCTs as the gold standard. Demographics were evaluated by hospital and data source using chi-square tests. We calculated sensitivity and positive-predictive value (PPV) of PHIS-defined events compared to the EMR gold standard at the patient level and only considered the first relapse and HSCT per patient. PHIS events were considered valid if the date was within ±14 days of the EMR. We estimated 5-year incidences of relapse and HSCT as first events for the entire PHIS cohort, infants (<1 year at diagnosis), and high-risk ALL (receipt of daunorubicin in Induction).
Of 395 patients in the CHOP EMR cohort, 362 matched with the PHIS ALL cohort. The TCH EMR cohort had 410 patients, matching 329 from PHIS. Age, sex, and Down syndrome were similar (Table 1). CHOP patients were more likely to be Black, and race distribution within each hospital was similar by data source. Fewer CHOP patients were Hispanic, and more had missing ethnicity. Fewer TCH patients were missing ethnicity regardless of data source, though PHIS had a higher proportion of missing data. Proportions of children with high- and low-risk B-ALL, T-ALL, infant ALL, and Induction daunorubicin were similar. Government primary insurance in the first admission was more common at TCH.
At CHOP, 39 relapses were identified in PHIS, and 45 by EMR (sensitivity 85.7%, PPV 100%). At TCH, 30/31 relapses were correctly identified in PHIS (sensitivity 96.6%, PPV 100%). Our PHIS algorithm identified 38 CHOP patients who underwent HSCT during the study period and 34 at TCH. All matched the EMR, with 100% sensitivity and PPV for both hospitals.
Table 2 shows five-year incidences of relapse and HSCT in the entire PHIS ALL cohort (N=10,162), including relapse estimates adjusted for sensitivity. Relapses and HSCTs were higher in infants and in children receiving daunorubicin.
We present novel approaches to identify relapse and HSCT events using administrative data, validated at two children's hospitals. Timing of events are matched within ±14 days. Relapse estimates are slightly lower than clinical trial data, but this approach has higher sensitivity than published administrative data reports, and sensitivity-adjusted rates approximate clinical trial data. Detected events are likely to be true based on the 100% PPV. Our relapse identification approach is complex and requires disease-specific clinical expertise to identify relapse-style chemotherapy patterns in children on therapy; however, this approach can capture early relapses in children outside of clinical trials.
Disclosures
Fisher: Merck: Research Funding; Pfizer: Research Funding.
Scrybe: A Secure Audit Trail for Clinical Trial Data Fusion
