A Comparison of Methods Examining Time-to-Readmission in the First Year of Life

Our cohort of births from the Pediatric Health Information System (PHIS) has been described in detail previously.⁶  We extracted birth discharge data and inpatient readmissions in the first year for singleton infants diagnosed with NOWS and those without NOWS or other major conditions who were born (or transferred on the day of birth) and discharged from the hospital from PHIS hospitals from January 1, 2016 to December 31, 2019.⁶  We identified newborns with NOWS by International Classification of Diseases, 10th Revision (ICD-10) diagnosis code (P96.1: neonatal withdrawal symptoms from maternal use of drugs of addiction) and infants without NOWS or other major conditions by All Patients Refined Diagnosis Related Groups (APR-DRGs) classification (APR-DRG 626: Neonate BWT 2000 – 2499 Grams, Normal Newborn or Neonate with Other Problem; and APR-DRG 640: Neonate BWT >2500 Grams, Normal Newborn or Neonate with Other Problem).⁶  Infants without NOWS or other major conditions were defined by ‘normal newborn’ APR-DRGs to construct a ‘healthy’ cohort of primarily routine births consisting of newborns with short birth stays and no major conditions at birth for comparison.⁶  As described in detail elsewhere and consistent with previous studies, we restricted our analysis to singleton infants and excluded newborns with additional comorbidities likely to contribute to readmission burden, including preterm or low birth weight, mechanical ventilation during birth hospitalization and congenital or genetic abnormalities.^6,11–13  Our analysis was restricted to hospitals with sufficient volumes of NOWS (≥15) and non-NOWS (≥150) birth discharges over the study period, similar to previous studies using PHIS.^6,11,13  Only inpatient readmissions to the PHIS hospital were included because PHIS does not capture readmissions to community or other hospitals.

The primary outcome was time-to-first-readmission within 1 year calculated relative to: (1) date of birth discharge and (2) date of birth. Figure 1 illustrates schematically the 2 methods for defining readmission timing for example newborns with and without NOWS.

Additional variables collected during birth hospitalization included LOS, NICU admission, birth weight, gestational age, race and ethnicity, sex, insurance payor, median household income based on ZIP code, and the presence of any complex chronic condition using flags provided by PHIS which are derived from ICD-10 diagnosis codes.¹⁴  Individual organ system chronic conditions (eg, cardiovascular, neurologic, respiratory, renal, metabolic, gastrointestinal, malignancy, hematologic, neonatal) were rare given our cohort was primarily otherwise healthy newborns; we, therefore, included the presence of any complex chronic condition rather than individual organ systems.

We classified the primary reason for readmission into major diagnoses using principal diagnosis codes and report the top reason for readmission at time points through the first year, by NOWS diagnosis and for each definition of readmission timing. Detailed substantive results from this exact cohort including comparisons between infants with versus without NOWS on sociodemographic and clinical factors as well as hospital use during readmissions, reasons for readmission, and adjusted association between NOWS diagnosis and readmissions have been reported fully in previous work.⁶ 

We report median (interquartile range; IQR) length of stay and compare by NOWS diagnosis using Wilcoxon rank test for continuous LOS and χ² test for discrete categories. We tested for differences in median time-to-readmission by NOWS diagnosis using Wilcoxon rank tests.

In our primary analysis, we compare crude rates and adjusted time-to-readmission by NOWS diagnosis for our 2 definitions of readmission timing. We fit Cox regression models predicting time-to-readmission: (1) from birth discharge; and (2) from birth, with left truncation allowing for delayed entry into the at-risk period at time of birth discharge.

Models accounted for hospital clustering and adjusted for birth sociodemographic factors (race and ethnicity, sex, insurance payor, median household income) and clinical factors (birth weight, gestational age, any complex chronic condition, NICU use) and birth length of stay. Missing values for sociodemographic factors (race and ethnicity, 8% missing; median household income, 1% missing) or clinical factors (birth weight, <0.1% missing; gestational age, 19% missing) were coded as “unknown” and retained in the sample and models. There were no missing data for any other predictors or outcomes. Given the skewed distribution of birth LOS as well as the high degree of correlation between NOWS diagnosis and birth LOS in our cohort, we applied a reciprocal transformation to birth LOS and included the transformed variable in adjusted models. A number of transformations were considered (eg, natural log, square root) and the reciprocal performed best with respect to visually assessing normality of the transformed variable. We assessed proportionality and time-varying effect of NOWS diagnosis, testing an interaction with log(time) in adjusted models. All analyses were performed in SAS (v9.4; Cary, NC) at an α-level of .05.

We included N = 155 885 birth discharges (n = 1467 with NOWS) from 17 hospitals. Newborns with NOWS had significantly longer birth LOS (median = 16.0 days, IQR = 15.0) compared to newborns without NOWS (median = 2.0 days, IQR = 1.0; P < .001). Ninety-eight percent without NOWS were discharged within 4 days, whereas 54% with NOWS stayed >14 days (P < .001).

Figure 1 illustrates schematically the 2 methods for defining readmission timing and the resulting differences in chronologic age at readmission. Newborn I without NOWS and newborn III with NOWS are both readmitted 7 days after birth discharge (Fig 1A). However, newborn I had a 3-day birth stay and is 10 days old at readmission, whereas newborn III stayed 21 days at birth and is, therefore, nearly 1 month old at readmission (Fig 1B).

Crude readmission rates by NOWS diagnosis and readmission timing definition are presented in Figure 2. The numbers of newborns at risk, along with crude readmission rates, and top reason for readmission by NOWS diagnosis and readmission timing definition for discrete time points through the first year are presented in Table 1. By 1 year since discharge or birth, 10% of infants with NOWS versus 6% without had been readmitted. Among newborns with versus without NOWS, median time-to-first-readmission was 76 (IQR = 119) vs 35 (IQR = 114) days since birth and 56 (IQR = 126) vs 33 (IQR = 115) days since discharge (both P < .001). Jaundice was the most common reason for readmission in the first week of life and first week since birth discharge for newborns without NOWS, whereas lower respiratory infections were the top reason for readmission for newborns with and without NOWS later into infancy (Table 1).

From birth discharge, readmission risk was highest in the first week for newborns without NOWS, whereas infants with NOWS were at higher risk during later postdischarge periods (Fig 2A). The time-varying NOWS effect resulted in a violation of proportional hazards (P < .001), an assumption required for Cox model validity. Survival curves and hazard functions crossed at 25 days and 14 days, respectively. To correct model specification because of violation of statistical assumptions, we constructed a piecewise model predicting early (<14 days) and late (≥14 days) readmissions by NOWS diagnosis. In adjusted models, infants with NOWS had significantly lower hazard of early readmissions (adjusted hazard ratio [aHR] = 0.56; 95% CI: 0.33–0.95) and significantly higher hazard of late readmissions through 1 year (aHR = 1.58; 95% CI: 1.20–2.08).

Defining readmission by day-of-life, risk was highest in the first 30-days since birth for infants without NOWS while infants with NOWS were at higher risk later into infancy (Fig 2B). Adjusted models indicated no violation of proportional hazards (P = .87). In adjusted models, infants with NOWS had a significantly higher hazard of 1-year readmissions (aHR = 1.56; 95% CI: 1.21–2.03).

Crude readmission risk differences through the first-year for infants with versus without NOWS vary by choice of index date (Fig 3). Methods differ substantially within the first 30 to 60 days, notably at 30 days in which estimated risk differences are inverted; ie, infants without NOWS are more likely to have been readmitted within 30 days since birth, whereas infants with NOWS have higher readmission rates at 30 days since discharge. Estimates from the 2 methods begin to converge after 60 days and are nearly identical by 1 year (Fig 3). The top reason for readmission also differs between methods, particularly at early time points (Table 1). Although jaundice was the top reason for readmission in the first week for newborns without NOWS regardless of index date, readmissions in the second week after birth discharge were most commonly for neonatal conditions, whereas other infections were most common in the second week of life (Table 1).

Our study highlights the importance of methodological decisions when studying newborn readmissions. Our 2 definitions of readmission timing lead to similar overall findings at 1 year with nearly identical adjusted 1-year hazards, but unsurprisingly there were important differences through early infancy driven by differences in chronologic age resulting in disparate risk profiles. We found differences between methods in both readmission rates and the top reason for readmission, especially at early time points which may be more likely to be unplanned readmissions attributable to quality of care during birth hospitalization.^1,2  This finding may have downstream clinical implications, particularly in identifying and targeting interventions for vulnerable age groups at highest risk for readmission or specific age-dependent causes.³  A different choice of index date may obscure critical chronologic age-based windows or nuances in reasons for readmission, which may lead to incorrect conclusions or ineffective interventions. A third approach to outcome definition might forego time-to-readmission in favor of a binary indicator and simpler logistic regression parameterization. However, readmission timing is a particularly important consideration for newborns, because even small differences in chronologic age are developmentally and clinically meaningful.^4,8 

Defining readmission by day-of-life is intuitive and required a less complex statistical model in our data, without issues of nonproportionality. Left-truncated Cox models are simple to implement in most statistical packages^15,16  and provide a reasonable compromise between complex time-varying parameterizations and oversimplification via logistic regression. Although a definition by date of discharge is commonly used to define readmissions, there are certain research questions for newborns in particular in which an alternative definition may be more appropriate. Indexing readmission timing by day-of-life is a suitable alternative for in-born or day-of-birth transfers. However, this method may not be appropriate for infants transferred to the hospital of interest later in the neonatal period.

Limitations of PHIS data are reported in detail elsewhere.⁶  Data from tertiary care children’s hospitals available in PHIS may not be representative of all delivery hospitals. Additionally, we restricted our analysis to hospitals with large numbers of uncomplicated term birth discharges with adjacent or affiliated birthing units and well-baby nurseries, which are not representative of all PHIS hospitals. Readmissions in PHIS are likely underestimated, because only readmissions to the same PHIS hospital are captured. We examined all-cause readmissions and did not distinguish between preventable versus unpreventable or expected readmissions, although the methods we applied are generalizable to all readmissions, regardless of preventability.^1,2  Our findings are limited to a single example comparing newborns with versus without a specific diagnosis. Because of the difference in age-dependent readmission risk profiles and reasons for readmission between newborns with and without NOWS, there is a potential for selection bias which may affect interpretation of our results. In this context, alternative methods may be appropriate to compare readmissions between infants with and without NOWS, eg, a nested matched case-control design or propensity score matching. We chose to compare newborns with versus without NOWS as a typical example comparing 2 relative extremes of birth LOS with very little overlap. Our findings are likely applicable to newborn readmissions more broadly, although more work is needed. For example, future studies might examine the measurement issue we describe among heterogeneous newborn cohorts and for other neonatal conditions. It is unknown how variation in average birth LOS between groups would influence model performance. Future simulation studies should compare model performance under a variety of conditions to determine the strength and direction of any bias.

Healthy newborns are at highest readmission risk shortly after birth, commonly for jaundice, whereas newborns with longer birth stays are likely not at risk for jaundice readmissions, because the physiologic window for this condition will have passed by the time of birth discharge, but are at higher risk later into infancy for readmissions because of other causes.^4–6,9,10  Therefore, studies comparing newborns with versus without comorbid conditions resulting in longer birth stays are likely to encounter similar methodologic issues related to differing risk profiles for age-dependent readmissions. Our findings emphasize the importance of thoughtful consideration when choosing readmission index date particularly for newborns, because results using different approaches may reach discordant conclusions in both readmission rates and reasons for readmission. These results highlight the measurement challenges of newborn readmissions and may have important methodologic implications for hospital benchmarking, national quality measures, and reimbursement penalties tied to readmissions for newborns as well as clinical implications in targeting age-based readmissions interventions.