Prognostic Value of Parenteral Nutrition Duration on Risk of Retinopathy of Prematurity

Key Points Question Does parenteral nutrition duration improve the sensitivity and maintain high specificity of DIGIROP models in predicting retinopathy of prematurity (ROP) treatment? Findings In this prognostic study, among 11 139 ROP-screened Swedish infants with 14 or more vs less than 14 days of parenteral nutrition, 64.0% vs 18.5%, respectively, had any ROP and 18.1% vs 1.6% had ROP treatment. DIGIROP 2.0 models were updated to include all ROP-screened infants regardless of gestational age and presented a 100% sensitivity, high specificity, and superiority to WINROP and G-ROP models. Meaning The validated DIGIROP 2.0 decision support tool is suggested to be an efficient individual prediction tool for safe release of infants from unnecessary ROP screening examinations.

R etinopathy of prematurity (ROP) is a multifactorial eye disease and a major cause of visual impairment in children. 1 Worldwide, ROP screening examinations detect and monitor the disease until it regresses or progresses to severe ROP needing treatment. 2 The most prominent risk factors for ROP are low gestational age (GA), low birth weight (BW), low early serum insulinlike growth factor-1 (IGF-1), poor early weight gain, fluctuating oxygen concentrations, infections, and comorbidities. 3 More parenteral nutrition and less human milk have also been identified as risk factors. 4 Enteral nutrition, particularly with mother's milk shortly after birth, promotes intestinal development and stimulates the cultivation of a healthier gut microbiome that is associated with lower risk of ROP. 5,6 Likewise, early attainment of full enteral nutrition is related to lower ROP risk. [7][8][9] Although life-saving for many infants, longer exposure to and higher volume of parenteral nutrition increase the risk of infections and reduce nutrient absorption in the premature baby. 10 Developed in Sweden, the Weight, IGF-1, Neonatal, and ROP (WINROP) model was, to our knowledge, the first ROP prediction model proposed to identify high-risk and low-risk infants. 11,12 It was simplified to include only GA, sex, and weekly weight gain. The Postnatal Growth and ROP (G-ROP) model, developed on approximately 7500 infants from the US and Canada, includes GA, BW, hydrocephalus, and weight gain for days 10 to 19, 20 to 29, and 30 to 39. 13,14 Further, we developed and validated 2 prediction models for ROP treatment based on approximately 7000 Swedish infants born at 24 to 30 weeks' GA. [15][16][17][18][19] The Digital ROP (DIGIROP) birth model includes GA, sex, and standardized BW. Additionally, the timing for the first ROP diagnosis is included in DIGIROP screen model. Both models were developed to require 100% sensitivity. The DIGIROP birth model showed specificity of approximately 50% and the DIGIROP screen model up to approximately 80% during screening. In a contemporary Swedish cohort, approximately 50% specificity at birth was maintained, but 4 infants with severe comorbidities of 57 with ROP treatment were identified as not needing ROP screening. 19 Inclusion of a clinical variable representing infants' comorbidity was warranted.
Therefore, we evaluated the prognostic value of parenteral nutrition duration (PND) on ROP in this study. Furthermore, DIGIROP prediction models for ROP treatment and their clinical decision support tool were updated to include all ROP-screened infants regardless of GA and to incorporate an early cutoff for PND as well as to perform internal and external validation. Additionally, the DIGIROP outcomes were compared with those of WINROP and G-ROP.

Study Population
The study population included infants born from 2007 to 2020 and registered in SWEDROP (N = 11 178). 23 SWEDROP collects information from prematurely born infants examined for ROP either routinely (ie, initially with a GA of less than 32 weeks, from 2012 with a GA of less than 31 weeks, and from 2020 with a GA of less than 30 weeks) or by indication. 24,25 Using unique personal identification numbers, SWEDROP was linked to Swedish Neonatal Quality Register to obtain PND. 26 Any mismatch of infants' GA between the 2 registers as well as any other missing or questioned data were checked in the medical records.
A total of 39 infants (0.3%) with missing BW data were excluded. In total, 11 139 infants were included ( Figure 1).

WINROP and G-ROP Validation Cohort
For 249 infants born from July 1, 2017, to December 31, 2020, who were routinely screened and/or treated at the Queen Silvia Children's Hospital in Gothenburg, Sweden, weekly weights were obtained from medical records to validate WINROP and G-ROP models and compare their predictive ability with DIGIROP models.

Key Points
Question Does parenteral nutrition duration improve the sensitivity and maintain high specificity of DIGIROP models in predicting retinopathy of prematurity (ROP) treatment?
Findings In this prognostic study, among 11 139 ROP-screened Swedish infants with 14 or more vs less than 14 days of parenteral nutrition, 64.0% vs 18.5%, respectively, had any ROP and 18.1% vs 1.6% had ROP treatment. DIGIROP 2.0 models were updated to include all ROP-screened infants regardless of gestational age and presented a 100% sensitivity, high specificity, and superiority to WINROP and G-ROP models.
Meaning The validated DIGIROP 2.0 decision support tool is suggested to be an efficient individual prediction tool for safe release of infants from unnecessary ROP screening examinations.

Study Procedures
The postnatal age (PNA), postmenstrual age, and GA (by fetal ultrasonography) were defined per the American Academy of Pediatrics policy. 27 BW SD scores (BWSDS) were calculated in infants with a GA of 24 weeks or more using the Swedish reference of approximately 800 000 singletons born from 1990 to 1999. 28

Study Outcomes and Predictors
The outcomes related to PND were any ROP, defined by the International Classification of ROP, and ROP treatment, as per the Early Treatment for ROP criteria, or based on the examining ophthalmologist's assessment. 29,30 The outcome for prediction models was ROP treatment.
PND reflects the number of days with parenteral protein and lipid supplementation. According to national and European guidelines, parenteral nutrition is initiated as early as possible after birth and is gradually increased during the following 3 to 4 days. 31,32 Enterally, infants in Sweden receive mother's own milk from day 1, if available, or otherwise pasteurized donor milk, increasing to an enteral target volume of 160 to 180 mL/kg per day depending on the infant's feeding tolerance. Healthy infants are expected to reach this target during the first 2 weeks postnatally.
Predictors used for development of the DIGIROP 2.0 prescreen model were GA, sex, BW, PND (less than 14 days, 14 days or more, or unknown), and important interactions. The DIGIROP 2.0 screen model included, similar to the original publication, the log-odds of the DIGIROP 2.0 prescreen model risk estimates (that includes PND), the age and presence or not of first detection of ROP at screening occasion, and important interactions.

Statistical Analysis
Descriptively, continuous variables were presented as means and SDs or medians and ranges, and categorical variables were presented as counts and percentages. Between-groups Fisher exact tests were used for dichotomous variables, Mantel-Haenszel χ 2 trend tests for ordered categorical variables, and Mann-Whitney U tests for continuous variables. Spearman correlation was used to study correlations between ROP severity and PND.
To identify an early cutoff of PND, receiver operating characteristic analysis was performed. The cutoffs investigated were at 7 to 28 days of PND, which were considered as meaningful for an early prediction of ROP treatment. The selected cutoff at 14 days had the highest area under the receiver operating characteristic curve (AUC) and maximized sensitivity and specificity (Youden index). The associations between PND and ROP were studied using logistic regression adjusting for GA, BW, and sex. Odds ratios (ORs), adjusted ORs (aORs), and 95% CI were calculated. In addition, risk differences were described in crude absolute terms using Meittinen-Nurminen 95% confidence limits.
The DIGIROP 2.0 prescreen model was developed including all ROP-screened infants using extended Poisson regression. 17,[33][34][35] The first model included variables from the DIGIROP 1.0 birth model, which was extended by including categorized PND and important interactions, and therefore renamed to the DIGIROP 2.0 prescreen model. The selected model had the lowest Akaike information criterion value. The parameter estimate, standard error, hazard ratio with 95% CIs, and P value were presented. The estimated probability for ROP treatment was calculated as 1 − survival probability. Survival probability was obtained by exp(−H[t]) and H(t) by numerical integration of the hazard function for 20 follow-up weeks.
The DIGIROP 2.0 screen model was developed including all ROP-screened infants using logistic regression models for PNA from 6 to 14 weeks. The same variables as those included in the original publication were used. 18 The models' predictive ability was described by sensitivity, specificity, cumulative specificity, positive predictive value, negative predictive value, accuracy, and AUC. For the DIGIROP screen model, the specificity for each week from weeks 6 to 14 was based on the number of infants discharged from ROP examinations that week or previously and was termed cumulative specificity. Calibration plots and Hosmer-Lemeshow test were performed to evaluate observed vs estimated probabilities. Internal validation of the models was performed using 10-fold cross-validation. External validation was performed on a temporally different Swedish cohort to evaluate the models' transportability in time. The model's sensitivity and specificity were compared with the WINROP model (2006 to 2009) and G-ROP model (2018 to 2020) in a subset of infants from the temporal validation cohort. 12,14 Superiority was evaluated by first comparing sensitivity, requiring achievement of 100%. Then, the Sign test was used to demonstrate superiority of one method over the other considering specificity. To obtain weights for postnatal days 10, 19, 20, 29, 30, and 39 in the G-ROP model, linear interpolation was applied. In case of missing data, the infant was deemed to need screening for both WINROP and G-ROP.
All tests were 2-sided. The significance level was P < .05. No adjustment for multiple comparisons was made. Only positive associations between PND and ROP were to be demon- strated. All analyses were performed using SAS software version 9.4 (SAS Institute) and R version 4.2.0 (The R Foundation).

PND and ROP
Among the whole cohort, infants received a mean (SD) of 10.8 (16.7) days of PND. A total of 7228 infants (64.9%) received PND for less than 14 days, 2308 (20.7%) received PND for 14 days or more, and 1603 (14.4%) had unknown PND ( Table 1). The proportion of infants with 14 days or more days of PND and number of days on parenteral nutrition increased gradually with ROP severity (Spearman r = 0.45; P < .001) (Figure 2A).
Compared with those with PND less than 14 days, infants with PND for 14 1). Overall and GA-stratified data are presented in Figure 2B and C and eFigure 1 in Supplement 1.
Disease progression from the first detection of ROP to first ROP treatment was faster in infants receiving PND for 14 days or more compared with less than 14 days (median  Table 1).

Update of the DIGIROP 2.0 Prescreen Prediction Model for ROP Treatment Including PND
In the DIGIROP 2.0 prescreen model, BWSDS was replaced by BW, including all ROP-screened infants. Categorized PND (less than 14 days, 14 days or more, and unknown) was added to capture high-risk infants who cannot be captured given only GA, BW, and sex. The interaction between sex and PND was significant (eFigure 2 in Supplement 1), showing more or similar ROP treatment received by girls compared with boys among infants who received 14 days or more of PND, as opposed to those who received less than 14 days of PND, where girls needed ROP treatment less than boys. The final model is presented in eTable 3 in Supplement 1 and the estimated probabilities in eFigure 3 in Supplement 1. The model was well calibrated (Hosmer-Lemeshow test, P = .57) (eTable 3 in Supplement 1), and the calibration plot of observed vs estimated probabilities well distributed around the diagonal (eFigure 4 in Supplement 1). The AUC was 0.93. Given the required 100% (95% CI, 99.2-100) sensitivity, the specificity was 48.5% (95% CI, 47.4-49.5) (Figure 3). The percentage of infants discharged from screening by GA is given in eFigure 5A in Supplement 1.

Internal and External Validation of the DIGIROP 2.0 Prescreen Prediction Model for ROP Treatment Including PND
Internal validation using cross-validation showed a specificity of 47.4% (eFigure 6 in Supplement 1). The obtained sensitivity on the temporally different Swedish validation cohort was 100% (95% CI, 97.6-100) and the specificity was 39.4% (95% CI, 37.3-41.5) ( Figure 3A). The lower specificity in the validation cohort was secondary to lower GA due to increased survival of more immature infants and fewer infants with a GA of 30 weeks (93 of   The temporal validation cohort included Swedish National Registry for ROP data collected from July 1, 2017,   Figure 3A). The DIGIROP models and G-ROP criteria were superior to WINROP considering sensitivity, and for specificity, DIGIROP models were superior to G-ROP criteria ( Figure 3B).

Discussion
Including all Swedish ROP-screened infants regardless of GA in the SWEDROP from 2007 to 2020, we showed a strong prognostic value of days of parenteral nutrition on any ROP and ROP needing treatment. After adjustment, infants with 14 days or more of PND had 84% higher odds of any ROP and 120% higher odds of ROP treatment than those with less than 14 days PND. Including GA, BW, sex, PND, and status and age at the first ROP diagnosis, DIGIROP 2.0 prediction models were updated and validated into a safe (sensitivity,  100%; 95% CI, 99-100) clinical decision support tool with a specificity of 47% (95% CI, 46-48) to 77% (95% CI, 76-78) that could be applied by physicians using the online application. 15 Vanhaesebrouck and colleagues 36 showed in 2008 among 412 infants (26% ROP treatment) that PND, GA, BW, and length of oxygen were independent predictors for any ROP. Niwald et al 37 showed that PND for more than 10 days was a predictor for ROP needing treatment among 118 infants. Petrachkova et al 38 developed a prognostic model for type 1 ROP (69 infants; 42% with type 1 ROP), with PND for more than 13 days being one of the significant predictors. Interestingly, the same cutoff for PND was selected as that in our cohort including more than 11 000 infants. We investigated days 7 to 28 to enable early prognosis; a 14-day cutoff had the highest predictive ability. Related to this, Porcelli and Weaver Jr 4 found that volume of parenteral nutrition during week 2, but not week 1, was related to severe ROP outcome.
PND was greater in girls than in boys, especially for lower GA, where ROP needing treatment is more frequent. Infants with longer PND had faster progression of the disease independently of GA. Further investigation of the mechanism behind early sex-specific effects of parenteral nutrition and its relation to intestinal and neurovascular development as well as accelerated progression of severe ROP is needed.

Strengths and Limitations
This study's strength is that it comprises the ROP-screened Swedish population from 2007 to 2020, excluding only 39 of 11 178 infants with missing data. However, this study has limitations. Register data may include potential editing errors. Additionally, due to the Swedish population being homogenous in terms of ethnicity, neonatal care, and socioeconomic status, the DIGIROP models need to be thoroughly validated on other populations and settings before being implemented.

Conclusions
Our study demonstrated a substantial prognostic value of PND on any ROP and ROP requiring treatment. Infants with PND of 14 days or more were at significantly higher risk of needing treatment. Continuous research on neonatal nutrition for premature babies is warranted.
Further, we updated and externally validated DIGIROP 2.0 prediction models and their clinical decision support tool to achieve 100% sensitivity and high specificity. Considering both sensitivity and specificity, the DIGIROP clinical decision support tool was shown to be superior to WINROP and G-ROP.