Correspondence to "Prediction of severe retinopathy of prematurity in 24-30 weeks gestation infants using birth characteristics".

To the Editor:

We read with great interest Journal Club publication entitled Prediction of severe retinopathy of prematurity in 24–30 weeks gestation infants using birth characteristics by Dr. R. E. Zackula and Dr. T. S. Raghuveer [1]. We are grateful for the thorough review of DIGIROP-Birth, our prediction model for ROP treatment (ROPT), and for having its appropriateness evaluated by the newly developed PROBAST instrument assessing potential risks of bias [2-4]. Below we provide justifications to the raised questions with highest concern.

Were all inclusions and exclusions of participants appropriate? (development and validation)

DIGIROP-Birth was based on 6947 infants born 2007–2017 at gestational age (GA) 24–30 weeks included in SWEDROP, the Swedish ROP registry. Of those, 289 (4.2%) had ROPT. From the development group, 94/7041 (1.3%) infants were excluded for missing data or date inconsistencies, 3 had ROPT. GA at birth and sex were similarly distributed in the excluded vs development group, 28.4 (SD 1.8) vs 28.3 (SD 1.9) weeks, and 47.9 vs 45.1% girls.

SWEDROP does not include race/ethnicity. A thorough validation of a model is a prerequisite for clinical implementation. If required, the model selection and/or parameter estimates might be re-evaluated for a specific population or clinical setting.

Were there a reasonable number of participants with the outcome? (validation)

Although separately evaluated in our publication, 153/2122 (7.2%) infants had ROPT in the Swedish, German and US validation datasets. Per the PROBAST instrument, validation with ≥100 events is recommended [4].

Were participants with missing data handled appropriately? (development and validation)

Reducing the effective sample by 1% (3/292 excluded events) and having no indication of infant selection in the excluded group, biased estimates were not expected. Internal validation including cross-validation and calibration plots, and external validations were affirmative.

Was selection of predictors based on univariable analysis avoided? (development)

DIGIROP-Birth aimed to include few well-known risk factors available for all infants at birth; GA, sex and birth weight (BW) (z-score). Hence, univariable analyses were not required. The model was consecutively extended, starting with GA.

Concern was raised regarding multicollinearity for GA and BW z-score. We expect high correlation between GA and BW, r = 0.79 in this cohort. However, BW z-score extracts rest of the immaturity effect beside GA (and sex), r = −0.06. Therefore, we did not expect multicollinearity problem.

Were relevant model performance measures evaluated appropriately? (validation)

In the DIGIROP-Birth publication the optimal cut-offs were not investigated. We identified cut-offs in our publication for the extended model incorporating ROP progression data (DIGIROP-Screen), proposing a clinical decision support tool [5]. Pre-specified cut-offs were defined using development group and 100% sensitivity. Applying those cut-offs on DIGIROP-Birth external validation cohort, sensitivity 149/153 (97.4%) and specificity 886/1969 (45.0%) were achieved. Further improvement of DIGIROP-Birth adding other known neo- and perinatal risk factors is planned in near future. Including more variables in a model increases the need for consideration of robust regression techniques that might reduce the potential risk for under- and overfitting.

We sincerely encourage and are thankful for all the efforts put on the validations and discussions of our work.