Ching-Wen Huang, Aldina Pivodic, Po-Nien Tsao, Ann Hellström, Po-Ting Yeh, Hung-Chieh Chou, Chien-Yi Chen, Ting-An Yen, Hsin-Chung Huang, Tso-Ting Lai
The DIGIROP 2.0 models demonstrated high sensitivities in identifying treatment-requiring ROP in this external validation cohort, particularly after recalibration, underscoring the importance of validation and modification when applying the models to different populations. Further large-scale studies are required to confirm the value of including PND in ROP screening models.
BACKGROUND: To evaluate the performance of the updated DIGIROP models (version 2.0), which incorporate parenteral nutrition duration (PND; protein and lipids) data, in predicting treatment-requiring retinopathy of prematurity (ROP).
METHODS: A retrospective validation study at a tertiary referral centre in Taiwan. The DIGIROP-Birth 2.0 and DIGIROP-Screen 2.0 models were applied to calculate individual risk estimates for treatment-requiring ROP. Model performance was evaluated by assessing sensitivity and specificity. Recalibration was performed based on the incidence of treatment-requiring ROP across different gestational age groups to account for differences between the study and the original developmental cohorts. The performance of the recalibrated models was then compared with that of the original DIGIROP 2.0 models.
RESULTS: Among the 303 infants included, 30 required treatment for ROP. The DIGIROP-Birth 2.0 model successfully identified 28 of the treated infants (sensitivity: 93.3% [95% CI: 77.9-99.2]; specificity: 48.0% [95% CI: 41.9-54.1]). The sensitivities of DIGIROP-Screen 2.0 ranged from 93.3% (95% CI: 77.9-99.2) to 100% (95% CI: 63.1-100) during postnatal weeks 6-14. The recalibrated DIGIROP-Birth 2.0 model correctly identified all infants with treatment-requiring ROP (sensitivity: 100.0% [95% CI: 88.4-100]; specificity: 33.3% [95% CI: 27.8-39.3]), as did the recalibrated DIGIROP-Screen 2.0 model across different postnatal weeks. A model developed using gestational age and PND in our cohort yielded 100% (95% CI: 88.4-100) sensitivity and highest specificity (66.7% [95% CI: 60.7-72.2]).
CONCLUSIONS: The DIGIROP 2.0 models demonstrated high sensitivities in identifying treatment-requiring ROP in this external validation cohort, particularly after recalibration, underscoring the importance of validation and modification when applying the models to different populations. Further large-scale studies are required to confirm the value of including PND in ROP screening models.