Prediction of retinopathy of prematurity development and treatment need with machine learning models
Ceren Durmaz Engin1,2, Taylan Ozturk3, Ozlem Ozkan4
1Department of Ophthalmology, Izmir Democracy University Buca Seyfi Demirsoy Education and Research Hospital, Kozagac Mah, Ozmen Cad No:147 Buca, Izmir, Turkey. cerendurmaz@gmail.com.
Insights
Machine learning models effectively predict retinopathy of prematurity (ROP) and treatment needs in preterm infants using early clinical data. These models enhance screening and clinical decisions for better infant care.
Area of Science:
- Neonatal Medicine
- Artificial Intelligence in Healthcare
- Predictive Analytics
Background:
- Retinopathy of prematurity (ROP) is a significant cause of visual impairment in preterm infants.
- Early prediction of ROP and the need for treatment is crucial for timely intervention.
- Current screening methods can be resource-intensive and may miss critical cases.
Purpose of the Study:
- To evaluate the effectiveness of machine learning (ML) models in predicting ROP occurrence.
- To assess the capability of ML models to predict the need for ROP treatment.
- To identify key predictive parameters for ROP development and treatment.
Main Methods:
- Developed four ML models using 49 parameters from birth and initial screening data.
- Employed five ML algorithms: logistic regression, decision tree, SVM, random forest, and XGBoost.
- Assessed model performance using balanced accuracy and identified top predictive parameters.
Main Results:
- Models predicting ROP development and treatment need achieved balanced accuracies of 80% and 81% using birth data.
- Models using screening data showed accuracies of 80% for ROP development and 66% for treatment need.
- Key predictors included gestational age, birth weight, APGAR scores, and erythrocyte counts.
Conclusions:
- ML models show significant potential for predicting ROP and treatment requirements.
- Integrating clinical and neonatal intensive care unit data improves ROP screening.
- ML-driven insights can enhance clinical decision-making for preterm infants at risk of ROP.
Background:
To evaluate the effectiveness of machine learning (ML) models in predicting the occurrence of retinopathy of prematurity (ROP) and treatment need.
Methods:
Four ML models were created using 49 parameters known within the first 24 h post-birth and obtained during the initial screening examination, encompassing demographic, maternal, clinical, and neonatal intensive care unit-related data. The models' performances were assessed using five machine learning (ML) classifier algorithms: logistic regression (LR), decision tree (DT), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost). Performance metrics were calculated, and the top ten parameters with the highest predictive value were identified.
Results:
In the cohort of 355 preterm infants, Model I, predicting ROP development using birth data, achieved a balanced accuracy of 80%, with gestational age (GA), birth weight (BW) and mean corpuscular volume (MCV) as the top predictive parameters. Model II, predicting treatment-requiring ROP using birth data, exhibited a balanced accuracy of 81%. Key predictive parameters included low GA, BW, 1-minute and 5-minute APGAR scores, and low erythrocyte counts. For Model III, predicting ROP using the first screening examination data, and Model IV, predicting treatment-requiring ROP using the same data, the accuracy values were 80% and 66%, respectively, with BW, daily weight gain, total O2 support duration, and platelet/lymphocyte ratio emerged as the most significant predictive parameters in both models.
Conclusion:
This study demonstrates the potential of ML models to predict ROP development and treatment need. Incorporating clinical and intensive care-related parameters can enhance ROP screening and clinical decision-making.
