Abstract
<title>Abstract</title> <p>Background Retinopathy of prematurity (ROP) is the leading cause of irreversible childhood blindness worldwide. The selection of surgical intervention timing and prognosis assessment for ROP are highly dependent on clinicians' experience and subjective interpretation of imaging findings. At present, there is a lack of studies that can systematically integrate multimodal data (including cardiovascular system indicators, perinatal metabolic indicators, and temporal fundus image features) to predict the need for surgery and the risk of reoperation. This study aimed to develop and internally evaluate a multimodal machine learning model to improve the objectivity of ROP treatment decision support. Methods A total of 213 preterm infant cases were enrolled in this study, with data including cardiovascular indicators (congenital heart disease (CHD), patent ductus arteriosus (PDA), persistent pulmonary hypertension of the newborn (PPHN), the lowest hemoglobin level during hospitalization), metabolic indicators (gestational diabetes mellitus), and a series of fundus images. Four multimodal fusion models (Transformer_Fusion, TwoTower, EarlyFusion_Attention, and LateFusion_LSTM) were trained to predict the requirement for surgical or interventional treatment and prognosis among patients requiring treatment. Model performance was evaluated using indicators including accuracy, F1-score, and area under the curve (AUC), and the optimal model was selected for final evaluation on the test set. Results Among the 213 infants, 179 did not require surgery and 34 required surgical intervention, of whom 4 required secondary surgery. The EarlyFusion_Attention model was selected as the best-performing model for detailed test-set analysis. In the test set, the model achieved a surgical decision accuracy of 62.79% and a weighted F1 score of 0.6829. The surgical decision ROC AUC was 0.80. The model correctly identified all 6 patients who required surgery, with a sensitivity of 1.000 and a specificity of 0.568. Among the 6 patients who actually required surgery, the prognosis prediction accuracy was 83.3%, the weighted prognosis F1 score was 0.8519, and the ROC AUC was 1.00. However, the prognosis result should be interpreted cautiously because the surgical-patient subgroup contained only one Secondary case. Conclusion The EarlyFusion_Attention model showed the best overall performance among the four multimodal fusion models and achieved high sensitivity for identifying patients requiring surgery. However, its specificity remained moderate, and the prognosis prediction results were limited by the small number of surgical and secondary-surgery cases. These findings suggest that multimodal machine learning may provide supportive information for ROP treatment decision-making, but external validation in larger cohorts is required.</p>