Exploring the Feasibility of a Deep Learning Algorithm for Postoperative Outcome Assessment in Unilateral Cleft Lip Repair: A Pilot Study

Primary surgical repair for a cleft lip is often performed around 3 to 4 months of age. Traditional outcome assessments rely heavily on subjective clinical judgment and structured rating tools such as the Cleft Aesthetic Rating Scale (CARS), both of which are limited by inter-rater variability and lack of scalability. This pilot study explores the feasibility of a deep learning (DL) algorithm to assist in postoperative outcome assessment and identify patients at increased risk for revision following unilateral cleft lip repair. The authors developed a convolutional neural network based on the EfficientNet-B1 architecture, trained on a class-balanced data set of 500 standardized postoperative facial photographs labeled using real-world revision outcomes. Rigorous data preprocessing, augmentation, and validation protocols were used to ensure model robustness. The model achieved an accuracy of 74% and an area under the ROC curve (AUC) of 0.79 on an independent test set, with a recall of 76% for identifying patients needing revision. Confidence score distribution and t-SNE feature space visualization demonstrated reliable class separation and interpretability. Our findings suggest that DL algorithms, when trained on structured clinical data, can offer meaningful support in cleft care. This model serves as an early prototype for AI-assisted triage systems that could aid in identifying revision candidates, particularly in resource-limited or outreach settings. While preliminary, this study highlights the potential for integrating artificial intelligence into routine cleft lip postoperative workflows to improve equity and consistency of care.