Machine learning-assisted prediction of speech outcomes at age 5 in patients with nonsyndromic cleft palate after double-opposing Z-plasty.

Abstract

Various risk factors have been identified for persistent speech impairment after cleft palate repair; however, early risk stratification tools are lacking. This retrospective cohort comprised 1117 patients with nonsyndromic cleft palate (NSCP) who underwent double-opposing Z-plasty (DOZ) by a single surgeon between 1988 and 2017. Perioperative variables included sex, age at surgery, Veau classification, cleft width measured intraoperatively at the bilateral maxillary tuberosities, average palatal length (mean of right and left side measurements), and presence of early palatal fistulae (<1 cm² within 1 month). Outcomes were assessed at age 5 years using standardized speech evaluations conducted by certified speech-language pathologists, focusing on 2 endpoints: the probability of hypernasality and need for speech therapy. A random forest classifier was trained on 80% of the dataset and tested on 20%, incorporating multiple imputation by chained equations, the Synthetic Minority Oversampling Technique, and five-fold cross-validation. The model achieved an area under the receiver operating characteristic (AUROC) of 0.747, area under the precision-recall curve (AUPRC) of 0.352, and F1-score of 0.435 for predicting hypernasality and an AUROC of 0.645, AUPRC of 0.545, and F1-score of 0.611 for predicting the need for therapy. Feature importance analysis identified sex, Veau classification, and cleft width as the most influential predictors. Early small fistulae contributed moderately to hypernasality prediction but were negligible for therapy prediction. This machine learning model, implemented as a web-based calculator (https://cp-speech.mdbcdss.com/), provides an accessible decision-support tool for early risk stratification of speech outcomes at age 5 after DOZ in patients with NSCP.