111. Unraveling Craniosynostosis Heterogeneity: Harnessing Machine Learning For Comprehensive Subgroup Discovery And Genotype-phenotype Correlations

PURPOSE: Craniosynostosis, a condition defined by the early fusion of cranial sutures, affects approximately 1 in 2,000 live births, presenting a highly variable clinical spectrum. Traditional diagnostic methods, while useful, often fail to capture the nuanced differences across subtypes, leading to challenges in accurate classification, treatment planning, and outcome prediction. The heterogeneity in cranial suture involvement, comorbidities, and genetic variants further complicates the identification of clinically meaningful subgroups. This study aims to apply Uniform Manifold Approximation and Projection (UMAP), a machine learning-based dimensionality reduction technique, to identify hidden subgroups within a large cohort of craniosynostosis patients, ultimately enhancing our understanding of genotype-phenotype correlations and informing more personalized treatment approaches. METHODS: A retrospective cohort study was conducted, analyzing 719 patients diagnosed with craniosynostosis over a 16-year period at our institution. For each patient, a comprehensive dataset encompassing 352 variables was compiled, including demographic information, clinical presentation, family history, comorbidities, genetic testing results, surgical interventions, and long-term outcomes. Data preprocessing involved cleaning, imputing missing values, and normalizing continuous data, followed by encoding categorical variables. UMAP was applied to reduce the high-dimensional dataset to two-dimensional and three-dimensional representations, revealing underlying patterns. Identified clusters were statistically analyzed using t-tests and χ² tests to evaluate clinical and genetic significance, with P-values < 0.05 considered significant. Logistic regression was utilized to further explore predictive factors. RESULTS: UMAP visualization revealed several distinct clusters of patients, suggesting previously unrecognized subgroups that extended beyond conventional clinical classifications. For instance, patients with bicoronal craniosynostosis demonstrated greater heterogeneity in clinical presentation and surgical outcomes compared to other subtypes. Additionally, the analysis uncovered significant genotype-phenotype correlations, particularly in patients harboring FGFR2 and FGFR3 mutations, which were linked to three distinct sub-clusters with varying clinical profiles. One notable sub-cluster of patients with metopic craniosynostosis exhibited a higher incidence of psychiatric comorbidities, including autism spectrum disorders, necessitating ongoing psychiatric follow-up. These findings highlight the utility of UMAP in identifying novel, clinically meaningful subgroups that may inform more tailored approaches to patient management. CONCLUSION: This study demonstrates the potential of UMAP as a powerful tool for uncovering hidden patterns in complex, high-dimensional datasets associated with craniosynostosis. By revealing cryptic subgroups and previously unrecognized genotype-phenotype correlations, UMAP offers the possibility of refining diagnostic precision and guiding personalized treatment strategies. Future research should focus on expanding the dataset to include multi-institutional cohorts. To facilitate collaboration, the code developed for this analysis will be publicly shared, enabling other institutions to replicate and build upon these findings.