<< Back to the abstract archive
Using Machine Learning to Quantify Severity of Sagittal Craniosynostosis in 195 Patients
Tobi J. Somorin, BS, Wenzheng Tao, MS, Janina Kueper, MD, Angel Dixon, BA, Nicolas Kass, BA, Nawazish Khan, MS, Krithika Iyer, BSE, Jake Wagoner, BS, Ashley Rodgers, MD, Ross Whitaker, PhD, Shireen Elhabian, PhD, Jesse A. Goldstein, MD.
University of Pittsburgh Medical Center, Children's Hospital of Pittsburgh
2025-01-09
Presenter: Tobi Somorin
Affidavit:
The entirety of this submission represents the original work of the research team.
Director Name: Jesse A. Goldstein
Author Category: Medical Student
Presentation Category: Clinical
Abstract Category: Craniomaxillofacial
Introduction: Sagittal craniosynostosis (SCS) presents with complex phenotypes requiring precise assessments to ensure quality care. Here we describe the development of the sagittal severity score (SSS) using validated machine learning (ML) techniques and describe the spectrum of severity across a large population of patients with SCS.
Methods: A survey with an internet-based rating portal was administered to expert surgeons. Respondents evaluated the severity of a series of SCS and normal control images. These ratings validated a principal-component-analysis ML tool to produce the SSS. A leave-one-out-cross-validation (LOOCV) analysis evaluated SSS against the prior gold standard, the cephalic index (CI). We assessed its performance in comparison with the cranial morphology deviation (CMD) score, which is based on unsupervised ML, and applied the SSS to a large population of patients with SCS.
Results: The survey was completed by 54 surgeons, each rating a random set of 20 patients (75% SCS:25% control), resulting in 1080 rating records. The SSS achieved a significantly higher accuracy in predicting expert responses than CI in LOOCV (p<0.01). The CMD score showed a strong correlation with the SSS (Pearson coefficient of 0.92, Spearman coefficient of 0.90, p< 0.01). We present the frequency distribution of SSS in 195 patients with SCS and 180 controls.
Conclusions: We introduce the SSS as a novel metric to quantify SCS severity and describe the distribution of phenotypic severity in a large cohort of patients. These methods pave the way for more objective, precise, and consistent evaluations in patients with SCS.
