310 A Deep Learning Approach to Predicting Neurosurgery Residency Interview Offers Using Neural Networks

View at publisher Summarise with VIRAT
Authors: Rommi Kashlan, Razan R. Faraj, Robert G. Briggs, David J. Coté, Ryan Chung, Angela P. Mihalic, Gabriel Zada, Jonathan A. Grossberg
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

INTRODUCTION: Residency application strategy remains opaque, with limited tools for predicting interview outcomes. We present a novel dual-tower deep learning framework that integrates applicant and program features to model neurosurgery interview offers, enhanced with a meta-classifier for stacked prediction to provide program-specific probabilities of interview succes METHODS: We extracted 2,240 applications from 460 applicants across over 100 U.S. neurosurgery programs (2022–2025) from the Texas Star database to train a dual-encoder architecture with separate neural towers for applicant and program features. input features included USMLE Step 2 CK score, research output, geographic ties, signal, away rotation, and program characteristics such as Doximity reputation and size. Additional engineered features included interaction terms (e.g., Step2CK × program rank) and normalized ratios (e.g., publications per research experience). Both MLP-based and dot-product-based towers were trained with binary cross-entropy loss and fine-tuned using pairwise ranking loss. A logistic regression meta-learner was trained on the sigmoid outputs of each base model. Evaluation was conducted on a 20% held-out validation set, filtered for programs with >50 applicants. RESULTS: The final stacked model achieved an accuracy of 70.3%, a macro-average F1 score of 0.65, and a ROC AUC of 0.83 on the validation set. For applicants who received interview offers, precision was 0.50 and recall was 0.83 (F1 = 0.62). For those who did not receive interviews, precision was 0.88 and recall was 0.60 (F1 = 0.71). The model correctly identified 668 of 879 applications that received an interview, and rejected 1,442 of 2,099 that did not. CONCLUSIONS: This model demonstrates the feasibility of using deep learning with structured application data to predict interview offer probability. It may serve as a powerful decision support tool to help neurosurgery applicants target programs more strategically and transparently. Ongoing work will incorporate deeper architecture to further enhance predictive performance and utility.

Fields of Study & Subjects

Medical Education and Admissions AI and HR Technologies Online Learning and Analytics

Publication details

Journal / source
Neurosurgery
Volume
72
Issue
Supplement_1
Pages
79–80
Publication date
26 March 2026
Language
EN