Deep learning to assess laryngoscope insertion depth during neonatal intubation with video laryngoscopy

TL;DR

This study uses deep learning to classify laryngoscope insertion depth during neonatal intubation, aiming to improve real-time guidance and training.

Contribution

A deep learning model was developed and validated for real-time insertion depth classification during neonatal video laryngoscopy.

Findings

The model achieved high F1 scores for detecting glottic zone and shallow insertions.

Deep insertion events were rare and poorly detected by the model.

Most clinicians preferred visual feedback over voice or haptic modalities.

Abstract

Video laryngoscopy (VL) improves glottic visualization during neonatal intubation, but real-time guidance on blade insertion depth is lacking. We developed a deep learning model to classify insertion depth during VL as shallow, glottic zone, or deep.

A deep learning model was trained on 298,955 annotated frames from 132 neonatal VL videos from two NICUs on one platform with fivefold cross-validation. 31 clinicians were surveyed regarding preferred device feedback modalities.

The model detected glottic zone and shallow insertion depths with F1 scores of 0.894 and 0.718, respectively. Deep insertion events were rare (2.7%) with low performance (F1 = 0.034). Most clinicians preferred visual, minimal prompts over voice or haptic feedback.

AI-enabled VL may support blade-insertion depth assessment and training. Given the rarity of deep events, conclusions about deep insertion and clinical impact are limited. Future multi-site studies should focus on clinical integration and assessing outcomes.

Figures