JAMA Pediatrics Author Interviews

Introduction

Host: Dimitri Christakis, Editor-in-Chief of JAMA Pediatrics
Guest: Dr. Bimal Desai, Chief Health Informatics Officer at Children’s Hospital of Philadelphia
Co-host: Dr. Alison Galbraith, Associate Editor and Professor of Pediatrics at Boston Medical Center
Topic: Artificial Intelligence (AI) in Pediatric Health and Healthcare

Title: Development and Validation of an Automated Classifier to Diagnose Acute Otitis Media in Children
Authors: Sheik and colleagues
Objective: Develop and validate an AI decision support tool for diagnosing otitis media from videos of tympanic membranes
Methods:
- Utilization of a medical grade smartphone app with a camera attached to an otoscope
- Videos from pediatric well and sick visits were collected and annotated
- Trained using a deep residual recurrent neural network (RNN) and a decision tree approach
- Trained a noise quality filter for video adequacy
Results:
- Over 1000 videos from 600+ children, mostly under 3 years old
- Deep RNN and Decision Tree had similar accuracy: Sensitivity ~94%, Specificity ~93-94%
- Both models outperformed primary care physicians and advanced practice clinicians
- Issues: 25% of videos excluded due to occlusion (e.g., wax)
- No patient demographic collection (unknown diversity)
Conclusion: AI algorithms can improve diagnostic accuracy and could be used in primary care settings by trained non-physicians

AI includes machine learning and deep learning
Different algorithms for image classification, clustering, and prediction
ChatGPT is a specific type of large language model used for natural language processing (NLP)
Comparison of ChatGPT to image classification models used in the discussed study

AI commonly used in various everyday technologies (e.g., lane assist, Amazon recommendations, smart doorbells, voice assistants)
Future AI applications in clinical practice:
- Drafting responses to patient messages
- Summarizing patient records
- Automation of non-value-added tasks
AI could eventually summarize complex patient details more efficiently than current practices

Caution needed regarding biases in AI models (e.g., ethnic diversity in training data)
Previous biases include misrepresentation based on utilization patterns
Multiple layers need to be validated: inputs, processing, and outputs
Real-world examples: overly general models or biased application (e.g., double-booking based on predicted no-show)

Importance of applying AI with the same rigor as other diagnostic tests
Examples of misapplied AI: Apple Watch atrial fibrillation algorithm not translating well to real-world prevalence rates
AI should enhance rather than shortcut standard clinical validation processes
Potential for AI to improve efficiency and accuracy significantly in clinical settings

The implications and future roles of AI in pediatric healthcare are vast, but proper validation, application, and bias monitoring are crucial.
Acknowledgment of potential, yet cautious optimism for AI-enhanced clinical practice.
Thank you to Dr. Bimal Desai for sharing his expertise.

Produced by Shelley Steffens at the JAMA Network. Visit jamanetworkaudio.com for more podcasts.