Evaluation and accurate diagnoses of pediatric diseases using artificial intelligence

Area of Science:

• Medical Informatics
• Computational Medicine
• Artificial Intelligence in Healthcare

Background:

• Machine learning classifiers (MLCs) excel in image-based diagnostics but struggle with diverse electronic health record (EHR) data.
• Analyzing large-scale EHRs presents challenges for traditional statistical methods.
• Physician-like reasoning is needed to extract meaningful insights from complex patient data.

Purpose of the Study:

• To develop and validate an AI-based framework for analyzing EHR data.
• To demonstrate the capability of MLCs in querying EHRs and identifying novel clinical associations.
• To assess the diagnostic accuracy of the AI model compared to experienced pediatricians.

Main Methods:

• An automated natural language processing (NLP) system using deep learning was employed to extract clinical information from EHRs.
• The framework was trained and validated on 101.6 million data points from 1,362,559 pediatric patient visits.
• MLCs were utilized to query EHRs, mimicking hypothetico-deductive reasoning.

Main Results:

• The AI model achieved high diagnostic accuracy across multiple organ systems.
• The model's diagnostic performance was comparable to that of experienced pediatricians in identifying common childhood diseases.
• Novel associations within EHR data, previously undiscovered by statistical methods, were unearthed.

Conclusions:

• AI-based systems can effectively augment physician diagnostic capabilities by analyzing large volumes of EHR data.
• The developed AI framework serves as a proof of concept for clinical decision support, particularly in complex or uncertain cases.
• This technology has the potential to universally benefit healthcare by improving diagnostic efficiency and accuracy.