Clinical Study on an Artificial Intelligence-Assisted Chest Radiograph Model Based on Big Data and Deep Learning for Early Detection of Kawasaki Disease

Overview

The goal of this observational study is to develop an AI-based early warning system for Kawasaki Disease (KD) using chest X-rays (CXR) in children diagnosed with Kawasaki Disease. The main question\[s\] it aims to answer are: 1. Can AI modeling of CXR features help identify high-risk KD patients earlier than current diagnostic methods? 2. Can the AI system predict the optimal IVIG treatment window and coronary artery risks in KD patients? Participants will: Provide retrospective data on chest X-rays and clinical data (CRP, coronary ultrasound, etc.) Allow analysis of CXR features using deep learning models to extract relevant patterns Have their data incorporated into a federated learning model to ensure privacy and data security

Detailed description

1. Research Background and Clinical Pain Points Kawasaki Disease (KD) is a leading cause of acquired heart disease in children. Traditional diagnosis relies on subjective symptoms such as fever lasting ≥5 days and rashes, leading to two major problems: delayed diagnosis, with 30% of atypical patients missing the optimal IVIG treatment window (fever duration of 5-10 days); and coronary artery damage: delaying treatment for ≥7 days increases the risk of coronary dilation by 47%. The current AHA standards have only a 35% sensitivity for children with fever ≤3 days, highlighting the urgent need to establish an objective early warning system. 2. Research Objectives and Technical Approach Core breakthrough: First time using routine chest X-rays (CXR) to develop an AI-based early warning model.

Technical path: Multi-center data integration, collection of CXR and clinical data (clinical symptoms, laboratory tests, coronary ultrasound, etc.), and a federated learning framework to ensure privacy and security. Exploration of imaging biomarkers and CXR features that are invisible to the human eye, as well as the development of a multi-modal dynamic early warning model.

Dual-path CNN to extract CXR features → Graph neural networks to integrate laboratory indicators → Diagnosis model to output the risk score of kawasaki disease.↑ 3. Innovation Advantages and Clinical Value Early-warning performance was strong by day 3 of fever, achieving a pre-trial sensitivity of 87.2% for Kawasaki disease, while providing individualized IVIG treatment windows and predicted coronary-artery risk. The lightweight model (less than 50MB) is adaptable for use in primary care settings.

Clinical pathway:

AI identifies high-risk children → Priority for echocardiography → IVIG treatment window advanced → Reduction in cardiovascular complications.

The ultimate goal is to shorten diagnosis time and reduce cardiovascular complications of KD patients in China. 4. Validation Plan and Results Translation

Three-phase validation:

Internal: 5-fold cross-validation (AUC ≥0.88) External: Blind testing in 3 hospitals (sensitivity \>85%, specificity \>80%) Clinical: Real-time deployment in emergency settings (response time ≤15 seconds) Results translation: 1-2 peer-reviewed journal publications and 1-2 patent filings; facilitating early identification of Kawasaki disease, thereby improving clinical outcomes. 5. Key Conclusion This study aims to decode objective biomarkers such as pulmonary artery vascular signs in CXR images and construct the AI-CXR early warning system for KD. It will break through the current reliance on fever duration and subjective symptoms, providing support for early diagnosis and improving patient outcomes.

Interventions

• Diagnostic test AI-Based Early Warning System for Kawasaki Disease
  This study utilizes an AI-based early warning system for Kawasaki Disease (KD) to predict the optimal IVIG treatment window and assess coronary risk. The system analyzes chest X-ray (CXR) images and integrates them with clinical data such as CRP levels and clinical symptoms. The intervention involves the development of a multi-modal dynamic prediction model that uses a dual-pathway convolutional neural network (CNN) to extract relevant CXR features and a graph neural network to integrate laborat

Primary outcome measures

• Area Under Curve [Time frame: Up to 14 days after fever onset]
• sensitivity [Time frame: Up to 14 days after fever onset]
• specificity [Time frame: Up to 14 days after fever onset]

Eligibility criteria

Inclusion criteria

• Case group

• The age of seeking medical treatment is less than or equal to 18 years old; ·The medical record system diagnosis contains the diagnosis of "Kawasaki Disease", "mucocutaneous lymph node syndrome" or "IVIG non-response Kawasaki disease"
• At least one complete chest X-ray examination data (images and reports) is available during the same hospitalization
• Control group

• The age of seeking medical treatment is less than or equal to 18 years old
• The same period as the case group
• Fever lasts for 3 days or more
• Rule out the possibility of diagnosing Kawasaki disease

Exclusion criteria

• Case group

• Chest X-ray quality issues: Severe artifacts, overexposure/underexposure leading to inability to assess key structures
• Incomplete clinical information, including lack of chest X-ray examination, laboratory tests, and unclear days of fever Inability to determine the final diagnosis (such as loss to follow-up, diagnosis in doubt)
• Control group

• Chest X-ray quality issues: Severe artifacts, overexposure/underexposure leading to inability to assess key structures
• Incomplete clinical information, including lack of chest X-ray examination, laboratory tests, and unclear days of fever
• Inability to make a clear final diagnosis (such as loss to follow-up, questionable diagnosis)

Criteria are shown verbatim from the registry (in English). Final eligibility is always assessed by the study center.

Healthy volunteers: Yes

Study design

Observational model

Other

Study locations

China · 1 center

• Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine — Shanghai

Identifiers

NCT: NCT07405658 · XH-25-010

Primary sources (government registries)

• ClinicalTrials.gov (US NIH)

View this study on ClinicalTrials.gov ↗