Essential Site Maintenance: Authorea-powered sites will be updated circa 15:00-17:00 Eastern on Tuesday 5 November.
There should be no interruption to normal services, but please contact us at [email protected] in case you face any issues.

loading page

Utilizing Medical Health Records and Imaging Data for Kawasaki Disease Diagnosis: A Multimodal Approach
  • +2
  • Zhixin Li,
  • Luo Gang,
  • Ji Zhixian,
  • Wang Sibao,
  • Pan Silin
Zhixin Li
Qingdao Women and Children's Hospital
Author Profile
Luo Gang
Qingdao Women and Children's Hospital
Author Profile
Ji Zhixian
Qingdao Women and Children's Hospital
Author Profile
Wang Sibao
Qingdao Women and Children's Hospital
Author Profile
Pan Silin
Qingdao Women and Children's Hospital

Corresponding Author:[email protected]

Author Profile

Abstract

Objective: To develop a multimodal artificial intelligence model based on medical health records, imaging information, and laboratory test indicators to assist in the diagnosis of Kawasaki disease. Methods: This study was conducted using the Kawasaki disease database from our hospital, retrospectively collecting medical information from a total of 500 children (both with Kawasaki disease and healthy). We designed a Chinese-BERT-Base module, a ResNet module, and a fully connected layer module to process medical records, image data, and laboratory test results, respectively. Subsequently, we utilized early fusion to concatenate the vectors and input them into a classifier for outputting classification results. We designed unimodal models and traditional machine learning models for comparative evaluation to assess the effectiveness of our model. We analyzed the attention of each module to the raw data to evaluate the interpretability of the model. Additionally, we collected data from another 100 children from a peer hospital as an external validation group. In a double-blind scenario, three senior doctors and the model performed classifications simultaneously for a human-machine comparison experiment. Results: The multimodal model developed in this study demonstrated significant improvements in accuracy (93%) and specificity (93%) compared to unimodal models and traditional machine learning models. Interpretability analysis showed that the attention of each module in the multimodal model largely aligned with the thought processes of human doctors. The human-machine comparison experiment indicated that the model’s classification performance (87%) still had a notable gap compared to that of the human doctors (98%). Conclusion: The multimodal model developed in this study effectively utilized clinical data, achieving good diagnostic performance and providing a relatively reliable tool to assist clinicians in diagnosing Kawasaki disease.
Submitted to Pediatric Allergy and Immunology
Submission Checks Completed
Assigned to Editor
Reviewer(s) Assigned
27 Jun 2024Review(s) Completed, Editorial Evaluation Pending