Kawasaki Disease (KD) 2021

What is Kawasaki Disease?

Kawasaki disease (KD) is relatively uncommon, mostly affecting children under the age of 5 years but can occur in older children. It is not known what causes KD and there is currently no diagnostic test, leaving doctors to diagnose the disease based on clinical criteria (such as presence of fever, rash, swollen lymph nodes and red eyes). The most serious complication of KD is damage to the coronary arteries, potentially requiring long-term management.


A need therefore exists for methods for diagnosis, prognosis, and/or monitoring of Kawasaki disease, and associated outcomes, including permanent coronary artery and heart damage in children.


Seattle Children’s Hospital identified Prevencio as having expertise in developing highly accurate, AI-driven, multiple protein, algorithmically scored blood tests. We developed a partnership and collaborated to develop a highly accurate multiple protein panel for the diagnosis of Kawasaki Disease.

International Kawasaki Disease Symposium (IKDS) Scientific Sessions

October 2021


Artificial Intelligence Derived Proteomic Panel to Diagnose Kawasaki Disease 

Highlights

  • Prospective blood samples were collected from children presenting to the emergency room with symptoms of fever ≥ 38.1 0 C to develop a diagnostic blood test for Kawasaki Disease (KD).

 

  • The final updated panel identified had three biomarkers: NT-proBNP, C-reactive protein (CRP) and T Uptake (TU).

  • The model had an AUC of 0.92 (95% C.I.:0.87, 0.96) for accurate diagnosis of KD. With the three-level risk score, scaled 1-3, diagnoses of low-risk patients (Score=1) had a NPV of 96%, and high-risk patients (Score=3) had a PPV of 86%.

 

  • This proteomic panel can provide rapid laboratory confirmation of KD diagnosis, thus decreasing time to diagnose and treat patients positive for KD. All 3 assays can be run on the same laboratory immunoassay platform with results inputted to the AI model

Study Overview


BACKGROUND

Kawasaki Disease (KD) diagnosis is currently based mainly on American Heart Association (AHA) algorithms. These include clinical symptoms, which occur commonly in multiple childhood diseases, and some laboratory values. No specific or sensitive blood test or panel exists to confirm the diagnosis. Thus, the diagnosis remains challenging for clinicians, particularly those rarely encountering KD.


GOAL

Using proteomics and machine learning, a subset of Artificial Intelligence (AI), develop an accurate blood panel to diagnose KD. Further, to have a commercially viable panel for existing clinical laboratory platforms and analytes already FDA approved and to optimize for 3 proteins or analytes to reduce complexity.


HYPOTHESIS

A multiple biomarker panel approach would provide accurate, biologically based diagnostic information to identify Kawasaki Disease. 


METHODS
Blood samples from children with symptoms of fever ≥ 38.10 C that presented to the emergency room (n=100 controls) and a cohort of children fulfilling KD diagnosis by American Heart Association criteria (n=50 cases). A machine learning method (least angle regression) was used to select protein panels accurate for KD diagnosis. Then a diagnostic model was trained on the 150 patient samples using LASSO with logistic regression. This model was evaluated using in-sample validation.

 


Key Findings

 

  • The final panel identified had three biomarkers: NT-proBNP, C-reactive protein (CRP) and T Uptake (TU).

  • With the three-level risk score, scaled 1-3, diagnoses of low-risk patients (Score=1) had a NPV of 96%, and high-risk patients (Score=3) had a PPV of 86%.

  • Using a protein-based approach and machine learning, we developed and internally validated a multiple blood protein panel with high accuracy for predicting the presence of KD.

 

  • This proteomic panel can provide rapid laboratory confirmation of KD diagnosis, thus decreasing time to diagnose and treat patients positive for KD. All 3 assays can be run on the same laboratory immunoassay platform with results inputted to the AI model

 


Conclusions

 

  • Using a protein/analyte-based approach and machine learning, we developed and internally validated a multiple blood assay panel with high accuracy for predicting the presence of KD.

 

  • NT-pro BNP and CRP individually were no surprise but do not provide clinical specificity needed for KD diagnosis.

 

  • AI and machine learning methods with addition of T Uptake (TU) provide a robust method of diagnosing KD.

 

  • This 3 assay panel can be performed on existing laboratory platforms with results inputted to the AI model and can provide a straightforward and rapid confirmation of KD diagnosis.

 

  • In the ED or practitioner office, this KD panel could enhance and decrease time to diagnose, eliminate multiple patient visits and treat those patients positive for KD.