Kawasaki Disease (KD) is a systemic vasculitis of largely unknown etiology and is the leading cause of acquired heart disease in developed nations.1 Several diagnostic criteria have been proposed since it was first reported in 1967,2 most recently with the updated American Heart Association’s (AHA) 2017 guidelines.3 Diagnosing KD can prove difficult, as the clinical criteria for KD are also found in other pediatric conditions, such as viral syndromes, inflammatory conditions, rheumatologic conditions, toxin-mediated syndromes, and rickettsial illnesses.4–6 Furthermore, for children with KD, clinical criteria may not always be present at the time of presentation, making historical symptoms or further monitoring for new signs on physical exam critical to diagnosis in some cases.3 Incomplete KD presents a unique diagnostic challenge given the presence of few clinical criteria in affected children, particularly in young infants who are more likely to present atypically and with more severe cardiac outcomes, including giant coronary artery aneurysm.3,7 For children with incomplete KD, studies have reported longer durations of fever at the time of diagnosis and a higher prevalence of coronary artery aneurysms after intravenous immunoglobulin treatment, as compared with complete KD, likely caused by delays in diagnosis.7
Clinical overlap between KD and other conditions may lead to both underdiagnosis and misdiagnosis of KD. Consequences of KD misdiagnosis include hospitalization, intravenous immune globulin cost and side effects, unnecessary imaging and follow up, and delayed vaccinations, whereas the main consequence of missed KD diagnosis is increased risk of coronary artery aneurysm, which occurs in 25% of untreated patients.8 Of note, KD experts concur with the diagnostic challenges inherent in the evaluation of KD and have highlighted the need for a reliable KD biomarker test to be developed for more than 20 years but without definitive success in existing studies.9,10
In this issue of Hospital Pediatrics, Portman et al sought to construct a novel, artificial intelligence-derived prediction rule from blood tests for the diagnosis of KD.11 In this single-center cohort study, the authors compared blood samples from 50 patients with KD with a control cohort of 100 children with ≥2 days of fever but without a diagnosis of KD. Out of 45 initial biomarkers tested, 11 were chosen for the model based on strength of association and commercial availability via machine learning techniques.
Machine learning was developed in the 1980s and allows data scientists to perform analyses using sophisticated algorithms within statistical software, which offers great computational power as compared with manually performing mathematical equations.12 In alignment with traditional approach, Portman et al used machine learning to develop the prediction rule from the data while leaving some of the data for validating the prediction rule. The authors used 3 methods to determine the accuracy of the prediction rule. First, they used 10-fold crossvalidation utilizing one-tenth of the dataset to determine accuracy. Subsequently, because of the small validation sample in the former method, they used Monte Carlo simulation with numerous iterations to randomly select one-third of the data set (“test” set) for each simulation to determine accuracy. Third, the authors used in-sample validation for their reduced panel where they used the same whole cohort to evaluate the accuracy of the prediction rule while performing other analyses to account for potential upward bias of the accuracy using this technique. The accuracy of the prediction model was noted to be similar using all 3 statistical approaches.
Specifically, the authors found that by using Youden’s index for cutoff to determine presence or absence of KD they were able to produce an 11-variable predictive model with an area under the curve (AUC) of 0.94 (95% confidence interval [CI]: 0.90–0.98), correctly diagnosing 47 of 50 (94%) of KD-positive patients and 88 of 97 (91%) of KD-negative patients. They further reduced the panel to 3 biomarkers, C-reactive protein, NT-proB-type natriuretic peptide, and thyroid hormone uptake, because of commercial availability that had a similar AUC of 0.92 (95% CI: 0.87–0.96). The strengths of the study included robust machine-learning methodology and identification of a small number of commercially available laboratory markers with high diagnostic accuracy for KD. The study is limited by its generalizability because of the single-center design and small cohort size and by potential bias from control group selection. For the latter, some of the children in the control group had alternate diagnoses or lack of clinical criterion for KD which may lend greater ease to distinguishing controls from children with KD and thus, may yield higher estimates of diagnostic accuracy than may otherwise occur in full populations of suspected KD.
Overall, the work by Portman et al marks a valuable step in identifying an objective biomarker model that could result in a timelier and more accurate KD diagnosis to avoid the harms of both misdiagnosis and underdiagnosis of KD. Numerous studies investigating the ability of inflammatory, immunologic, proteomic, and genetic biomarkers to predict KD have been conducted but have been limited by sample size, generalizability, lack of proper control group selection, conflicting results, and poor clinical translation.10
Development of such a model has been hindered by several factors. First, because there is no gold standard for KD diagnosis, studies are forced to rely on the AHA guidelines to identify KD cohorts,3 and the diagnostic clinical criteria for KD do not exclusively exist in children with KD. Consequentially, patients with conditions that mimic KD may be diagnosed and treated for KD because of meeting clinical criteria and concern for possible sequelae of nontreatment, and thereby be included in KD study groups.
Second, and perhaps most difficult, KD is an uncommon diagnosis with variable presentation across age groups and uncommon long-term cardiac effects, making large, prospective studies difficult to conduct. Major cardiac adverse events also represent a very small percentage of children with KD and >90% of nongiant aneurysms resolve over time in patients with timely treatment.13 This makes its clinical significance difficult to assess and forces many researchers to use surrogate variables to infer clinical relevance. Additionally, accurate diagnostic tests are most needed for young infants with KD, who are more likely to present atypically and develop giant, persistent aneurysms, yet they make up a small percentage of the cohort in most studies (5 of 50 [10%] in this study).
Lastly, because the inflammatory host response in KD diagnostic tests uses similar inflammatory pathways to other acute febrile illnesses of childhood,14 proper control selection is imperative to clinical utility.15 Although earlier studies were limited by control selection, more recent studies have focused on using controls that present with conditions that mimic KD, representing the most useful clinical scenario.8 First, Tremoulet et al produced models using 11 and 16 commercially available tests among three cohorts of patients who resulted in AUC of 0.81 to 0.96, depending on cohort.15 Another more recent study among 48 KD patients and 105 febrile controls identified a model using 2 neutrophil-derived proteins and CRP as a biomarker panel that resulted in an AUC of 0.84 (95% CI: 0.80–0.88), with a sensitivity of 74% and a specificity of 83%.8 Although the study by Portman et al produced a more accurate predictive model, only 20% of the febrile control group had at least 1 criterion for KD, which could have led to an overestimation of accuracy.
Although the authors present a promising KD diagnostic model with strong statistical support, we are still a long way out from having a usable test to replace current guidelines. Large, prospectively designed studies will be needed to assess external validation as well as diagnostic performance and effect on clinical outcomes relative to the current AHA guidelines before implementation.8 Additionally, experts recommend that prospective assessment of pretest probability of KD will need to be included in such studies to determine posttest probability and utility of the novel tests undergoing investigation.15 Until then, all physicians caring for children should familiarize themselves with the AHA diagnostic criteria for complete and incomplete KD, paying close attention to the young infants who are at the highest risk of poor outcomes.
FUNDING: No external funding.
CONFLICT OF INTEREST DISCLOSURES: The authors have indicated they have no conflicts of interest to disclose.
COMPANION PAPER: A companion to this article can be found at www.hosppeds.org/cgi/doi/10.1542/hpeds.2022-006868.
Comments