Closed Loop Computer-based Artificial Intelligence Model "Maverik" to Help Diagnose and Differential Diagnoses of Childhood Onset Chronic Nonbacterial Osteomyelitis: Pilot Study.

BACKGROUND

Childhood-onset chronic nonbacterial osteomyelitis (CNO) is an inflammatory bone disease that has become better defined in the last 2 decades and is frequently encountered in pediatric rheumatology. As the disease is still not well known and is often confused with malignancy and growth pains, it can easily be missed in clinical practice. We aimed to develop and evaluate a computer-aided, physician-friendly model for detecting CNO using closed-loop artificial intelligence (AI).

METHODS

Python software language, TensorFlow AI library, and Recurrent Neural Network were used to develop the model. Data from 83 cases of CNO, 9 cases of growth pain (GP), 9 cases of bone tumors, 9 cases of juvenile idiopathic arthritis, and 30 healthy controls (HCs) were used to train the model. The medical data for the cases were digitized as 1 (abnormal), 0 (normal), and -1 (abnormal). The dataset was scaled by 20 to reach 2800 cases, with 80% used for training and 20% for testing. A dataset of 30 cases, unknown to the model and pediatric rheumatologist, was presented, and the results were compared.

RESULTS

The error rate was ~0.5 in the first few minutes of model training. In the next generation of Maverik, this rate decreased to 0.028. The training took 62 minutes. The model correctly identified the CNO, GP, and HCs.

CONCLUSIONS

Our study is the first pilot study in the literature to develop and test an AI model as a diagnostic tool for CNO. We recommend creating the model using real-time participant data from a larger population with multicenter participation and then testing its applicability.