Machine-learning algorithms for predicting colchicine resistance in Familial Mediterranean Fever

FMF is a genetic autoinflammatory condition marked by recurrent fever and serositis. Colchicine is the mainstay treatment. However, 5% to 10% of patients exhibit colchicine resistance, requiring alternative therapies. Early identification of resistance is vital. While some scoring systems exist for paediatric FMF, artificial intelligence's (AI's) potential to predict colchicine resistance in adult patients with FMF has not been systematically explored. This study aimed to utilize machine-learning (ML) and deep-learning (DL) algorithms to predict colchicine resistance in adult patients with FMF.

We retrospectively analysed data from 965 adult patients with FMF diagnosed according to the Tel Hashomer criteria with genetically confirmed FMF and at least 1 year of follow-up. Data for features including mutation type, specific MEFV mutations, presence of arthritis, arthralgia, oligoarthritis, age at diagnosis, and attack frequency were selected. The data were split 80:20 for training and testing. We developed a logistic regression model and a fully connected neural network. Model performance was assessed using the Area Under the Curve (AUC) of the receiver operating characteristic (ROC) curve and other performance metrics.

Both the logistic-regression and DL models achieved an AUC of 0.79. Statistically significant differences between colchicine-resistant and non-resistant groups were found for the homozygous mutation type (P = 0.0005), presence of recurrent arthritis (P = 0.0033), presence of chronic arthralgia (P = 0.0286), age of diagnosis (P = 0.0022), and frequency of attacks (P = 0.0436).

Our findings suggest that AI-based algorithms, particularly DL models, show significant potential in predicting colchicine resistance in adult patients with FMF. These models could assist in early treatment decision-making, facilitating tailored therapeutic strategies for resistant patients.