Prediction of FMF Symptoms with Deep Learning Using Genetic Mutations and Clinical Findings

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Familial Mediterranean Fever is a hereditary autoinflammatory disorder characterized by recurrent episodes of peritonitis, pleuritis, arthritis, fever, and abdominal pain, with genetic mutations playing a critical role in disease manifestation as well as symptom type and severity. This study aimed to resolve the genotype–phenotype uncertainty in Familial Mediterranean Fever by integrating genetic mutations, demographic characteristics, and clinical data through deep learning methods, thereby enhancing the predictive understanding of symptom development and disease-related risks. A total of 1452 individuals diagnosed with Familial Mediterranean Fever at the Department of Medical Genetics, Bolu Abant İzzet Baysal University Training and Research Hospital between 2016 and 2019 were included. Genetic testing was performed via Real-Time Polymerase Chain Reaction, and clinical data were analyzed using deep learning approaches. Models applied using convolutional neural network, Residual Blocks, and Multi-Head Attention showed strong potential to predict symptom occurrence with high accuracy; specifically, they achieved the highest accuracy and receiver operating characteristic area under the curve for skin rash/redness (accuracy: 90%, Area Under the Curve: 0.873) and joint swelling (accuracy: 88.6%, Area Under the Curve: 0.871). Prediction for chest pain was also strong (Area Under the Curve 0.827), while abdominal pain, joint pain, and fever demonstrated moderate predictive performance (Area Under the Curve 0.719–0.751). Synthetic Minority Over-sampling Technique enhanced prediction for rare symptoms in imbalanced datasets. These findings suggest that deep learning models could contribute to clarifying the relationships between genetic mutations and Familial Mediterranean Fever clinical manifestations; furthermore, they demonstrate a potential to improve the understanding of genotype–phenotype correlations by predicting symptoms and risks, thereby helping to enhance the clinical utility of personalized Familial Mediterranean Fever management.