The role of "red flags" in the diagnostic work-up of hereditary transthyretin amyloidosis: a study using a machine-learning approach.

INTRODUCTION: Hereditary transthyretin amyloidosis (ATTRv) is a rare progressive, potentially life-threatening multisystem disorder caused by mutations in the transthyretin (TTR) gene, with variable penetrance and heterogeneous phenotypes, often leading to diagnostic delays, particularly in non-endemic regions. Identifying clinical "red flags" is crucial to shorten diagnostic latency. Machine learning (ML), a branch of artificial intelligence (AI), together with explainable artificial intelligence (XAI), offers novel opportunities to refine diagnostic algorithms and prioritize predictive features in ATTRv. MATERIALS AND METHODS: A total of 452 patients who underwent TTR genetic testing between 2019 and 2024 in Sicily were retrospectively analyzed. Genetic testing was performed via polymerase chain reaction (PCR) and sequencing of TTR exons 2-4. Patients were stratified into Western and Eastern Sicily sub-cohorts to train and validate supervised ML models. Multiple algorithms were compared with hyperparameter tuning via GridSearch with cross-validation. Model interpretability was ensured using SHapley Additive exPlanations (SHAP) values and permutation importance. RESULTS: Among 452 patients, 68 (15%) carried a TTR mutation, 51.5% of whom were symptomatic. The most frequent red flags were sensory neuropathy (62.6%) and family history of cardiomyopathy (51.8%). Tree-based models outperformed other algorithms, with Random Forest selected for its optimal balance between precision and recall. Bilateral carpal tunnel syndrome, family history of neuropathy, and ataxia emerged as the most informative predictors. DISCUSSION: These findings suggest that integrating ML with clinical red flags may support the diagnostic decision-making process in patients referred for suspected ATTRv. However, these results should be considered exploratory and require validation in independent cohorts before clinical implementation.