Artificial intelligence for the detection of systemic amyloidosis

DELFINA CIRELLI; NICOLAS QUIROS; CARRETERO, MARCELINA; MARIA LOURDES POSADAS MARTINEZ

Simposio; International Symposium on Amyloidosis 2024; 2024

ID 105Artificial intelligence for the detection of systemic amyloidosisDelfina Cirelli,Nicolas Hugo Quiroz,Maria Victoria Anconetani,Marcelina Carretero,erika.brulc@hospitalitaliano.org.ar,MARIA ADELA AGUIRRE,Gabriela Alejandra Blugerman,ELSA NUCIFORA,Marcelo Raul Risk,MARIA LOURDES POSADAS MARTINEZKeywords: Diagnosis, Artificial intelligence, AmyloidosiBackground: Early detection of systemic amyloidosis is crucial for prognosis and quality of life. Artificial intelligence (AI) predictive models offer the potential for timely identification. Studies using medical records for diagnostic algorithms are lacking.Objective: To develop an AI-driven predictive model for systemic amyloidosis identification.Materials & methods: Dynamic retrospective cohort study. All adult patients affiliated to the Medical Care Program at the Hospital Italiano de Buenos Aires were included. For each case of amyloidosis, up to four controls without amyloidosis of the same age, gender, and affiliation period were matched. A retrospective collection of all available information was conducted for each case and their respective controls spanned from six months before diagnosis. The collected data included demographic, clinical, laboratory, and imaging information. Seven machine-learning models, employing resampling techniques, were applied to training and testing cohorts with internal validation through cross-validation.Results: Among 961 patients, 218 cases and 743 controls met the selection criteria. The mean age was 65 years, with 63% male. The predictive variables that best identify amyloidosis included carpal tunnel syndrome history, arrhythmia, abnormalities in free light chains, and imaging. When assessing performance details, the random forest and decision tree models exhibited the best AUC (over 90%) consistently in both cohorts.Summary & conclusion: Medical record features demonstrated strong performance in early amyloidosis identification using the developed AI-driven model, particularly with the random forest and decision tree algorithms.