Machine Learning Aplicado a Triagem de Osteoporose: modelo baseado na atenuação de ondas eletromagnéticas

Abstract

Osteoporosis is a silent and still underdiagnosed condition, with a mortality rate higher than several types of cancer, especially when patients suffer fractures. The gold standard equipment for the diagnosis, Dual-energy X-ray absorptiometry (DXA, or DEXA), which uses ionizing radiation and is expensive, is scarce in countries considered middle or low-income, thus hindering timely access to diagnosis. In this context, a portable device, Osseus, was developed for the screening of patients who need the densitometry exam, i.e., to qualify the referrals of exams to the DEXA equipment. The thesis aimed to validate the Osseus device using machine learning techniques. For this, the planning and data collection of 505 patients who underwent the exam at DEXA and Osseus. 21.8% of them were healthy and 78.2% were diseased (they had low bone mineral density or osteoporosis). The dataset was separated into 80% for training and validation (5-fold cross-validation) and 20% for testing. The performance obtained in the test base with the best model (Random Forest) corresponded to sensitivity=0.853, specificity=0.871, and F1(harmonic average of precision and sensitivity rate)=0.859. The results showed that the most relevant variables to indicate the individual health status were age, body mass index (BMI), and the attenuation of the signal emitted and detected by the Osseus device.When compared to the results of DEXA scans, the model has proven to be effective and consistent in screening individuals with osteoporosis and facilitating early diagnosis of the disease, which consequently entails improved productivity and reduced costs for surgery, treatment, and hospitalization. Thus, by qualifying the referral of patients from primary care to the specialized network, Osseus can impact the reduction of waiting lines of the Brazilian National Health System.