Análise de classificação multirrótulo em desfechos da doença falciforme.

Abstract

Sickle cell disease is a genetic condition that affects hemoglobin, resulting in systemic consequences that influence patients' quality of life and survival. Hemoglobin electrophoresis is gold standard technique for diagnosing the disease and is difficult to access in remote areas of Brazil and several developing countries, such as the African continent. Clinical heterogeneity and disease severity stratification have been tested in medical practice, so predicting results involves analyzing a series of clinical and laboratory markers.

Several studies have discussed the application of artificial intelligence models in diagnosing sickle cell disease and its monitoring, with the authors focusing on the traditional classification, where a single stage is developed. This condition limits some health problems where there is more than one stage per patient. The purpose of this study was to use multi-label algorithms to observe the relationships between the outcomes of sickle cell disease and to evaluate the hypothesis that using artificial intelligence models can allow for better stratification of patients with the disease, considering the multiple clinical outcomes.

A database with 609 patients and 63 variables was used. The database was obtained from experimental cross-sectional cohort studies generated by the Anemia Research Laboratory of the School of Pharmacy (UFBA) and the Translational Genetics and Hematology Research Laboratory (LIGHT - FIOCRUZ-BA). Multi-label algorithms Binary Relevance (BR), Classifier Chains (CC), Label Powerset (LP), and Random k-label sets (RAKEL) were used to predict the development of vaso-occlusion, pain crises, hospitalization, and infections. These algorithms are used in multi-label transformation problems, which transform the multi-label problem into a series of independent, multi-class, and ensemble binary classification problems.

The results showed that when the order of the stages is relevant, the RAkELo model with Random Forest stands out, presenting a Label Ranking Loss of 0.52. On the other hand, when ordering is not an important clinical factor, the SVM and Random Forest-based models, in all approaches (BR, CC, LP, RAkELo), presented an optimal combination of low Hamming Loss (0.20) and high microaverage f1-score (0.89).