RESUMO
BACKGROUND: Cognitive and functional decline are common problems in older adults, especially in those 75+ years old. Currently, there is no specific plasma biomarker able to predict this decline in healthy old-age people. Machine learning (ML) is a subarea of artificial intelligence (AI), which can be used to predict outcomes Aim: This study aimed to evaluate routine laboratory variables able to predict cognitive and functional impairment, using ML algorithms, in a cohort aged 75+ years, in a one-year follow-up study. METHOD: One hundred and thirty-two older adults aged 75+ years were selected through a community-health public program or from long-term-care institutions. Their functional and cognitive performances were evaluated at baseline and one year later using a functional activities questionnaire, Mini-Mental State Examination, and the Brief Cognitive Screening Battery. Routine laboratory tests were performed at baseline. ML algorithms-random forest, support vector machine (SVM), and XGBoost-were applied in order to describe the best model able to predict cognitive and functional decline using routine tests as features. RESULTS: The random forest model showed better accuracy than other algorithms and included triglycerides, glucose, hematocrit, red cell distribution width (RDW), albumin, hemoglobin, globulin, high-density lipoprotein cholesterol (HDL-c), thyroid-stimulating hormone (TSH), creatinine, lymphocyte, erythrocyte, platelet/leucocyte (PLR), and neutrophil/leucocyte (NLR) ratios, and alanine transaminase (ALT), leukocyte, low-density lipoprotein cholesterol (LDL-c), cortisol, gamma-glutamyl transferase (GGT), and eosinophil as features to predict cognitive decline (accuracy = 0.79). For functional decline, the most important features were platelet, PLR and NLR, hemoglobin, globulin, cortisol, RDW, glucose, basophil, B12 vitamin, creatinine, GGT, ALT, aspartate transferase (AST), eosinophil, hematocrit, erythrocyte, triglycerides, HDL-c, and monocyte (accuracy = 0.92). CONCLUSIONS: Routine laboratory variables could be applied to predict cognitive and functional decline in oldest-old populations using ML algorithms.
RESUMO
O relacionamento probabilístico de registros tem sido utilizado para integrar dados dos Sistemas de Informação do Sistema Único de Saúde (SUS). Contudo, ainda são necessários mais estudos dedicados à estimativa de parâmetros para o relacionamento e a validação de seus resultados. Neste trabalho, foram relacionados os registros de dois grandes sistemas de informações do SUS: o Sistema de Informações Hospitalares (SIH) e as Autorizações de Procedimentos de Alta Complexidade (Apac) do Sistema de Informações Ambulatoriais (SIA-SUS), na modalidade Terapia Renal Substitutiva (TRS). Foram relacionados 39.448.139 registros do SIH com 645.338 da Apac/SIA-SUS. No processo foram utilizadas três técnicas para estimar os parâmetros do relacionamento, dentre elas o algoritmo EM. Para validar os resultados e definir o ponto de corte, construiu-se uma curva precision-recall (PR), utilizando-se, como padrão ouro a revisão manual por dois revisores independentes. A sensibilidade, a especificidade, o valor preditivo positivo e o valor preditivo negativo para o ponto de corte selecionado foram, respectivamente, de 0,957; 0,999; 0,962; 0,999. A concordância entre os dois revisores foi excelente (Kappa=0,956). Ao final, foram identificadas 418.336 internações referentes a 104.109 indivíduos.
Record linkage has been used to integrate data from Information System of Single Health System (SUS, acronym in Portuguese). However, studies dedicated to parameter estimation and validation of the results are still necessary. The present study described the record linkage of two Brazilian health information systems, concerning patients under renal replacement therapy: the hospital information system (SHI ? Sistema de Informações Hospitalares) and the outpatient information system (SIA ? Sistema de Informações Ambulatoriais) of SUS Overall, 39,448,139 records from SIH were linked to 645,338 records from Apac/SIA/SUS. In the process, three techniques were used to estimate the linkage parameters, including the EM algorithm. To validate the results and define the cut-off, a precision-recall curve was plotted, using as gold-standard the manual review by two independent examiners. Sensibility, specificity, positive predictive value and negative predictive value where, respectively, 0.957; 0.999; 0.962; 0.999. The agreement rate between the two reviewers was considered excellent (Kappa=0.956). As a result 418,336 hospitalizations of 104,109 patients were identified.