Dados do Trabalho

Título

Sleep variables and comorbidity indicators in depression in middle-aged and elderly adults: a machine learning approach

Introdução

Depressive Disorder (DD) is highly prevalent in middle-aged and elderly individuals, especially when associated with multimorbidities such as obesity and diabetes. Also, high levels of excessive daytime sleepiness (EDS) and low levels of physical activity (PA) are markers of depression. In this context, diagnosing DD takes time through the use of complex subjective tools, which urges the need for efficient screening practices for the identification of depressive symptoms, including depression biomarkers related to comorbidities, sleep habits and lifestyle.

Objetivo

The aim of this study was to develop models based on Machine Learning (ML) to predict the occurrence of depressive symptoms in middle-aged and elderly adults, based on sleep variables, presence of EDS, physical activity level, anthropometric measurements, and biomarkers of obesity and diabetes.

Métodos

Data from the US National Health and Nutrition Examination Survey (NHANES) database, were used, including the Global Physical Activity Questionnaire (GPAQ - physical activity level), the Patient Health Questionnaire (PHQ-9), and the sleep habits questionnaire (sleep onset, offset, duration and EDS). In addition, anthropometric measurements (waist circumference and BMI), plasmatic biomarkers of obesity (C-Reactive Protein, total cholesterol and HDL), and diabetes (glycated hemoglobin) were included to estimate the presence of depressive symptoms by the implementation of three ML supervised learning algorithms: Penalized Logistic Regression (PLR), Random Forest (RF) and Gradient Boosting (GB).

Resultados

2161 individuals of both sexes (52.9% women), between 40 and 80 years of age were classified according to the presence or absence of depressive symptoms. Among the three models, Random Forest showed the best performance with an area under the curve (AUC) of 0.88 and an accuracy (AC) of 0.83, when compared to GB (AUC = 0.77 and AC = 0.71) and PLR (AUC = 0.71) and AC = 0.68). The most prominent variables in predicting depressive symptoms were EDS, sleep onset, offset and sleep duration.

Conclusões

Sleep variables are essential to predict the occurrence of depressive symptoms in middle-aged and elderly subjects, achieving an accuracy superior to 80%.

Palavras-chave

Machine Learning; comorbidities; sleep; daytime sleepiness; depressive symptoms