Vigilância da atividade física: análise de tendências e perspectivas com inteligência artificial e estudos longitudinais

Abstract

Traditional surveillance based on aggregate prevalence is insufficient to guide public policies. Therefore, this thesis seeks to analyse the contemporary challenges of physical activity (PA) surveillance through the use of diverse methods, such as time trend analyses, artificial intelligence for small-area prevalence estimation, and longitudinal analyses, contributing to discussions about territorial inequalities and local factors influencing PA practice. The first study describes the general situation of PA indicators in Brazil from 2006 to 2021; the second estimates the prevalence of leisure-time physical activity (LTPA) in small areas in the city of Belo Horizonte from 2009 to 2013 and from 2014 to 2018, observing inequalities between areas and between periods according to the Health Vulnerability Index (HVI); the third one analyses environmental factors associated with different PA behavior trajectories among adults aged 45 to 60 years over a 10-year follow-up period in the city of Brisbane, Australia. The database for the first and second studies was the Surveillance System for Risk and Protective Factors for Chronic Diseases by Telephone Survey in Brazil (Vigitel). For time trend analysis, Prais-Winsten regression was applied to several PA indicators, presented for the total population and by sex, age, education level, and region. The main findings showed that, for the total population, the trends in PA were positive and significant (β = 0.614; p = 0.010), as well as for total screen time (β = 1.319; p = 0.001), demonstrating an increasing trend in the prevalence of these two indicators in the Brazilian population during the analysed period. Subsequently, spatial estimates were performed in small areas for the city of Belo Horizonte, based on an artificial intelligence (AI) technique called k-means, which allowed for a more cohesive and homogeneous regrouping of the city's census tracts according to territorial characteristics described by the HVI. After the formation of subgroups or clusters derived from the HVI groups, LTPA prevalence estimates were calculated using the post-stratification weights of the Vigitel sample. Differences in prevalence between periods and between clusters were calculated in percentage points (p.p.). In the first period, LTPA prevalence ranged from 23.70% in a cluster derived from the very high-risk stratum to 45.55% in a cluster derived from the low-risk stratum (difference of 21.9 p.p.). In the second period, from 2014 to 2018, prevalences also ranged from 31.82% to 52.81% in the highest and lowest vulnerability strata, respectively (difference of 21.0 p.p.). Estimation in small areas allowed for the observation of patterns of inequality between areas. It was observed that, despite the increase in the prevalence of LTPA in the clusters between the two periods, the difference between them did not decrease from the first to the second period analysed, especially between the groups derived from the low HVI risk compared to those of medium, high, and very high risk. Finally, the third study used a longitudinal database from Australia, HABITAT, or "How Brisbane Areas Influence Health and Physical Activity," and investigated the relationship between environmental determinants and LTPA trajectories, that is, the factors related to belonging to one or another behavioral trajectory for PA over time. The definition of trajectories for the individuals in the sample was performed using the Group-Based Trajectory Modeling (GBTM) method. More details of this method are in the corresponding section. For association analysis, a robust multinomial logistic regression model was applied. All environmental variables (analysed for small areas or neighborhoods) were categorized into tertiles, including: bike path length, area of park, residential density, street connectivity, land use mix, and street light count per neighborhood. The model was adjusted for sex, age, education, income, living arrangements, time living at current address, and neighborhood socioeconomic classification. Statistically significant associations were observed between the middle tertile of bike path length and a lower probability of belonging to the low-level of LTPA over time [OR = 0.55 (95% CI: 0.36 – 0.84)]; and between the average tertile of street connectivity and a lower probability of belonging to low-level [OR = 0.67 (0.51 – 0.87)] and increasing-level [OR = 0.76 (95% CI: 0.61 – 0.94)] LTPA trajectories. Thus, some environmental characteristics were associated with better patterns of LTPA during the transition from adulthood to old age. In short, this thesis started from cross-sectional national analyses and moved on to more granular and longitudinal territorial analyses, proposing a more sensitive surveillance to territorial inequalities. Local surveillance allows for the formulation of more assertive public policies.