A Multi-Pollutant and Meteorological Analysis of Cardiorespiratory Mortality among the Elderly in São Paulo, Brazil—An Artificial Neural Networks Approach

Abstract

Traditionally, studies that associate air pollution with health effects relate individual pollutants to outcomes such as mortality or hospital admissions. However, models capable of analyzing the effects resulting from the atmosphere mixture are demanded. In this study, multilayer perceptron neural networks were evaluated to associate PM10, NO2, and SO2 concentrations, temperature, wind speed, and relative air humidity with cardiorespiratory mortality among the elderly in São Paulo, Brazil. Daily data from 2007 to 2019 were considered and different numbers of neurons on the hidden layer, algorithms, and a combination of activation functions were tested. The best-fitted artificial neural network (ANN) resulted in a MAPE equal to 13.46%. When individual season data were analyzed, the MAPE decreased to 11%. The most influential variables in cardiorespiratory mortality among the elderly were PM10 and NO2 concentrations. The relative humidity variable is more important during the dry season, and temperature is more important during the rainy season. The models were not subjected to the multicollinearity issue as with classical regression models. The use of ANNs to relate air quality to health outcomes is still very incipient, and this work highlights that it is a powerful tool that should be further explored.

Exposomics and Cardiovascular Diseases: A Scoping Review of Machine Learning Approaches

“…Proceeding with the remaining 30% of pertinent literature that does not conduct performance comparison whatsoever, a part of research opted for a non-linear algorithm, 42,43,56,57 while another part has focused on an ensemble algorithm of their selection. 50,51,58,59 As it becomes obvious from Figure 4 ensemble algorithms are the most popular choice encountered in pertinent literature, 35,[39][40][41][44][45][46][47][48][49][50][51][52][53][54][55]58,[60][61][62][63][66][67][68] closely followed by non-linear algorithms.…”

“…50,51,58,59 As it becomes obvious from Figure 4 ensemble algorithms are the most popular choice encountered in pertinent literature, 35,[39][40][41][44][45][46][47][48][49][50][51][52][53][54][55]58,[60][61][62][63][66][67][68] closely followed by non-linear algorithms. 35,[41][42][43][44][45][46][47][48][49]54,59,60,62,[64][65][66][67][68][69][70][71]...…”

“…The most widely used environmental parameters include air pollutant concentrations and meteorological parameters. 43,45,46,53,64,69,72 Additionally but less often, parameters reflecting noise pollution, 39,51 green spaces 39,46 and traffic proximity/volume 41,66 are also included in the exploited feature space. In very few cases, variables reflecting daily or work exposures such as number of smokers at home, 11,39,52 biohazardous materials 11 and heavy metals concentration based on human scalp hair analysis 49 , blood or urine samples are exploited 68 .…”

“…The authors conclude that interventions to reduce premature cardiovascular mortality should be targeted to geographic areas with high-risk phenotypes of premature cardiovascular mortality. Leirião et al have generated forecasts of cardiorespiratory mortality in the elderly aiming at providing decision-makers with a powerful tool for the evaluation of environmental public health-related policies [43]. Martin-Morales et al predicted CVD mortality with a focus on identifying associated risk factors from a pool of significant nutritional variables [44].…”

“…Proceeding with the remaining 30% of pertinent literature that does not conduct performance comparison whatsoever, a part of research opted for a non-linear algorithm [42], [43], [56], [57], while another part has focused on an ensemble algorithm of their selection [50], [51], [58], [59].…”