The comparison between multiple linear regression and random forest model in predicting environmental noise and its...

Environmental noise exposure is suggested to be linked with various chronic diseases' development, including cardiovascular diseases and cognitive decline. However, traditional acoustic models are not flexible enough in estimating participants' noise exposure with diverse exposure sources in epidemiological studies. The Land-use regression (LUR) model has been recommended as a better alternative in predicting noise levels with less data input. The comparison of multiple linear regression (MLR) and machine learning algorithms in developing the LUR model is required to understand how machine learning algorithms overcome the non-linearity issue of predictor variables that the MLR model cannot solve. Random forest is the more favorable machine learning algorithm in developing LUR model due to its ability in handling outliers and overfitting. Summer and winter measurements for A-weighted equivalent sound pressure levels over 24 h (L_eq,24h) and nighttime (L_night) with the frequency components were conducted at 102 sites during 2019-2020. The noise parameters together with various traffic, population and land-use variables were used to estimate the spatial variability of environmental noise in Hong Kong. Random forest (RF) models performed better than the MLR models with more predictors and in higher leave-one-out-cross-validation R² in L_eq,24h (RF: 0.79, MLR: 0.70). MLR model attained a better performance than RF in fitted R² for all night noise parameters. Additionally, MLR outperformed RF in cross-validation mean absolute error and root-mean-square error in some noise frequencies at night. Vegetation, industrial area and buses were mostly included in both models of high importance. Our findings demonstrate that the RF model can generate credible predictions comparable to those of the MLR model for future epidemiological studies focusing on noise-related health outcomes.