Lessons learned from the application of machine learning to studies on plant response to radio-frequency.

This paper applies Machine Learning (ML) algorithms to peer-reviewed publications in order to discern whether there are consistent biological impacts of exposure to non-thermal low power radio-frequency electromagnetic fields (RF-EMF). Expanding on previous analysis that identified sensitive plant species, we extracted data from 45 articles published between 1996 and 2016 that included 169 experimental case studies of plant response to RF-EMF. Raw-data from these case studies included six different attributes: frequency, specific absorption rate (SAR), power flux density, electric field strength, exposure time and plant type (species). This dataset has been tested with two different classification algorithms: k-Nearest Neighbor (kNN) and Random Forest (RF). The outputs are estimated using k-fold cross-validation method to identify and compare classifier mean accuracy and computation time. We also developed an optimization technique to distinguish the trade-off between prediction accuracy and computation time based on the classification algorithm. Our analysis illustrates kNN (91.17%) and RF (89.41%) perform similarly in terms of mean accuracy, nonetheless, kNN takes less computation time (3.38 s) to train a model compared to RF (248.12 s). Very strong correlations were observed between SAR and frequency, and SAR with power flux density and electric field strength. Despite the low sample size (169 reported experimental case studies), that limits statistical power, nevertheless, this analysis indicates that ML algorithms applied to bioelectromagnetics literature predict impacts of key plant health parameters from specific RF-EMF exposures. This paper addresses both questions of the methodological importance and relative value of different methods of ML and the specific finding of impacts of RF-EMF on specific measures of plant growth and health. Recognizing the importance of standardizing nomenclature for EMF-RF, we conclude that Machine Learning provides innovative and efficient RF-EMF exposure prediction tools, and we propose future applications in occupational and environmental epidemiology and public health.