Indoor air quality pollutants predicting approach using unified labelling process-based multi-criteria decision making and machine learning techniques

Indoor air quality (IAQ) refers to the conditions found within buildings that can impact respiratory health. Good IAQ conditions for hospital facilities are essential, especially for patients and medical staff. Recently, several concerns have been outlined and require an urgent solution in identifying IAQ pollutants and related thresholds and ways to provide a knowledge-based method for labelling pollution levels. To this end, a systematic review should be conducted first to construct new taxonomy research on internet of things-based IAQ sensory technology in hospital facilities to identify a research gap. Thus, the present study aims to develop an IAQ methodology that includes the recommended nine pollutants for hospitals and facilities: Carbon monoxide, Carbon dioxide, Nitrogen Dioxide, Ozone, Formaldehyde, Volatile organic compound, particulate matter (PM) and air humidity and temperature. The developed methodology utilised actual and simulated IAQ pollutant datasets to predict the pollution levels within hospital facilities in three distinct phases. In the first phase, two IAQ datasets (real and large-scale simulated datasets) are identified. The second phase includes the following: First is utilising the Interval type 2 trapezoidal-fuzzy weighted with zero inconsistency (IT2TR-FWZIC) method from the Multi-Criteria Decision Making theory for providing the required weights to the nine pollutants. Second is developing a new method, the Unified Process for Labelling Pollutants Dataset (UPLPD), consisting of six processes based on the IT2TR-FWZIC method. The UPLPD can classify the pollution levels into four levels and assign the required labels within the two datasets. Third is applying the labelled datasets to the developed machine learning model using eight algorithms. The third phase includes the model evaluation using five metrics in terms of accuracy, Area under the Curve, F1-score, precision and recall. For the actual dataset, the best three algorithms' results are Support Vector Machine, Logistic Regression and Decision Tree (DT), which achieved the highest accuracy of 99.813, 99.259 and 98.182%, respectively, with performance metrics. The simulated dataset, the Random Forest, DT and AdaBoost achieved the highest accuracy of 90.094, 88.964 and 87.735%, respectively, with performance metrics. The results satisfied the challenges and overcame the issues, and experimental results confirmed the efficacy of the predictive model.