Predicting the Occurrence of Respiratory Diseases Based on Campus Indoor Air Quality

Air quality is known to be strongly correlated with respiratory diseases. Indoor air quality considerably affects human health, especially in spaces such as classrooms, where students gather and interact for long periods. Most schools are located in relatively old buildings, where suitable ventilation systems are difficult to implement. The consequent lack of a standard ventilation rate increases the risk of cluster infections in classrooms. Accordingly, this article proposes a classroom respiratory disease occurrence prediction method based on indoor air-quality data (CROP-IAQ). Early warnings provided by CROP-IAQ will enable authorities to implement measures such as ventilation and isolation that reduce the risk of cluster infections in school campuses. Data on indoor temperature, relative humidity, particulate matter (PM) concentrations (PM1.0, PM2.5, and PM10, referring to the concentrations of particles with diameters of , , and micrometer, respectively), carbon dioxide concentration, total volatile organic compound concentration, and luminosity in classrooms were collected using a MAPS V6.0 airbox. The air-quality data corresponding to potential cluster infections were identified from the aforementioned data and records of student epidemic prevention leaves in each class. Because most of the collected air-quality data did not correspond to potential cluster infections (that is, the dataset was imbalanced), synthetic data samples were generated using a synthetic minority oversampling technique. Four neural network models were constructed for predicting the possibility of disease occurrence and alerting authorities to classrooms at the risk of cluster infections: a convolutional neural network model, the inception model, a residual network model, and a residual network with external transformations model. The predictive capabilities of these models were only slightly improved after implementing a squeeze-and-excitation (SE) module. Experimental results indicated that the inception model with the SE module achieved the best results among the four models, with an F1 score and sensitivity of 0.72 and 0.76, respectively.