以校園室內空氣品質預測呼吸道疾病的發生

Abstract

人們長時間待在室內環境中，室內空氣品質對人體的影響更直接，已經有相當多的研究證實空氣品質與呼吸道疾病之間存在著關聯性，尤其是學校教室這類長時間人群聚集且互動頻繁的密閉空間，對於學生的健康以及學習力都會受到影響。在高度工業化發展的環境下，容易有空氣品質不佳的問題產生，而大多數的學校屬於較老舊的建築，很難有適當的通風設備讓教室維持在標準通風率下，使得校園群聚感染的風險上升。本研究提出了使用神經網路以室內空氣品質預測教室呼吸道疾病的發生(CROP-IAQ)的方法。本研究使用MAPS V6.0空氣盒子佈建於國立臺灣師範大學林口校區教室內，蒐集室內的溫濕度、PM10、PM2.5、PM1.0、二氧化碳、TVOC和光度等數據，以各班級學生防疫假的日期標記可能有疾病影響的空氣品質資料，透過神經網路模型建構的分類器，預測疾病發生的可能性，並提供示警。CROP-IAQ的示警可以讓校園及早針對高風險教室做通風或隔離等措施，以減少校園群聚感染發生的風險。研究中所使用的數據加入了以SMOTE生成的相似資料，做為模型訓練時的輸入資料，模型共有四種包含Basic CNN model、Inception model、ResNet model及ResNeXt model，四種模型也再加入Squeeze-and-Excitation module提高模型預測能力。根據實驗結果，最終是Inception-SE model有最好的正確率，F1score為0.72、靈敏度為0.76。People staying indoors for long periods are more directly affected by indoor air quality. Numerous studies have already confirmed the correlation between air quality and respiratory diseases. Particularly in the case of spaces like school classrooms, where people gather for long durations and interact frequently, students' health is significantly impacted. Especially in most modernized environments, the issue of poor air quality is prone to arise. In addition, most schools are in relatively old buildings; it is challenging to implement adequate ventilation systems to maintain classrooms at the standard ventilation rate, thereby increasing the risk of cluster infections in classrooms. In this paper, we proposed a Classroom Respiratory disease Occurrence Prediction method with Indoor Air Quality data (CROP-IAQ) using Neural Networks.This study used MAPS V6.0 Airbox deployed in classrooms to collect data on indoor temperature, relative humidity, PM1.0, PM2.5, PM10, CO2, TVOC, and luminosity. The data, combined with the records of student epidemic prevention leave in each class, were used to mark the air quality data potentially affected by diseases. A classifier is constructed using a neural network model to predict the possibility of disease occurrence and to provide alerts for classrooms with a risk of clustered infection. CROP-IAQ can reduce the risk of clustered infections on campus with early warnings for authorities to implement measures such as ventilation and isolation.In addition to the data collected, the Synthetic Minority Over-sampling TEchnique (SMOTE) method was employed to generate similar data for input into the models during training. The models included the Basic Convolutional Neural Networks (CNN) model, the Inception model, the Residual Network (ResNet) model, and the ResNeXt model. These four models were enhanced with a Squeeze-and-Excitation (SE) module to improve the predictive capabilities. Based on the experimental results, the Inception model with the SE module achieved the highest accuracy, with an F1-score of 0.72 and a Sensitivity of 0.76.