Research on risk scorecard of sick building syndrome based on machine learning

Sick building syndrome (SBS) negatively affects numerous aspects of human health. To date, few studies have used machine learning to establish SBS risk prediction models. In this study, we investigated 2370 buildings in Chongqing, China and collected 29,785 questionnaires as the basic dataset. Logistic regression was used as the machine learning algorithm. Based on the weight of evidence and information value, features with significant effects were selected, and prediction models were established for mucosal irritation, dermal symptoms, neuro-toxic symptoms, and respiratory symptoms. Through verification based on the accuracy, receiver operating characteristic curve, and Kolmogorov-Smirnov values, the model performed well on the test set. Thus, the prediction models were further transformed into a scorecard that could provide quantitative results for each environment, and the impact of each level of each feature on the results could be observed. The results indicated that when there was an indoor pollution source, the risks of the four SBS models increased. In addition, the indoor thermal environment had a significant impact on SBS. If the indoor thermal environment deviated from the range of thermal comfort, the prevalence of SBS increased, particularly when the indoor and outdoor thermal environments had opposite thermal sensations to the occupants. Finally, we initially established four SBS rating systems based on quantiles, which can provide a reference for establishing a rating system for healthy indoor environments.