Goal: Thermography-based infection screening at international airports plays an important role in the prevention of pandemics. However, studies show that thermography suffers from low sensitivity and specificity. To achieve higher screening accuracy, we developed a screening system based on the acquisition of multiple vital-signs. This multimodal approach increases accuracy, but introduces the need for sophisticated classification methods. This paper presents a comprehensive analysis of the multimodal approach to infection screening from a machine learning perspective.
Methods: We conduct an empirical study applying six classification algorithms to measurements from the multimodal screening system and comparing their performance among each other, as well as to the performance of thermography. In addition, we provide an information theoretic view on the use of multiple vital-signs for infection screening. The classification methods are tested using the same clinical data, which has been analyzed in our previous study using linear discriminant analysis. A total of 92 subjects were recruited for influenza screening using the system, consisting of 57 inpatients diagnosed to have seasonal influenza and 35 healthy controls.
Results: Our study revealed that the multimodal screening system reduces the misclassification rate by more than 50% compared to thermography. At the same time, none of the multimodal classifiers needed more than 6 ms for classification, which is negligible for practical purposes.
Conclusion: Among the tested classifiers k-nearest neighbors, support vector machine and quadratic discriminant analysis achieved the highest cross-validated sensitivity score of 93%.
Significance: Multimodal infection screening might be able to address the shortcomings of thermography.