CHI ’25, Yokohama, Japan

Recent pandemics (SARS, MERS, COVID-19) have demonstrated the important need for rapid non-contact screening tools for respiratory infections. While forehead "spot check" thermometers and thermal imaging have been widely used in recent years, their utility has been limited to identification of people with a fever (T>100.4°F), which misses a large number of infectious cases. In this paper, we present several results that demonstrate how deep learning can be used to greatly enhance thermal imaging tools. In our first experiment, we present results from a CNN ResNet50 transfer learning model trained on thermal images collected from (N=45) patients who had mild respiratory infections (COVID-19, tuberculosis, LRI) who did not present with a fever. When compared with (N=85) healthy controls, the deep learning model demonstrated a median accuracy of AUC=0.80 for a binary classifier (infected vs non-infected). In our second experiment, we tested the ability of thermal imaging with deep learning to distinguish between (N=25) COVID-POSITIVE patients from (N=49) COVID-NEGATIVE patients who had other symptomatic infections. Using a similar deep learning model, we discovered that thermal imaging alone had a limited ability (AUC accuracy = 0.61) to detect COVID-POSITIVE patients from among other infected patients; however, by adding a clinical questionnaire, this accuracy was increased to AUC=0.77, and the overall screening accuracy (thermal+questionnaire) was increased to AUC=0.90. These two experimental results suggest that current COVID-19 screening could be significantly improved by first employing rapid non-contact thermal imaging to quickly detect individuals with some type of respiratory infection, and then following up each positive thermal test with a simple questionnaire to determine if the infection is likely to be COVID-19. To the best of our knowledge, this is the first reported result demonstrating thermal imaging and deep learning among patients with various respiratory infections.

Pulmonary diseases are a leading cause of mortality and disability, but lack of simple low-cost tools to help diagnose and screen for such diseases. In this paper, we present results from a preliminary study exploring the use of thermal imaging as a possible diagnostic tool for several common pulmonary diseases including Asthma, COPD, ILD, Allergic Rhinitis, and Respiratory Infection. As part of a global health study, thermal images of the face were collected from 125 pulmonary disease patients as well as 11 healthy controls. All subjects were evaluated using a full pulmonary function test (PFT) and diagnosed by an experienced chest physician. For each pulmonary disease, we developed a separate naïve 2-layer CNN model as well as a transfer learning CNN model, using a more complex pre-trained ResNet50 model. The naïve CNN models demonstrated an accuracy of AUC = 0.75 for respiratory infection and an AUC=0.76 for COPD, but lacked any significant predictive value for other pulmonary diseases. The transfer learning CNN models demonstrated an accuracy of AUC = 0.82 for respiratory infection and AUC=0.81 for COPD, but exhibited poor performance for other pulmonary diseases. From these results, we conclude that a facial thermal image can be a useful tool to help identify respiratory infections as well as COPD. It is also important to note that none of the patients in our study had a significant fever (T >100.4 °F) that would be predictive of infection, and our CNN models were also able to distinguish Respiratory Infection from other pulmonary diseases including COPD. Given that thermal imaging is a non-contact measurement, such a tool could be of tremendous value in low resource settings or global health.

Society increasingly relies on machine learning models for automated decision making. Yet, efficiency gains from automation have come paired with concern for algorithmic discrimination that can systematize inequality. Recent work has proposed optimal post-processing methods that randomize classification decisions for a fraction of individuals, in order to achieve fairness measures related to parity in errors and calibration. These methods, however, have raised concern due to the information inefficiency, intra-group unfairness, and Pareto sub-optimality they entail. The present work proposes an alternativeactive framework for fair classification, where, in deployment, a decision-maker adaptively acquires information according to the needs of different groups or individuals, towards balancing disparities in classification performance. We propose two such methods, where information collection is adapted to group- and individual-level needs respectively. We show on real-world datasets that these can achieve: 1) calibration and single error parity (e.g.,equal opportunity ); and 2) parity in both false positive and false negative rates (i.e.,equal odds ). Moreover, we show that by leveraging their additional degree of freedom,active approaches can substantially outperform randomization-based classifiers previously considered optimal, while avoiding limitations such as intra-group unfairness.