Paper: https://dl.acm.org/doi/abs/10.1145/3381014 https://dl.acm.org/doi/abs/10.1145/3381014 From the abstract: "We conducted an IRB-approved 7 month-long stud

Paper: https://dl.acm.org/doi/abs/10.1145/3381014 https://dl.acm.org/doi/abs/10.1145/3381014 From the abstract: "We conducted an IRB-approved 7 month-long study from December 10, 2018 to July 12, 2019 where we deployed FluSense in four public waiting areas within the hospital of a large university. During this period, the FluSense platform collected and analyzed more than 350,000 waiting room thermal images and 21 million non-speech audio samples from the hospital waiting areas." Huh, my ethics board would ask me if I'm okay if I proposed that. Field work with people that are likely in a vulnerable position, to further some "AI" proposal of unproven effectiveness, great. Also from the abstract: "FluSense can accurately predict daily patient counts with a Pearson correlation coefficient of 0.95." Odd measure, buzzwordy press release article, let's look at the paper. From their contributions (page 3): "Thermal imaging-based crowd density estimates were found to have a strong correlation with the reference daily waiting room patient counts from hospital records (with a Pearson correlation coefficient of 0.95)" You can count people on thermal images? It's slightly more interesting of an experiment than that but okay. Not really surprising result. So that's the number from the abstract. Might make sense if daily patient count is roughly normally distributed. The distribution of patients during a pandemic is likely to be completely different than that of "regular" times. Luckily, the article doesn't mention Corona / covid, that seems to be added fluff for the psychology today version. The term "pandemic" only appears in their references. Though it might have made for far more interesting reading if they waited a month and investigated how this behaviour changed in the onset of the pandemic. "Our results suggest that multiple and complementary sets of FluSense signals exhibited strong correlations with both ILI counts and laboratory-confirmed influenza infections on the same campus." Multiple, okay (table 7 looks like they tried every easily accessible method for ensembles in sklearn). So how about the cough model / audio stuff the article boasts? "This real-world cough model shows resilience to complex noises and diverse settings and achieves an F1 score of 88%" Fair enough, sounds reasonable for a CNN pipeline based on 1 second samples. Their model seems to really have problems with door screeches. Their privacy assumptions appear somewhat lacking. The thing is a raspberry pi with physical access to basically everybody in the vicinity, making everything else almost irrelevant. They say they use "two-stage encryption" to store non-speech samples to disk, the raspberry pi they are using contains no modules that would make this secure to physical attacks. Section 3.1 mentions that they have a "high-fidelity" classifier to discard speech snippets in this sensitive area, that appears to be 93% for speech detection with "augmentation" (background noise etc.). In a waiting room I would not want a 7% failure rate here. They mention HIPAA on page 13 and say that their real system, from the rather small sample of 2500 snippets only contained 79 unintelligible speech samples, not sure why that would be okay but hey, I can put my own SD card in there anyway so who cares right? Their cough model seems to work well in the lab setting (public datasets suited for the purpose). Their ground truth for real world validation appears to be 2500 one-second long samples they annotated manually (5.4), evaluation is against a human annotator (classes: cough/ no cough). Table 4 seems to be where the 88% f1 score come from, it says "after transfer learning". Page 11 says "We used approximately half of our labeled data for the transfer-learning.", leaving 1250 test samples? This real world model is then evaluated on 200 samples from days under different conditions (location, how crowded the day was), achieving F1 scores of .8 (less crowded days), .88 (crowded days). For covid,, my country's situational report says that only 53% of clinical cases have cough symptoms in the first place. That seems to leave much room for any real world use to "detect pandemic indicators". The paper itself is not uninteresting but it seems somewhat... unclear on a few of the assumptions made throughout, which imho puts some doubt on the efficacy of the final output of their overall model in a real setting. The link to covid this PT article feels to need to put in there would be highly questionable but luckily the authors refrain from making that connection.