COVID-19 Instant Screening test, page-543

Veryvalid points @SteadyEddie. We will wait and see re accuracy and admittedly is a large assumption of the analysis. Albeit, the assumption is not without merit and I believe there is good reason we will surpass the 80% benchmark.

I agree also with you regarding the limitations of using a cough sound topinpoint a virus among the thousands that exist. Obviously a cough will not beable to separate the magnitude of viruses if all lined up in a row.

However, given covid's high prevalence, it is going to be safe toassume/predict that if it sounds like covid, its likely to becovid. The model would theoretically work until a new virus with asimilar signature became prevalent. This would confuse themodel/algorithm rendering it unable to differentiate between the twoviruses and the real world accuracy would decline.

ALLpredictive models rely on the assumption that the past is very similar to thefuture. When the system changes, accuracy declines and you get modeldegradation. My call would be that the covid diagnosis will be accurate enoughin the short term, then degrade over time when the virus becomes less prevalentor is superseded by a new virus. But that doesn't mean the algorithm doesn'thave utility in the short term. If you said that just because it cant predictwhen the system changes that it is useless, that would denigrate the entiremachine learning/AI discipline. To demonstrate this with a driverless carexample, the algorithm is trained with a particular set of road rules. Ifthe road rules happen to change and new stop signs areintroduced, then the model needs to be retrained. It doesn't mean the previousmodel was inaccurate, just that it was only fit for purpose during thatparticular period. The same goes for RAP, this algorithm will only bevalid for a particular period. In a year or two time when covid runs itscourse, we aren't going to be running around getting tested for covid.Therefore the degradation arguably inconsequential.

To argue the model is only accurate when what what you are predicting isprevalent may seem like cheating, but it’s not dissimilar to what the RATis doing. The RAT is not sequencing Covid RNA, it’s analysing a response tocovid by the immune system. The accuracy of these tests will also decline asCovid becomes less prevalent - and even more so if another virus emergesto which the immune system has a similar response. The RAT will share a similarfate where the test accuracy will decline due to confusion.

Currently im building a predictive model which classifies whether a customerwill leave my employer. I will build this model for the current day and willmonitor it. The model I deploy will degrade and there will be a time where Iwill need to retrain the model. I may have new datasets which have morepowerful predictor attributes or some factors which were important, may nolonger important. To relate this back to RAP, when covid is prevalent, expectgreat results, when its not, it will degrade.

Im expecting good enough results (hopefully over 80%) that will have utility inthe short term and then degrade over time. The degradation wont matter much asthe algorithm would have served its short term purpose. If the virus keeps onevolving, maybe RAP need to go back and retrain.