"By now its spectrogram cough signals will have been refined further to higher resolution by inclusion of another six years of collected audio datasets" – There are two problems with that statement.
Firstly an audio sample is only valuable for algorithm training if it is accompanied by an accurate, independent diagnosis that the algorithm can be trained or tested against. Without that the sample is useless for refining the algorithm. And we have learned that accurate diagnosis is extraordinarily difficult to achieve until well after the patient presents, so it is necessary to capture follow-up data that isn't normally available in a clinical setting. In many cases the diagnosis is never known; the patient simply gets better and there is no definitive answer.
Secondly Resapp have pointed out many times that beyond a certain level increasing the size of the training dataset produces very little improvement in predictive accuracy. Part of the reason for this is that as the accuracy increases there is less room for improvement. It's easier to go from 70% to 90% accuracy than it is to go from 90% to 95% simply because there are fewer wrong results to fix. The other part of the reason is the difficulty of getting an accurate diagnosis in the first place. If the results in the training data are only 80% accurate then you can't possibly train an algorithm to be better than 80% accurate.
- Forums
- ASX - By Stock
- RAP
- Ann: Publication on challenges of diagnosing respiratory disease
Ann: Publication on challenges of diagnosing respiratory disease, page-45
Currently unlisted. Proposed listing date: 4 SEPTEMBER 2024 #