For researchers reading this page, I have to add that accuracies are calculated using the medically proven inter-patient separation scheme for all the tasks: 1 - MIT BIH Arrhythmias database, the AAMI suggested which patients to use for training and which to use for testing. Accuracies that do not respect this separation are not counted towards the state of the art. 2 - AFIB the dataset is public and used in a conference paper competition, similarly to MIT-BIH conference owners suggested patients for training and other patients for testing. 3 - The respiratory disease audio database has a extremely unbalanced and difficult inter-patient separation scheme, original authors provided the patients to train on and the patients used for testing, the accuracy in state of art with the original separation scheme ranges fom 66% to 78% we hit 95% on the same train test split deeply respecting the inter-patient separation scheme.
Once again, in a race for accuracies , interpatient separation scheme is THE GOLD STANDARD when it cames to medical algorithms for CAD (computer aided diagnosis).
OUR ALGORITHMS ALWAYS ANSWER THE QUESTION: IS THIS PERSON' BEATS OR BREATHE AFFECTED BY THIS PATHOLOGY?
DISCUSSION: Algorithms with classical 70/30 split won't be kept into consideration and do not count towards the state of the art. Why? because using a mixed intrapatient separation schemes, beats or breath cycles of paople used for training will be used, partially, also for testing. This is called bleeding of information. Most importantly, the algorithm is not answering the question: is this person affected by this pathology ? but is answering another question: what is the owner of these beats and what is the pathology?
Log in or sign up for Devpost to join the conversation.