Question about 1D-CNN ECG models #24
Description
Hi there,
I am researching the current state of the art for neuromorphic applications in ECG. The Aikda is one of the few chips that I can actually get for empirical evaluation. I was very excited to see that you already have some models for anomaly detection and classification.
However; there is no information regarding the input or output structure of the models. I gather from #22 that you do not wish to include the training or evaluation pipelines, which I understand. Though I was hoping that you would be able to shed some light on how these models can be used.
For example, I can see from netron.app that the ECG Classifier (ECG_conv_float32.h5) takes an input of 50x5x1 and outputs 4 classes. It also uses Conv2D with 5x5 Kernels throughout; which is strange given it is labeled as a 1D-CNN.
I have the MIT-BIH dataset that this classifier was trained with and the dimensionality does not match. Each recording is of 2 leads; and consist of a very large set of readings (around 650k data points). How this being fragmented and reshaped into 50x5 is beyond me.
Likewise; MIT-BIH is typically classified with 5 classes; Normal, Supraventricular Ectopic, Ventricular Ectopic, Fusion, and Unknown. I expect that this model is simply omitting the unknown class; but I can otherwise only guess at which classes the 4 outputs of this model correspond to.
I would be very interested to hear anything you could tell me about using these models.