In the project I am studying I have a set of records (each record is a 1D signal) with variable length. For normalizing length and data augmentation, I am segmenting them in fixed length samples so I can input them in my CNN. I will then obtain sample wise classification. I would also like to obtain record wise classification by combining the predictions from each record's samples.
I am thinking of pursuing an approach similar to ensemble voting in my binary classification, so either by averaging the probabilities predicted (and if the mean exceeds 0.5 the record will be labeled as 1, and 0 otherwise), or using a hard voting strategy in which I classify each record (using a 0.5 threshold) and I get the majority, or even by determining the maximum and see if it's above 0.5. My problem is related to the medical area and the response will be if the record is normal (0) or abnormal (1).
What do you think it's the best approach? I haven't found papers that apply these strategies, usually when they divide into samples, they don't focus on record wise classification. This is similar to ensemble voting but instead of different classifiers, it's the same classifier for different samples.