predict w/MTL vs. static-feat included, then based on results, visualize based on metadata labels + performance enhancement

[nealgravindra]

• [x] need to be able to put metadata in the y_file, like patient ID, etc. so should instead remove the label part, split based off GA and all metadata, and then rename series2modeltable to df2y and add the label option there, so that can always go back and forth between metadata and train/test/val splits (metadata-containing one can be Y_train, one for model can be y_train)
• [ ] define activity by comparison with NHANES data... define sleep quality from actigraphy as well, then use these extracted features to correlate or use as feature in T2LD to examine its relation ("activity score" and "sleep score") by downloading NHANES and starting with sleep detection from pyactigraphy
• [x] update class balanced metric after a bit more research and try AU-PRC_model / AU-PRC_random or AU-PRC_model - AU-PRC_random (1/AU-PRC_random or class ratio, e.g., if balanced, PRC of random clf = class ratio at precision, AU-PRC = 0.5, so model - random at best is 0.5, now 0.5 / 1/0.5 = 0.5 / 1/ 1/2 = 20.5 = 1)
• [x] fix the training scheme and re-run with the cleaned up metric
• [x] consider running on the one-hot encoded matrix instead, as it may be interesting to consistently predict the uncommon class in a OvR scheme (e.g., can we predict the death of the baby relative to class no and class missing?) and also include composities, e.g., edinburgh score
• [ ] run with the timeseries thingy to also use the time (read paper to see if it's useful)
• [ ] hyperparam exploration

[nealgravindra]

• [ ] make a graph for each target label and cluster, annotating modules... underlying structure of graph should represent relationships such as correlation between label and activity data
• [ ] triplet net (try bidirectional lstm for embedding, take only last hidden via: https://stackoverflow.com/questions/55907234/pytorch-lstm-grad-only-on-last-output) and double-check for leftward-lstm net

[nealgravindra]

Need to review the classification scheme but this is on pause because we don't want to run hundreds of models before optimizing the main model architecture a bit more. As in, optimize for GA prediction, then apply to metadata variables

[nealgravindra]

May be worthwhile to also try class imbalance adjustments to the loss function, such as: https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/class_balanced_loss.py

[nealgravindra]

Cleaning up the figures:

• [x] viz of time-series, and cohort table, stratified by ptb (or any var)
• [ ] add more md vars from the larger raw table as well as the composite scores (edinburgh and sleep qual)
• [ ] hyperparam search training w/fixed epoch, simple storage, and save the eval for later after getting model output
• [ ] data augmentation scheme (or at least noise injection) for training

[nealgravindra]

Other enhancements:

• [ ] add aggregated metrics to metadata, then pull that metadata into for labeling with a cleaned up data pipeline. More efficient data storage of actigraphy data to allow for scalability (e.g., not saving trainer data each time but only indices)
• [ ] eval (and train) of model to output to dict and pkl instead of csv to avoid annoyances with pandas ast.literal_eval