Try alternative ways of showing GPT HPO annotation consistency

[bschilder]

Currently shown as stacked bar plots filled by proportion of each response category.

Could instead try:

• confusion matrices (1 per metric)
• sankey plots (1 per metric)

Additional tasks:

• [x] Add N of unique phenotypes per metric
• [x] Bin responses into two categories: (always/often) vs. (sometimes/never)

[bschilder]

HPOExplorer::gpt_annot_check()

• consistency: binned consistency omitting NAs and comparing "sometimes","often","always" vs. "never","rarely"
• consistency_stringent: non-binned consistency where each response type is considered unique.
• true_pos: true positive rate according to ground-truth HPO branches (e.g. neoplasms phenotypes causing cancer)

Text annotations within each column indicate the number of phenotypes were were able to use when computing the respective metric.

[bschilder]

Made this version where the color is the number of phenotypes, and the label shows the exact yaxis value:

[KittyMurphy]

Updated version: