Last Experiment: Identifying Unknown Patient Status

[annaritz]

The Goal: show how this can be useful for unknown-status patients. One example: take patients with little data (or unknown end-status), predict drugs, and then see if (a) those drugs were administered or (b) if those drugs weren't.

• Look into datafiles & formatting.
• Go into big original file. (a) Find a patient & format data, run code, etc. (b) find multiple patients to do this & show results in some table.
• Just want code to stop at the prediction (instead doing the validation).

[annaritz]

Ran method on one patient w/ unknown status -- predicted same drug that was administered!
Next up, do this for multiple patients - first, figure out about how many patients have unknown status. Keep track of (drug & confidence score) for every patient. We can then calculate:

• The % of patients where predicted drug was administered
• The confidence score of the predictions where the administered & predicted drugs were same.

[annaritz]

One idea: have ds_pipeline() function STOP after fitting the model and return model. Then, you can have a test_random_patient() function, and a test_specific_patient() function. In test_specific_patient(), you want to pass in patientID.

[annaritz]

Getting Results into the paper:

• [ ] Send Anna the list of patients that didn't recover & the MAF files (compress the directory & send it through Google Drive). Anna will calculate sex & age stats. Anna will summarize the "unknown patients"
• [ ] Write up the fact that ALL patients had predicted drug that was administered. [ How common is the drug that's predicted? how many other patients had taken it? Just look up some info about that specific drug.]
• [ ] Calculate a histogram of the prediction confidence scores. Even though all patients had this drug predicted, predicted a middling confidence levels.