Run PLSR plot #101
Run PLSR plot #101
Conversation
|
@JacksonLChin This is the code I'm using to run the PLSR plots. Let me know if you find any glaring issues! I based this file mostly by your D8 file. |
| covid_acc = score( | ||
| labels.loc[meta_data.loc[:, "patient_category"] == "COVID-19"], | ||
| probabilities.loc[meta_data.loc[:, "patient_category"] == "COVID-19"], | ||
| labels.loc[meta_data.loc[:, "patient_category"] == "COVID-19"].to_numpy().astype(int), | ||
| probabilities.loc[meta_data.loc[:, "patient_category"] == "COVID-19"].to_numpy(), | ||
| ) |
There was a problem hiding this comment.
To avoid typing headaches later, let's use to_numpy() for either all of the probabilities variables or none of them.
|
@andrewram4287 Looks good! Just one minor change for type handling. |
|
Other thing to note--not sure if this is the same plot you showed in lab meeting yesterday, but the |
Okay that makes sense I can look into what the prediction if we actually combine all the patient samples. Yes, the code creates the same plot I showed in lab meeting.... So that means accuracy does decrease with more PLSR components... |
|
Yup! We may want to swap to the 1-component model for each then. As Dr. Meyer suggested yesterday, I think we could move to a single scores plot with the COVID and non-COVID PLSR components as the x and y axes for interpretation. |
I find it weird though that when you ran the PLSR model before that was not the case... |
|
In the past, we've found the non-COVID model worked best at one component, and the COVID one worked slightly better at two. I think this is mostly in line with what we've seen previously--I'm guessing that the change to how we normalize factors prior to PLSR changed this a little bit. |
|
@JacksonLChin
|
|
@andrewram4287 Cool! Thanks for looking into this. It's interesting that 1-component seems better for accuracy while 2 is better for AUC-ROC--given that they're comparable, though, I think we should continue with the 1-component. |
No description provided.