Integrating diabetic retinopathy

[thewati]

Creating the DR module and associated files. Fixes #1595

[mnjowe]

Thanks for creating this PR @thewati. Happy to contribute

[thewati]

Hi @tbhallett and @mnjowe,
Please find the attached plots for DALYs. I excluded deaths since Diabetic Retinopathy and Diabetic Macular Oedema do not cause death. Diebetes Mellitus does. Ready for code review
DR Visual Acuity DALYs.docx

[mnjowe]

Hi @thewati. Thanks for your great work on this PR. Before we start reviewing it, can you first look at the conflicts and see if you can resolve them?

Thanks

[thewati]

Hi @thewati. Thanks for your great work on this PR. Before we start reviewing it, can you first look at the conflicts and see if you can resolve them?

Thanks

Thanks @mnjowe. Resolved. Its now ready

[mnjowe]

Hi @thewati,

Thank you again for the great work on the Diabetic Retinopathy PR. I’m still reviewing the tests and some other remaining files, but I thought it would be helpful to share my initial set of comments now to make tracking and follow-up easier.

Please feel free to question any of the comments or ask for clarification where needed. I’ll continue reviewing the remaining files and will share additional feedback if anything comes up.

Thanks.

[mnjowe]

It looks like you're using probabilities and not rates? if yes please change rate_ to prob_ i.e. rate_onset_to_mild_or_moderate_dr to prob_onset_to_mild_or_moderate_dr and so on

[mnjowe]

for consistency

Suggested change

	"p_laser_success": Parameter(Types.REAL, "Diabetic retinopathy treatment/medication effectiveness"),
	"prob_laser_success": Parameter(Types.REAL, "Diabetic retinopathy treatment/medication effectiveness"),

[mnjowe]

I can't see these parameters in the resourcefile. If you have no plans to use them please delete them

[thewati]

I kept this just in case Cova and Shaffi might want us to include this later. From the previous meetings we've had, the debate on this may not be over. It may also depend on whether they find data to inform regression (which is rare)

[mnjowe]

is this a probability?

[thewati]

Yes, I believe it is.

[mnjowe]

Just wondering, is there any reason we are using "mild_or_moderate" as a category instead of just choosing to go by one?

[thewati]

Yes, I can't quite remember the exact reason. I think its something to do with the treatment of the two being the same if I am not mistaken. Probably something to ask Cova in the Epi call tomorrow

[mnjowe]

I think the 0.5 should be a parameter? I have seen it in a number of places an I'm not sure what it means

[thewati]

So when a person has clinically_significant_macular_oedema (CSMO), the treatment given is either Anti-VEGF or Focal Laser as seen in the attached diagram. Right now, we don't know what proportion of people goes for each. So I just made it so that the chance of going for either is 0.5 (i.e., there's a 50% chance that a patient will have have to go for Anti-VEGF and a 50% chance that a patient will have to go for Focal Laser). Will make it a parameter.

[mnjowe]

Great!
yes, please include it both in the parameters section and in the resource file. It will be updated once you agree on the right value

[mnjowe]

This is a bug.
np.where(will_progress)[0] are positions from 0 to len(will_progress)-1 and you are applying such positions to df.index which is a full population dataframe. What follows is the selection of completely different people i.e. it picks rows at those positions in the full df. This result is tests like the below to fail
assert (df.loc[df.is_alive & ~df.nc_diabetes].dr_status == 'none').all()

Please use the commented block. its safe that way
will_progress_idx = will_progress[will_progress].index

Once you have done the above you may want to set to false squeeze_single_row_output argument in all your linear models. This ensures a single row is returned as a dataframe instead of just a boolean value. This practice also avoid the object has no index errors when doing will_progress_idx = will_progress[will_progress].index

see an example below on how you set squeeze single row to false

lm['onset_mild_or_moderate_dr'].predict(
diabetes_and_alive_nodr, 
squeeze_single_row_output = False,
self.module.rng
)

[mnjowe]

same as my comment above, this is a bug. please use the commented code. you just need to update moderate_to_severe to mildmoderate_to_severe i.e.
moderate_to_severe_idx = mildmoderate_to_severe[mildmoderate_to_severe].index

[mnjowe]

same as my comment above. please use the commented out code. you just have to update it
severe_to_proliferative_idx = severe_to_proliferative[severe_to_proliferative].index

[mnjowe]

Hi @thewati,

Thank you again for the great work on the Diabetic Retinopathy PR. I’m still reviewing the tests and some other remaining files, but I thought it would be helpful to share my initial set of comments now to make tracking and follow-up easier.

Please feel free to question any of the comments or ask for clarification where needed. I’ll continue reviewing the remaining files and will share additional feedback if anything comes up.

Thanks.

[thewati]

Hi @mnjowe,
Right now, the code is using nc_diabetes_date_diagnosis from the cardiometabolicdisorders module to determine how long a person has had diabetes. This duration is used as a risk factor for DR (especially for sight threatening DR). However, nc_diabetes_date_diagnosis only captures individuals who have been diagnosed which is not ideal. We need something like nc_diabetes_date_onset which would capture when an individual actually starts suffering from diabetes. I haven't been able to find such a property in the cardiometabolicdisorders module. Is it a property we would have to create in the cardiometabolicdisorders module or is there an alternative?

[mnjowe]

Hi @mnjowe, Right now, the code is using nc_diabetes_date_diagnosis from the cardiometabolicdisorders module to determine how long a person has had diabetes. This duration is used as a risk factor for DR (especially for sight threatening DR). However, nc_diabetes_date_diagnosis only captures individuals who have been diagnosed which is not ideal. We need something like nc_diabetes_date_onset which would capture when an individual actually starts suffering from diabetes. I haven't been able to find such a property in the cardiometabolicdisorders module. Is it a property we would have to create in the cardiometabolicdisorders module or is there an alternative?

@thewati I hope this has now been addressed in the meeting yesterday

[thewati]

Hi @mnjowe, Right now, the code is using nc_diabetes_date_diagnosis from the cardiometabolicdisorders module to determine how long a person has had diabetes. This duration is used as a risk factor for DR (especially for sight threatening DR). However, nc_diabetes_date_diagnosis only captures individuals who have been diagnosed which is not ideal. We need something like nc_diabetes_date_onset which would capture when an individual actually starts suffering from diabetes. I haven't been able to find such a property in the cardiometabolicdisorders module. Is it a property we would have to create in the cardiometabolicdisorders module or is there an alternative?

@thewati I hope this has now been addressed in the meeting yesterday

Yes.