Update risk.py

[drozzy]

Fixing #63

[jinyu-loopmind]

In general, this is a big change on how to compute risk index. It might have big impact on people who are already used to the old scales, especially the risk index is no longer the sum of LBGI and HBGI. Risk index was computed as the sum of LBGI and HBGI for a long time in the original UVa/Padova Simulator.

To minimize the impact, I am thinking about introducing your new risk index computation as an alternative instead of completely replacing them. Maybe a more gentle way to do this is to issue deprecation warning, and advice people to use the new risk index instead. Let me know what you think on this. Thanks!

[drozzy]

Do you know where in UVa/Padova Simulator they define it? I would be curios to see their implementation.

Regarding current use, I can't imagine people have used it because it produces wrong output at extreme values.

For example, here is the current risk if I let the glucose fall indefinitely:

As you can see the function just becomes undefined at some point and gets set to zero.

Here is the proposed risk function for the same trajectory:

Regarding compatibility, the new risk behaves the same for values in the acceptable range.
Here is the current risk:

and here is proposed risk:

For large glucose values, the current risk is not a big problem as it does increase in the desired direction, however it is not bounded and goes to infinity:

the proposed risk caps this at 100:

I welcome more testing and feedback of my proposal, but believe it is especially important for automated optimizers, where such a drop may lead to incorrect learned behavior.

[jxx123]

Do you know where in UVa/Padova Simulator they define it? I would be curios to see their implementation.

Regarding current use, I can't imagine people have used it because it produces wrong output at extreme values.

For example, here is the current risk if I let the glucose fall indefinitely:

As you can see the function just becomes undefined at some point and gets set to zero.

Here is the proposed risk function for the same trajectory:

Regarding compatibility, the new risk behaves the same for values in the acceptable range. Here is the current risk:

and here is proposed risk:

For large glucose values, the current risk is not a big problem as it does increase in the desired direction, however it is not bounded and goes to infinity:

the proposed risk caps this at 100:

I welcome more testing and feedback of my proposal, but believe it is especially important for automated optimizers, where such a drop may lead to incorrect learned behavior.

Wow, thank you so much for the detailed and compelling analysis. Yes, the current risk computation is definitely problematic.

FYI, here is how the original UVa/Padova Simulator implements the risk index (they sum up LBGI and HBGI), where i means i-th patient.

I think I will approve your change anyway since the existing one is definitely wrong. Thank you so much!

[drozzy]

Out of curiosity, where did you find this code? Is it accessible somewhere?

[jxx123]

Out of curiosity, where did you find this code? Is it accessible somewhere?

I had it as in the 2008 UVa/Padova Simulator Academic version. You might be able to request a version from epsilon group https://tegvirginia.com/software/t1dms-2014/, but I heard stories that there was no reply from them.