soda-inria / hazardous

Competing Risks and Survival Analysis
https://soda-inria.github.io/hazardous/
MIT License
56 stars 12 forks source link

Add the accuracy in time evaluation metric #74

Open Vincent-Maladiere opened 2 months ago

Vincent-Maladiere commented 2 months ago

As presented in the paper, for competing risks, the accuracy in time answers the question "Does my model accurately predict which event will occur first to individuals at a given horizon?".

It differentiates from the C-index which answers an alternative question: "Does my model accurately rank individuals by their incidence probability for a specific event of interest?

A model with good accuracy in time is useful for the physicist to determine where to focus the treatment for a given individual. A model with a good C-index on the other hand is useful to allocate treatment resources to individuals that need it the most.

In the paper, we introduced the accuracy in time as:

Screenshot 2024-08-23 at 16 12 58

where

Screenshot 2024-08-23 at 16 15 15

is the most probable event at time horizon $\zeta$,

Screenshot 2024-08-23 at 16 32 07

is the event (or censoring) observed at $\zeta$, and $n_{nc}$ represents the number of individuals uncensored at $\zeta$.