Include MTLR model for direct estimation of survival probabilities and supporting feature effects varying over time

[schelhorn]

Feature

Hi there, this is a hugely welcome addition to the R stack for the pharma community! Thanks so much for working on this.

Apart from the issues which are still on your ToDo list, I would like to highlight a particular model which has been shown to work very well in benchmarks as well as in practice, Multi-task Logistic Regression (MTLR).

The approach is originally from NIPS 2011 and already available on CRAN by @haiderstats; it's prediction modes are displayed in the Vignette. The package is actively maintained on Github. There is a very detailed Rmd by the package author that compares different R models for estimating per-patient survival probabilities.

The following resources contain the original paper and subsequent benchmarks and applications; MTLR does very well in the benchmarks:

• 2011 NeurIPS paper
• 2018 ArXiv paper
• 2019 master thesis by the package author
• 2020 JMLR paper
• Benchmark R code for 2020 JMLR paper (maybe a good dataset to include in tidymodels?)
• 2020 BMC application paper
• 2020 AAI application paper
• 2021 PMLR benchmark paper
• 2021 ArXiv paper for SurvBenchmark (including a large benchmarking dataset and comparison of 20 methods)
• SurvBenchmark R package GitHub repo
• SurvBenchmark benchmark code for paper GitHub repo

The methods's main advantages are that it provides per-patient survival probabilities (i.e., it predicts the survival function itself) across a range of times (i.e., it allows for feature effects to vary through time). Also, it works with many different forms of censoring events, which makes it very versatile (although it still does not allow for missing values, unfortunately).

To quote from the R package vignette:

MTLR was specifically designed to give survival probabilities across a range of times for individual observations . This differs from models which produce risk scores (such as those given by Cox proportional hazards), single time probability models (such as the Gail model), and population wide models ( e.g. Kaplan-Meier curves). Producing survival probabilities over a range of times gives a more holistic view of survival to patients and physicians which may be critical in making healthcare decisions.

MTLR can be used for survival data containing right, left, interval, or no censoring. In addition, these types of censoring can be mixed in the same dataset.

(I am not involved with any of these papers nor do I know the authors of the papers or the package maintainer)

[hfrick]

Thank you for the detailed issue! :100:

[schelhorn]

You are welcome.

(I may also add that we are RStudio Enterprise customers with an air-gapped environment; so having {censored} on CRAN and thus RSPM is something we are eagerly awaiting to be able to use the package in our R&D projects. But quality comes first, of course :wink:).