Instead of simply measuring the health state as the count of binary health deficits as in a Frailty Index, we estimate damage and repair processes. Paper available at https://elifesciences.org/articles/77632.
A python package to extract counts of damaged and repaired binary health variables between longitudinal time-points (as in the above schematic) is available in damagerepair-counts/. It is installed with setup.py.
To use, provide input data as a pandas dataframe "data_raw", a list of deficits "deficits", and specify strings for individual id_column, time varaible time_column, and optiaonal death information.
from damagerepair_counting import counting
data = counting(data_raw, deficits, id_column, time_column,
deaths = None, death_column = None, status_column = None)The output of this approach is a pandas dataframe with columns "repair.count" and "damage.count" representing the number of new deficits repaired or damage between measurement times (with time-interval in column "delta.t"). This data is then used as input for damage/repair count models.
Joint models of survival and longitudinal repair and damage counts written in Stan are available in models/. For each of the datasets, these are fit by running fit_mouse_1.r, fit_mouse_2.r, and fit_mouse_3.r. A longitudinal model for human repair and damage counts is fit by running fit_human.r, which takes a command line argument for the chain ID.
Fitting the mouse models takes a few hours each (with all chains running parallel in time). Fitting the human data takes 12-15 hours using 40 CPU cores for a single chain.
Bayesian spline models for interval censored survival data written in Stan are also available in models/. These are fit in plotting_code/mouse_timescales.r and plotting_code/human_timescales.r.
R Packages used:
R 4.1.0
abind 1.4-5
caret 6.0-92
cmdstanr 0.4.0
cowplot 1.1.1
dplyr 1.0.9
ggplot2 3.3.6
ggsci 2.9
ggfortify 0.4.14
gridExtra 2.3
glrt 2.0
latex2exp 0.9.4
pammtools 0.5.8
posterior 1.2.1
rstan 2.21.5
splines2 0.4.5
statmod 1.4.36
survival 3.2-11
survminer 0.4.9
tidybayes 3.0.2
tidyr 1.2.0
Stan version 2.26.0 was used.
python packages used
python 3.8.10
pyreadstat 1.1.0
pandas 1.4.2
numpy 1.22.4
matplotlib 3.5.2
Three different mice datasets are used, from https://doi.org/10.1093/gerona/gly219, https://doi.org/10.1093/gerona/glab297, and https://doi.org/10.1038/s41467-020-18446-0. Dataset 3 is freely available from https://github.com/SinclairLab/frailty as the "Frailty data from Schultz_Kane_et_al.xlsx" file. The authors of the other papers should be contacted to request access to mouse datasets 1 and 2.
Human data from the English Longitudinal Study of Aging are used, which can be accessed here https://www.elsa-project.ac.uk/accessing-elsa-data by registering. We use waves 1-9 from edition 31.
The code used to clean and pre-process these datasets is available in clean_datasets/.
Running setup_demo.sh in demo/ will download and prepare mouse dataset 3 and save it in the datasets/ folder. The model for this dataset can then be fit by running fit_mouse_3.r. Fitting this model may take a couple hours.
Please cite if this is useful
@article {Farrell2022elife,
article_type = {journal},
title = {Measurements of damage and repair of binary health attributes in aging mice and humans reveal that robustness and resilience decrease with age, operate over broad timescales, and are affected differently by interventions},
author = {Farrell, Spencer and Kane, Alice E and Bisset, Elise and Howlett, Susan E and Rutenberg, Andrew David},
volume = 11,
year = 2022,
month = {nov},
pages = {e77632},
citation = {eLife 2022;11:e77632},
doi = {10.7554/eLife.77632},
url = {https://doi.org/10.7554/eLife.77632},
journal = {eLife},
issn = {2050-084X},
publisher = {eLife Sciences Publications, Ltd},
}