efi_familypractice.Rmd

---
title: "The patient behind the illness: electronic frailty index as an early 
        indicator of healthcare utilization."
author: 'Alex Bokov ^1,✉^, Sara Espinoza ^1^, Chandana Tripathy ^1^,
  ..., ..., ..., and Kathleen R. Stevens ^1^'
css: "production.css"
abstract: 'We have found that the electronic frailty index (EFI), a risk
  score developed using the Rockwood deficit-accumulation framework, is
  a strong predictor of patient safety events without relying on any predictors
  other than diagnoses, vital signs, and laboratory results from the electronic
  health record. Previous studies using electronic frailty indexes on claims 
  data or electronic health records focused on older patients, but we found that
  for most of the patient safety outcomes we examined EFI is also predictive 
  for patients in the 45-65 age range and even in the 18-45 age range.'
documentclass: article
clean: false
bibliography: efi_paper.bib
csl: harvard-cite-them-right.csl
output:
  bookdown::html_document2:
    fig_caption: true
    self_contained: true
    number_sections: false
    keep_md: true
    clean: false
  bookdown::word_document2:
    reference_docx: efi_template.docx
    fig_caption: true
    self_contained: true
    number_sections: false
---

```{r load_deps, echo=FALSE, message=FALSE, warning=FALSE,results='hide'}
.projpackages <- c('GGally','pander','dplyr','ggplot2','data.table'
                   ,'survival','broom','forcats','table1','english');
.deps <- c( 'analysis.R' );
.debug <- 0;
.junk<-capture.output(source('./scripts/global.R',chdir=TRUE,echo=FALSE
                             ,local=TRUE));
# Set some formatting options for this document
panderOptions('table.alignment.default','right');
panderOptions('table.alignment.rownames','right');
panderOptions('table.split.table',Inf);
panderOptions('p.wrap','');
panderOptions('p.copula',', and ');
panderOptions('graph.fontfamily','serif');
# theme_get() gives default theme
theme_set(theme_bw(base_family = 'serif',base_size=14) +
            theme(strip.background = element_rect(fill=NA,color=NA)
                  ,strip.text = element_text(size=15)));
knitr::opts_chunk$set(echo=.debug>0, warning=.debug>0, message=.debug>0);

# detect current output format
.outfmt <- knitr::opts_knit$get('rmarkdown.pandoc.to');
message('.outfmt = ',.outfmt);
if(is.null(.outfmt)){
  .outfmt <- if(knitr::is_html_output()) 'html' else {
    if(knitr::is_latex_output()) 'latex' else {
      if(interactive()) 'html' else {
        'unknown';
    }}}};
message('.outfmt = ',.outfmt);

.currentscript <- current_scriptname('efi_familypractice.Rmd');
.outcomenames <- v(c_fampractice);
snip_outcomes <- rname(.outcomenames);

# limit the data to items of interest
tb2 <- subset(tb2,outcome %in% .outcomenames & predictor %in%
                c('Frailty','Frailty, age:18-45','Frailty, age:45-65'
                  ,'Frailty, age:65+'));
tb3 <- subset(tb3,Predictor %in% c('Frailty','Frailty, age:18-45'
                                   ,'Frailty, age:45-65','Frailty, age:65+') &
                Outcome %in% .outcomenames);
fits <- fits[.outcomenames];
# snippets ----
snip_resultsfold <- unique(resultsfold00(tb2));
snip_dat <- c(pat=length(unique(dat03$patient_num)),vis=nrow(dat03)) %>% 
  format(big.mark=',');
```


# Introduction

The primary diagnosis does not exist in a vacuum; it can interact in complex
ways with the patient's overall physiological state. A promising approach are 
frailty scores that use the deficit accumulation methodology 
[@rockwood07; @mitnitski01a]. Such indexes has been used to predict disability, 
injurious falls and mortality [@pajewski19], and long-term care admissions 
[@pajewski19; @rockwood06; @rockwood05].



# Methods

## Population

A random 1% sample (N=14,844) was drawn from the deidentified patient
records of a large academic health center and its teaching hospital
partner. Visits during which patient age was less than 18 years old were
excluded and then patients who had fewer than three visits in the
remaining data were excluded. To avoid bias, for each patient an index
visit date was chosen and only data recorded on or after that date was
used in analysis. To avoid distorted results in patients with sparse
visit histories, those who had fewer than two visit-dates after index
visit assignment were removed from the sample as were patients whose EFI
was never higher than 0. Most patients  had visits at the tails of their
histories during which EFI was 0 and no diagnoses, procedures, nor lab results
were recorded. These may have been missed appointments or accesses to patient
data that did not physically involve the patient, so they were excluded also.
Finally, the patients were randomly assigned to a development cohort 
(N=`r snip_dat['pat']` patients, `r snip_dat['vis']` visit-days) or a testing
cohort (N=3,220 patients, 56,320 visit-days). Sensitivity analysis was
done to see the effect of leaving in data from all visits by adult
patients and the overall direction of EFI's effect was the same but
`r fs('[the model performance was inflated, see supplement]')`.
All decisions about data processing and statistical analysis were made
using only the development cohort and blinded to the testing cohort. For
publication, the same analysis scripts were run on the testing cohort
and used to create all results reported here (the development version of
each table and figure is available in the supplemental materials). The
baseline characteristics of the testing cohort are shown in Table
\@ref(tab:table1).

***

```{r table1header}
    pander(rbind(c('.','','','','','','',''))
           ,caption='(\\#tab:table1) Cohort demographics\n\n');
```


```{r table1,results='asis'}
if(.outfmt == 'html'){
  .tb1formula <- c('sex_cd','race_cd','BLANK0',.outcomenames
                   ,'Patient age (years)','Frailty','Number of Visits') %>% 
    paste0('`',.,'`',collapse='+') %>% paste('~',.,'|`Frailty Stage`') %>% 
    formula;
  
  tb1 <- table1(.tb1formula,data=dat04,render.categorical=table1cat00) %>%
    rnameshort(method='partial') %>% gsub('BLANK.','<br/>',.) %>%
    gsub('BLANK.','</br>',.) %>% gsub('Patient age','<br/>Patient age',.);
  
  print(tb1);
};
```


## Electronic Frailty Index

@clegg16a developed an electronic frailty index (EFI) for UK health systems
following the methodology of Mitnitsky, Rockwood, et al. [@mitnitski01a;
@song10; @searle08a]. Recently, @pajewski19 adapted the EFI to use ICD10 and
ICD9. We further built on this work by also mapping laboratory tests to specific
LOINC codes, de-duplicating several items, and omitting data elements which are
not available in some EMR systems. We made the mapping tables and sample SQL 
code available for public use and discussion [@bokov20]. For each patient visit,
all distinct diagnoses and abnormal lab results over the preceding two-year
window for that patient were aggregated into a single EFI, a numeric value that
can range from 0 to 1 (but in practice seldom exceeded 0.6). As mentioned above,
we omitted all visits prior to a randomly select index visit for each patient.
This did not interfere with EFI calculation because those EFI values were
calculated separately for every distinct patient-date in our health system, and
then joined to the EHR data.

## Outcomes

The primary outcomes we predicted with EFI were `r pander(snip_outcomes)`.



## Statistical Analysis

For each outcome of interest, we used a Cox proportional hazard model to
estimate the risk of the first occurrences of the outcome after the patients'
respective index visits using EFI as the predictor. Unlike earlier studies, we
treated EFI as a time-varying numeric predictor with multiple followups per
patient. Since only the first occurrence was being predicted, a recurring event
model was not necessary. Patients were followed up to the first occurrences of
the respective outcomes, so the standard form of the Cox model was sufficient.
In order to determine whether predictive accuracy of EFI was only limited to
some age subgroups, we repeated the analysis separately for the following age
ranges: 18-45 years old, 45-65 years old, and 65 or above. These split the
dataset into three roughly equal subgroups.

# Results

Table \@ref(tab:tb2) shows the results of Cox proportional hazard models for
each of the responses, with EFI as the predictor. For each 0.1 increase in
EFI, we found at least a doubling of risk: `r pander(snip_resultsfold)`.
The p-values shown have been adjusted for multiple comparisons
(`r as.english(nrow(tb2)) %>% as.character` outcomes reported in
one study) using the @holm79 method and in all cases are highly
significant. 


```{r tb2}

tb2[,c('predictor','Outcome','β^ (95% CI)','fold-change (95% CI)','SE','Z','P, adjusted','# Events','# Visits')] %>%
  pander(digits=3, row.names=FALSE,caption='(\\#tab:tb2) Cox-proportional hazards with EFI as a predictor');
```

For each outcome, we re-fit the same Cox proportional on subsets of the data
separated by age at visit: 18-45 years old, 45-65 years old, and 65 or older.
We found that for most of the primary outcomes (as well as for most of the 
analyzable outcomes, see supplementary table S1) EFI was a good predictor in
all three age groups.

```{r tb3}
tb3 %>% mutate(Outcome = rname(Outcome)) %>%
         pander(caption='(\\#tab:tb3) Performance of EFI models');
```


In Table \@ref(tab:tb3) we report the performance of EFI as the predictor for
each combination of age group and outcome, as well as for the cohort overall
(Frailty). 

In figures \@ref(fig:kmplots) a-`r letters[length(fits)]` we show Kaplan-Meier 
plots for each of the outcomes stratified by whether EFI is greater than 0.19 
(Frail=TRUE) or less (Frail=FALSE), the cut-point established by [@stow18].


```{r kmplots, fig.show="hold", fig.height=4, fig.width=12, fig.cap=" - "}
# . kmplots ----
.subfig <- 0;
for(jj in fits) {.subfig <- .subfig + 1;
with(jj,{
  message(dispname);
  print(ggmatrix(multiplot,nrow=1,ncol=3,legend=grab_legend(plot),title=dispname
           ,xAxisLabels = names(multidata)
           ,xlab=plot$label$x
           ,ylab=plot$label$y) + 
          labs(caption=paste0('Figure 1 ',letters[.subfig],'. \n')));
  cat('\n\n\n');
  # pander(modelsummary);
  # cat("\n******\n");
  })};

```



# Discussion

Our data shares the fundamental limitation of the EHR system from which it was
obtained: like all EHR systems, it only has information that providers and
coders put into it. Events taking place outside the health system or at
un-connected health systems are not visible to our analysis. On the other hand,
providers who rely on EHR systems at point of care are also working under these
limitations on. The data we used is representative of this scenario, and despite
the limitations EFI provides accurate predictions of poor patient outcomes.
Because our implementation of the Rockwood index real EHR data, it is more
directly transferable to clinical use than implementations based on curated
registries and more immediate than claims data. EFI is most accurate for
patients who have accumulated a reasonable in-system visit-history but further
work is needed to find a more precise relationship between the length of a
patient's visit-history and the accuracy of EFI and to better distinguish
genuinely non-frail patients from those who have a lot of missing data because
they are often seen outside the researchers' health system.

There is no gold standard method to assess Frailty in clinical practice.
Currently available frailty assessment tools used in geriatric practice have
good validity (for example @fried01a) but these are time
intensive and often difficult to implement in a busy general practice.
Assessing frailty helps clinicians identify high risk patients and tailor
interventions to prevent health decline and poor outcomes. Because of the
simplicity of the Rockwood Frailty Index, it is more likely to be adopted by
clinicians in a busy practice.

# Conclusions

`r fs('TBD')`

# Acknowledgments

This work was supported by NIH/NCATS UL1TR001120 (AFB, KRS), the Long School
of Medicine KL2 Award (AFB), `r fs('...',class='note2team')`



# References



---
date: '2020-09-22'

---