In chjackson/adverse: Data on adverse events of adjuvant breast cancer treatment

Bisphosphonate effects on events compared between cancer and healthy patients.

Just two studies of bisphosphonates for preventing osteoperosis in "healthy" women without cancer: FIT and Horizon.

HORIZON: two papers reported on the same population, one reporting acute events following the start of treatment, and one with longer follow-up. It's unclear whether some of the same events are repeated between the two reports.
FIT: two papers, one reporting the full trial population and the other reporting a subgroup. The only events reported were abdominal pain, oesophagitis, dyspepsia, gastrointestinal (unspecified) and nausea.

clean_aetype <- function(aetype){
    aetype <- stringr::str_to_sentence(aetype)
    aetype[aetype=="Ggt increased"] <-"GGT increased"
    aetype[aetype=="Cpk increased"] <-"CPK increased"
    aetype[aetype=="Inr increased"] <-"INR increased"
    aetype[aetype=="Withdrawals because of aes"] <-"Withdrawals because of AEs"
    aetype
}

library(flextable)
library(meta)
library(plotly)
library(binom) 
library(dplyr)
filter <- dplyr::filter
load_all("../../gemtc/gemtc")
load_all("..")
source("theme.r")
load("ors.rda")

hae <- unique(bpaeprevma$aetype) 
bpaeprevma$patient <- "Healthy"
bpcma <-
  bpaeriskdiff %>%
  filter(trtcon %in% c("100","101","103")) %>%
  filter(aetype %in% hae) %>%
  rename(ract=count, nact=N) %>% 
  mutate(patient = "Cancer") %>% 
  select(names(bpaeprevma)) %>% 
  rbind(bpaeprevma) %>%
  arrange(patient, study, aetype) %>%
  mutate(p0 = (rcon+0.5)/(ncon+1),
         p1 = (ract+0.5)/(nact+1),
         or = p1/(1-p1) / (p0/(1-p0)),
         logor = log(or), 
         selogor = sqrt(1/(rcon+0.5) + 1/(ncon-rcon+0.5) +
                        1/(ract+0.5) + 1/(nact-ract+0.5)),
         orl = exp(log(or) - qnorm(0.975)*selogor),
         oru = exp(log(or) + qnorm(0.975)*selogor)) %>%
  mutate(`Events` = sprintf("\nTreatment %s/%s\nControl %s/%s",
                          ract, nact, rcon, ncon))

The table below compares the odds ratios from the network meta-analysis of bisphosphonates in women with cancer with the equivalent odds ratios from a simple meta-analysis of HORIZON and FIT, for the events that are significant in the cancer analysis and also reported in HORIZON and FIT.

knitr::opts_chunk$set(echo=FALSE,
                      message=FALSE,
                      fig.width=15,
                      fig.height=15)

eventsh <- bpaeprevma %>% filter(rcon>0|ract>0) %>% select(aetype) %>% unlist %>% as.vector %>% unique
mares <- data.frame(event=eventsh)
mares$or <- mares$orl <- mares$oru <- NA
mares$rd <- mares$rdl <- mares$rdu <- NA

for (i in seq(along=eventsh)){ 
    bpma <- bpaeprevma %>% filter(aetype==eventsh[i])
    if (nrow(bpma) > 0){
        mres <- metabin(ract, nact, rcon, ncon, studlab=study, data=bpma, sm="OR")
        mares[i,c("or","orl","oru")] <- exp(unlist(mres[c("TE.fixed","lower.fixed","upper.fixed")]))
        mresrd <- metabin(ract, nact, rcon, ncon, studlab=study, data=bpma, sm="RD")
        mares[i,c("rd","rdl","rdu")] <- unlist(mresrd[c("TE.fixed","lower.fixed","upper.fixed")])
    }
}

mares <- mares %>%
  mutate(actlab = "Bisphosphonate",
         patient = "Healthy",
         event_trt = sprintf("%s: Bisphosphonate", event),
         sig = (or>1.5 & orl>1) | (rd>0.02 & rdl>0)
         )

sigevents <- ortab %>% filter(sig) %>% pull(event) %>% unique
sigevent_trts <- ortab %>% filter(sig) %>% mutate(event_trt = paste(event, actlab, sep=": ")) %>% pull(event_trt) %>% unique

orsh <- ors %>%
  filter(measure=="or" & model=="NMA") %>%
  mutate(event_trt = paste(event, actlab, sep=": ")) %>%
  filter(event_trt %in% sigevent_trts) %>% 
  select(event, actlab, est, lower, upper) %>%
  mutate(patient="Cancer") %>%
  rename(or=est, orl=lower, oru=upper) %>%
  mutate(event_trt = sprintf("%s: %s",event, actlab))

maress <- mares %>% filter(event %in% sigevents)
orsh <- orsh %>% filter(event %in% pull(maress, event))
maress <- maress %>% filter(event %in% pull(orsh, event))
orsh <- orsh %>% full_join(maress)

## Why no ONJ, cancer? 
## Cos that's only sig under direct
## Just remove it then. 

ggplot(orsh, aes(y=or, x=event, col=patient, lty=actlab)) +
  coord_flip(ylim=c(0.7, 30)) +
  geom_hline(yintercept=1, col="gray") + 
  geom_point(position=position_dodge(0.6), size=4) +
  geom_errorbar(aes(ymin = orl, ymax=oru), position=position_dodge(0.6), size=2) +
  scale_y_continuous(trans="log", breaks=c(0.7, 1, 1.5, 2, 5, 10, 20, 30)) +
  ylab("Odds ratio") +
  xlab("") +
  labs(col="Patients", lty="Bisphosphonate treatment") +
  paper_theme +
  theme(plot.margin = margin(t = 0, r = 100, b = 0, l = 0, unit = "pt"),
        legend.key.width = unit(3, "cm"))

marestab <- maress %>%
  mutate(orhealthy = sprintf("%s (%s, %s)", round(or,2), round(orl,2), round(oru,2)))
orsh <- ortab %>% filter(sig) %>% select(categ, event, actlab, NMA) %>%
  left_join(marestab, by="event") %>% filter(!is.na(orhealthy)) %>%
  rename(actlab="actlab.x", treatmenth="actlab.y") %>% 
  as_tibble %>%
  select(categ, event, actlab, NMA, treatmenth, orhealthy) %>%
  as_grouped_data(groups="categ") %>%
  flextable %>% 
  set_header_labels(categ="", event="",
                    actlab="Treatment (cancer NMA)",
                    NMA="OR (Cancer, NMA)",
                    treatmenth="Treatment (healthy MA)",
                    orhealthy="OR (healthy MA)"
                    ) %>% 
  flextable::bold(j = 1, i = ~ !is.na(categ), bold = TRUE, part = "body" ) %>%
  flextable::bold(part = "header", bold = TRUE )  %>%
  width(width = 2)
orsh

Events deemed significant in a fixed effects meta analysis of HORIZON/FIT, using same criterion as in the main analysis: statistically significant OR > 1.5 or risk difference > 0.02.

mares %>% filter(sig) %>%
  mutate(orstr = sprintf("%s (%s, %s)", round(or,2), round(orl,2), round(oru,2))) %>%
  mutate(rdstr = sprintf("%s (%s, %s)", round(rd,2), round(rdl,2), round(rdu,2))) %>%
  select(event, orstr, rdstr) %>%
  flextable %>% 
  set_header_labels(event="",
                    orstr="Odds ratio",
                    rdstr="Risk difference") %>% 
  width(width = 2)

The study-specific odds ratios underlying these results are shown below.

For each event, the odds ratios from HORIZON and FIT are not surprising given the variability between studies.

knitr::opts_chunk$set(echo=FALSE,
                      message=FALSE,
                      fig.width=15,
                      fig.height=40)

p <- bpcma %>%
  filter(aetype %in% unique(ortab$event[ortab$sig])) %>% 
ggplot(aes(y=or, x=study, col=patient, label=Events)) +
                                        #  coord_flip(ylim=c(0.1, 15)) +
  coord_flip(ylim=c(0.1, 20)) +
  geom_hline(yintercept = 1, col="gray") + 
  geom_point() +
  geom_errorbar(aes(ymin = orl, ymax = oru), width=0) +
  facet_wrap(~ aetype, ncol=3) + 
  scale_y_continuous(trans="log", 
                     breaks=c(0.1, 0.2, 0.5, 1, 2, 5, 10, 20)) +
  ylab("Odds ratio (bisphosphonates / control)") +
  xlab("") +
  scale_colour_manual(values=c("black","red")) +
  paper_theme

#ggplotly(p, tooltip=c("label"), height=5000)
p

knitr::opts_chunk$set(echo=FALSE,
                      message=FALSE,
                      fig.width=15,
                      fig.height=12)

bpplot <- bpcma %>%
  filter(aetype %in% unique(ortab$event[ortab$sig])) %>%
  mutate(aetype = clean_aetype(aetype))

bpplot$aetype[bpplot$aetype=="Influenza-like symptoms"] <- "Influenza-like"

evsig <- unique(bpplot$aetype)

bsize <- 20
supp_theme <-
  theme(axis.text = element_text(size=bsize),
        strip.text = element_text(size=bsize),
        text = element_text(size=bsize),
        legend.text = element_text(size=bsize),
        axis.title = element_text(size=bsize, margin=margin(t=0.2, b=0.2)),
        axis.title.x = element_text(size=bsize,
                                    margin=margin(t=10, b=10)))

plotfn <- function(ev){ 
    ggplot(bpplot[bpplot$aetype %in% ev,],
           aes(y=or, x=study, col=patient, label=Events)) +
      coord_flip(ylim=c(0.1, 20)) +
      geom_hline(yintercept = 1, col="gray",size=2) + 
      geom_point(size=2) +
      geom_errorbar(aes(ymin = orl, ymax = oru), width=0) +
      facet_grid(cols=vars(aetype)) + 
      scale_y_continuous(trans="log", 
                         breaks=c(0.1, 0.5, 1, 5)) +
      ylab("Odds ratio (bisphosphonates / control)") +
      xlab("") +
      scale_colour_manual(values=c("black","red")) +
      supp_theme
}

plotfn(evsig[1:3])
plotfn(evsig[4:6])
plotfn(evsig[7:9])
plotfn(evsig[10:13])