In mbadge/hipsMultimodal: Analysis of MSH Hip Data

Objective

Evaluate a model trained on the cross-sectional population on case-control cohorts with variable matching.

library(knitr); library(pander); library(tibble); library(caret)
library(AnalysisToolkit)
devtools::load_all()

knitr::opts_chunk$set(comment="#>",
                      fig.show='hold', fig.align="center", fig.height=8, fig.width=8,
                      message=FALSE,
                      warning=FALSE, cache=FALSE, rownames.print=FALSE)

ggplot2::theme_set(vizR::theme_())

set.seed(123)

Report Options

kRUN_STATS <- TRUE

Load Models

kDIR_BY_COHORT <- Fp_ml_dir("by_cohort")
InFp <- function(fn) {
    file.path(kDIR_BY_COHORT, fn)
}

models <- readRDS(file = InFp("trained_models.rds"))
load(InFp("test_cohorts.Rdata"))

data("caseControlCohorts", package = "hips")

Originally, I had been using models that trained on variable patient populations, but after discussing with the team, I'm going to just the model trained on the cross-sectional population.

model <- models$full
test_cohort <- test_cohorts$full

Evaluation

Test Cohorts

For the test sets, I'll have to take the cross-sectional test set and then filter by membership in different cohorts.

matched_test_cohorts <- map(caseControlCohorts, ~filter(test_cohort, img %in% .x))

test_cohorts <- c(list(test_cohort), matched_test_cohorts)
names(test_cohorts) <- c("crossSectional", "caseControl_matchNone", "caseControl_matchDem", "caseControl_matchPt", "caseControl_matchAll")

save(test_cohorts, file = file.path(kDIR_BY_COHORT, "by_test_cohorts.Rdata"))

Inference

# Model inference ----
pY_lst <- map(test_cohorts, ~predict_pY(model, newdata=.x))
Y_lst <- map(test_cohorts, "fx")


# Assert disjoint image partitions
train_imgs <- model %>% train_eg_ids()
test_img_sets <- test_cohorts %>% map("img")
map(test_img_sets, ~ .x %in% train_imgs) %>% 
    map_lgl(any) %>% 
    map_lgl(any) %>% 
    `!` %>% 
    stopifnot()

cCs <- map2(pY_lst, Y_lst,
     ~ClassifierCurve(pY = ..1, Y = ..2))

Tables

pretty_perf <- cCs %>% 
    glance_pretty() %>% 
    mutate(Classifier = mapvalues(Classifier,
                                  from = names(AES_COHORT_LABELS),
                                  to = unname(AES_COHORT_LABELS))) %>% 
    mapnames("Classifier", "Test Cohort") %>% 
    arrange(desc(auc))

pretty_perf %>% 
    kable(caption = "Fracture prediction with pretrained image models applied to various cohorts.")

Tbl(pretty_perf, bn = "SuppTable6_byCohortPerf")

Bootstrap Comparison

# Compare method uses spc separator
comp_tbl <- hips::compare_cCs(cCs, pilot = FALSE) 

cCs %>% map(cC2roc) %>% map(pROC::ci.auc)
cCs %>% map(cC2roc) %>% map(pROC::ci.auc, method = "bootstrap")


comp_tbl %<>% 
    map_df( ~.x[c("p.value", "method", "statistic")], .id = "cC_pair") %>%
    tidyr::extract("cC_pair", c("cC1", "cC2"), "(\\w+)-(\\w+)")

comp_tbl %>% 
    arrange(desc(p.value)) %>% 
    mutate(cC1 = mapvalues(cC1,
                           from = names(AES_COHORT_LABELS) %>% str_case_snake(),
                           to = unname(AES_COHORT_LABELS)),
           cC2 = mapvalues(cC2,
                           from = names(AES_COHORT_LABELS) %>% str_case_snake(),
                           to = unname(AES_COHORT_LABELS))) 

comp_tbl %>% 
    mutate(p.value = ifelse(p.value < 0.05,
                      yes = kableExtra::cell_spec(p.value, color = "green", bold = TRUE),
                      kableExtra::cell_spec(p.value, color = "red", bold = FALSE))) %>% 
    mutate(p.value = str_replace(p.value, "([0-9]\\.[0-9]{3})[0-9]+", "\\1")) %>% 
    mutate(p.value = str_replace(p.value, "([0-9])\\.[0-9]{3}e", "\\1e")) %>% 
    select("Classifier 1" = cC1, "Classifier 2" = cC2, "Bootstrap Test p-value"=p.value) %>% 
    knitr::kable(format = "markdown", caption = "Comparing Fracture Models predicted by Various Predictor Sets.") %>%
    kableExtra::kable_styling(bootstrap_options = c("condensed", "hover", "striped"), full_width = FALSE)

comp_tbl %>% 
    mutate(p.value = str_replace(p.value, "([0-9]\\.[0-9]{3})[0-9]+", "\\1")) %>% 
    mutate(p.value = str_replace(p.value, "([0-9])\\.[0-9]{3}e", "\\1e")) %>% 
    select("Classifier 1" = cC1, "Classifier 2" = cC2, "Bootstrap Test p-value"=p.value) %>% 
    Tbl(bn = "SuppTable7_byCohortComp")

comp_tbl %>% 
    mutate(p.value = ifelse(p.value < 0.05,
                      yes = kableExtra::cell_spec(p.value, color = "green", bold = TRUE),
                      kableExtra::cell_spec(p.value, color = "red", bold = FALSE))) %>% 
    mutate(p.value = str_replace(p.value, "([0-9]\\.[0-9]{3})[0-9]+", "\\1")) %>% 
    mutate(p.value = str_replace(p.value, "([0-9])\\.[0-9]{3}e", "\\1e")) %>% 
    select("Classifier 1" = cC1, "Classifier 2" = cC2, "Bootstrap Test p-value"=p.value) %>% 
    knitr::kable(format = "html", escape = FALSE, caption = "Comparing Fracture Models predicted by Various Predictor Sets.") %>%
    kableExtra::kable_styling(bootstrap_options = c("condensed", "hover", "striped"), full_width = FALSE) %>% 
    Tbl(bn = "SuppTable7_byCohortComp", tbl_type = "html")

DATA  <- glance(cCs) %>% 
    mapnames("Classifier", "cohort")
DATA %<>% arrange(desc(auc))


glue::glue("We then evaluated the image-only classifier for fracture on each test set (Figure 3B-D, Supplementary Tables 6&7).  The area under the Precision Recall Curve (PRC) is dependent on the disease prevalence, and since the original population had a 3% fracture prevalence but case-control cohorts have a 50% prevalence, the PRC is significantly higher for case-control cohorts.  Random subsampling had no effect on the primary evaluation metric, AUC ({DATA %>% filter(cohort=='crossSectional') %$% auc %>% round(digits=2)} vs {DATA %>% filter(cohort=='caseControl_matchNone') %$% auc %>% round(digits=2)}, p={comp_tbl[1, 'p.value'] %>% round(digits=2)}).  Balancing patient demographics (age and gender) also made no difference to model performance (p={comp_tbl[3, 'p.value'] %>% round(digits=2)}).  Model performance was significantly lower when evaluating on a test cohort matched by patient demographics, bmi and symptoms (AUC={DATA %>% filter(cohort=='caseControl_matchPt') %$% auc %>% round(digits=2)}, p={comp_tbl[6, 'p.value'] %>% round(digits=2)}).  When evaluated on a test cohort matched by all covariates, the fracture detector was no longer better than random (AUC={DATA %>% filter(cohort=='caseControl_matchAll') %$% auc %>% round(digits=2)}, 95% CI {DATA %>% filter(cohort=='caseControl_matchAll') %>% with(., glue('{auc_lower %>% round(digits=2)} - {auc_upper %>% round(digits=2)}'))}) and significantly worse than when assessed on all other test cohorts.")

Views

# Target learnability
# ROC by target

DATA$cohort %<>% fct_reorder(DATA$auc)
PLT_DAT_ROC_SUM <- DATA  # Save intermediate plot data

gg_roc <- ggplot(DATA,
                 aes(x = fct_reorder(cohort, auc),
                     y = auc,
                     ymin = auc_lower,
                     ymax = auc_upper)) +
    geom_errorbar(width = 0.25, position = Pd()) +
    geom_point(position = Pd()) +
    geom_text(aes(label = is_sig), nudge_x = 0.2) +
    aes_cohort_col() +
    scale_x_cohort()
gg_roc <- gg_roc +
    geom_hline(yintercept = 0.5, alpha = 0.5, linetype = 2) +
    geom_hline(yintercept = 1, alpha = 0.5) +
    theme(axis.text.x = element_text(angle = 90)) +
    labs(y = "AUROC +/- 95% bootstrap CI") +
    theme(strip.text.y = element_text(angle = 0),
          axis.text.x = element_text(vjust = 0.5)) +
    scale_y_continuous(breaks = c(0.5, 0.75, 1), labels = c(0.5, 0.75, 1)) +
    theme(legend.position = "bottom", legend.direction = "vertical")
GG_ROC <- gg_roc %+%
    coord_flip() +
    labs(title = "Different methods of subsampling a test-set lead to markedly different performance scores") +
    theme(plot.title = element_text(hjust = 1))
GG_ROC

# ROC
XY_byClassifier <- purrr::map(cCs, gg_data_roc)
XY_byGeom <- purrr::transpose(XY_byClassifier)
xy_geoms <- purrr::map(XY_byGeom, purrr::lift_dl(dplyr::bind_rows, .id = "cohort"))


PLT_DAT_ROC <- xy_geoms  # Save intermediate plot data

gg_roc <- ggplot2::ggplot(xy_geoms$lines, ggplot2::aes(x=x, y=y)) +
    ggplot2::geom_line() +
    ggplot2::geom_point(data = xy_geoms$points, shape=10, size = 3) +
    ggplot2::geom_rug(data = xy_geoms$points) +
    ggplot2::geom_text(data = xy_geoms$points, aes(label = is_sig), size = 5, nudge_y = 0.03, show.legend=FALSE) +
    gg_style_roc +
    aes_cohort_col()


# PRROC
XY_byGeom <- map(cCs, gg_data_prc) %>% transpose()
xy_geoms <- map(XY_byGeom, lift_dl(bind_rows, .id="cohort"))

PLT_DAT_PRC <- xy_geoms  # Save intermediate plot data

gg_prc <- ggplot2::ggplot(xy_geoms$lines, ggplot2::aes(x=x, y=y)) +
    ggplot2::geom_line() +
    ggplot2::geom_point(data = xy_geoms$points, shape=10, size = 3) +
    ggplot2::geom_rug(data = xy_geoms$points) +
    gg_style_prc +
    aes_cohort_col()


gg_prc %<>% `+`(list(theme(legend.direction = "vertical", legend.position = "bottom")))

gg_prc %<>% `+`(guides(color = guide_legend(title = "Cohort", ncol = 2, title.hjust = 0.5)))
leg <- cowplot::get_legend(gg_prc)
gg_prc %<>% `+`(NL)
gg_roc %<>% `+`(NL)

GG <- cowplot::plot_grid(
    cowplot::plot_grid(gg_roc, gg_prc, ncol = 2, align='hv'),
    cowplot::plot_grid(NULL, leg, NULL, ncol = 3, rel_widths = c(1, .1, 1)),
    ncol = 1, rel_heights = c(.6, 0.4))
GG

# cowplot::save_plot("analysis/figures/ROC_PRROC.png", GG, base_width = 4, base_height = 6)

Fx Associations

• Tabulate number of significant associations
• View ORs

cohort_imgs <- map(test_cohorts, "img")
mshBinary <- hipsCohort(mutating = binary)
mshBinCohorts <- map(cohort_imgs, ~filter(mshBinary, img %in% .x))

OR_DATA <- mshBinCohorts %>% 
    OddsRatios(grp_chr = "cohort")
OR_DATA %>% 
    ggOddsRatios() %+% 
    aes_cohort_col()

sig_ors <- OR_DATA %>%
    mutate(is_sig = map_lgl(p.value, ~ .x < 0.05)) %>%
    group_by(cohort) %>%
    filter(is_sig)

sig_ors %>% 
    mutate(target = map_chr(target, PrettyTargets)) %>% 
    summarise(`Sig. Assoc` = str_x(target)) %>% 
    kable(caption = "Significant fracture-covariate associations in each cohort")
sig_ors %>%
    summarise(n_sig = n()) %>% 
    kable(caption = "Number of significant fracture-covariate associations in each cohort")

Model training stats

models %>% map(train_terms)
models %>% map(train_n_eg)

FP_OUT_INT_PLT_DATA <- file.path(kDIR_BY_COHORT, "plot_data.Rdata")
save(PLT_DAT_ROC_SUM, PLT_DAT_ROC, PLT_DAT_PRC, file = FP_OUT_INT_PLT_DATA)

Codebase State: r dotfileR::gitState() on r date()

mbadge/hipsMultimodal documentation built on May 9, 2019, 12:05 a.m.