R/perfCalc.R

In BaderLab/netDx: Network-based patient classifier

#' Computes variety of predictor evaluation measures based on the confusion
#' matrix
#'
#' @param dat (data.frame): 5 columns: score, tp, fp, tn, fn. 
#' One row per cutoff
#' score for feature selection
#' @return (list)
#' stats (data.frame): score, f1, ppv, precision and recall. One row
#' per cutoff for feature selection
#' auc (numeric between 0 and 1): AUC of overall ROC curve
#' prauc (numeric between 0 and 1): AUC of overall precision-recall curve
#' @importFrom pracma trapz
#' @examples
#' data(confmat)
#' x <- perfCalc(confmat)
#' @export
perfCalc <- function(dat) {
    dat <- na.omit(dat)
    # F1 - harmonic mean of precision recall resolves to the formula below
    tp2 <- 2 * dat$tp
    f1 <- tp2/(tp2 + dat$fp + dat$fn)
    
    # precision recall curve
    
    # precision = positive predictive value (pr = ppv)
    ppv <- dat$tp/(dat$tp + dat$fp)
    rec <- dat$tp/(dat$tp + dat$fn)
    # trapz integrates from right to left, so you need to apply rev() 
		# otherwise you

    # get a negative area.
    prauc <- pracma::trapz(rev(rec), rev(ppv))
    
    # roc auc
    x <- dat$fp/(dat$fp + dat$tn)
    y <- dat$tp/(dat$tp + dat$fn)
    
    x <- c(0, rev(x), 1)
    y <- c(0, rev(y), 1)
    auc <- pracma::trapz(x, y)
    out <- data.frame(score = dat$score, ppv = ppv, f1 = f1, rec = rec)
    
    return(list(stats = out, auc = auc, prauc = prauc))
}