merge main to dev

R/metric_generators.R

 stop("rf (ranger) prediction type not supported or unknown prediction structure.") 
 stop("rf (ranger) prediction type not supported or unknown prediction structure.") 
@@ -149,16 +149,34 @@ predict_custom <- function(x, y = NULL, fit, model, type = "response") {
     }
 
     # Survival: ranger returns a matrix of survival probabilities by timepoint
-    if (is.matrix(preds) && inherits(fit, "ranger") && fit$treetype == "Surv") {
-      # If user asks for survival probabilities, return the matrix
+    if (is.matrix(pr$survival) && inherits(fit, "ranger") && fit$treetype == "Survival") {
+      # If user asks for survival probabilities, return the survival matrix
       if (type == "survival") {
-        return(preds)
+
+        times <- pr$unique.death.times
+        surv_matrix <- pr$survival
+
+        return(surv_matrix)
       }
-      # For linear predictor / risk score, convert survival curves to a summary risk:
-      # pragmatic approach: use negative mean survival (higher => worse risk)
+      # For linear predictor / risk score, convert survival to linear predictor using
+      # logit = (S(t) / 1- S(t)) or
+      # lp = log(H(t)) where H(t) = sum(h(t))
       if (type == "lp") {
-        lp <- -rowMeans(preds, na.rm = TRUE)
-        return(as.numeric(lp))
+
+        #times <- pr$unique.death.times
+        #surv_matrix <- pr$survival
+
+        # obtain the survival at an ith observation unique time
+        #surv_vec <- sapply(1:nrow(x_df), function(i) {
+        #   t_i <- x_df$time[i]
+        #   idx <- which.min(abs(times - t_i))
+        #  surv_matrix[i, idx]
+        # })
+
+        # get lp from survival vector + adjust S(t) = 1 by subtracting 1e-8
+        # surv_vec_to_lp <- qlogis(surv_vec - 1e-8)
+        # chf n x times matrix: summing over times gives cumulative hazard
+        return(log(rowSums(pr$chf)))
       }
     }
 
@@ -306,7 +324,9 @@ continuous_calib_itl <- function(data, fit, model) {
 
 survival_cindex <- function(data, fit, model) {
   y_surv <- survival::Surv(data$time, data$event)
-  x <- data[, !(names(data) %in% c("time", "event", "id")), drop = FALSE]
+
+    x <- data[, !(names(data) %in% c("time", "event", "id")), drop = FALSE]
+
   # request linear predictor / risk score
   y_hat <- try(predict_custom(x, NULL, fit, model, type = "lp"), silent = TRUE)
   if (inherits(y_hat, "try-error")) {
@@ -322,7 +342,9 @@ survival_cindex <- function(data, fit, model) {
 # Cox-like calibration slope (uses linear predictor)
 survival_calib_slope <- function(data, fit, model) {
   y_surv <- survival::Surv(data$time, data$event)
+
   x <- data[, !(names(data) %in% c("time", "event", "id")), drop = FALSE]
+
   y_hat <- try(predict_custom(x, NULL, fit, model, type = "lp"), silent = TRUE)
   if (inherits(y_hat, "try-error")) {
     return(NaN)

R/model_generators.R

@@ -25,13 +25,29 @@ default_models <- list(
       nthreads <- ncores - 2
 
       x <- d[, -1, drop = FALSE]
-      y <- d[, 1]
+      y <- as.factor(d[, 1])
+
+      #### new
+
+      ff <- NULL
+      invisible(
+        capture.output(
+          ff <- randomForest::tuneRF(x, y, trace = FALSE, plot = FALSE)
+        )
+      )
+
+      bestmtry <- data.frame(ff)
+      mtry_best <- bestmtry$mtry[which.min(bestmtry$OOBError)]
+
       ranger::ranger(
         x = x,
         y = y,
+        mtry = mtry_best,
         probability = TRUE,
-        num.threads = nthreads
+        num.trees = 300,
+        num.threads = nthreads 
       )
+
     },
     xgboost = function(d, nrounds = 100, params = list(objective = "binary:logistic", eval_metric = "logloss")) {
       # expects first column y (0/1), remaining columns predictors
@@ -141,7 +157,7 @@ default_models <- list(
 
       stopifnot(all(c("time", "event") %in% colnames(d)))
       formula <- stats::as.formula("survival::Surv(time, event) ~ .")
-      ranger::ranger(formula, data = d, num.threads = nthreads)
+      ranger::ranger(formula, data = d,  num.trees = 300, num.threads = nthreads)
     },
     xgboost = function(d, nrounds = 100, params = list(objective = "survival:cox", eval_metric = "cox-nloglik")) {
       # XGBoost Cox objective: uses times as label but does not directly take a censoring vector.

R/simulate_wrappers.R

@@ -364,7 +364,7 @@ simulate_survival <- function(
   )
 
   # Tune the data-generating function
-  tune_param <- binary_tuning(
+  tune_param <- survival_tuning(
     target_prevalence = 1 - censoring_rate,
     target_performance = large_sample_cindex,
     candidate_features = signal_parameters,