diff --git a/inst/tinytest/test_utils_anomaly_score.R b/inst/tinytest/test_utils_anomaly_score.R
index 70959e05..7bc72f29 100644
--- a/inst/tinytest/test_utils_anomaly_score.R
+++ b/inst/tinytest/test_utils_anomaly_score.R
@@ -383,5 +383,30 @@ duplicate_metrics = run_quality_metrics(
 # The last 5 rows (with high values) should have lower mean anomaly scores
 # Since they are all clumped between 2 and 4, whereas 0.1 is by itself
 expect_true(mean(duplicate_metrics$AnomalyScores[6:10]) < mean(duplicate_metrics$AnomalyScores[1:5]),
-            info = "Rows 6-10 (values clumped 2-4) should have lower 
+            info = "Rows 6-10 (values clumped 2-4) should have lower
             anomaly scores than rows 1-5 (isolated value of 0.1)")
+
+nan_first_row_df = create_base_df(5)
+nan_first_row_df$QualityMetric.mean_increase = c(NA, 0.2, 0.4, 0.6, 0.8)
+
+nan_first_row_result = tryCatch({
+    MSstatsConvert:::.runAnomalyModel(
+        nan_first_row_df,
+        n_trees = 100,
+        max_depth = "auto",
+        cores = 1,
+        split_column = "PSM",
+        quality_metrics = c("QualityMetric.mean_increase"))
+}, error = function(e) e)
+
+expect_false(inherits(nan_first_row_result, "error"),
+    info = paste(
+        "Anomaly model should not crash/error when a quality metric has a",
+        "leading NA/NaN value within a PSM group.",
+        if (inherits(nan_first_row_result, "error"))
+            paste("Got error:", conditionMessage(nan_first_row_result)) else ""))
+
+if (!inherits(nan_first_row_result, "error")) {
+    expect_true(all(is.finite(nan_first_row_result$AnomalyScores)),
+        info = "Anomaly scores should be finite even when a quality metric has a leading missing value")
+}
diff --git a/src/isolation_forest.cpp b/src/isolation_forest.cpp
index 7ac6dae9..ad82975d 100644
--- a/src/isolation_forest.cpp
+++ b/src/isolation_forest.cpp
@@ -78,22 +78,30 @@ std::unique_ptr<IsolationTreeNode> isolation_tree(
   std::string split_feature = features[feature_dist(gen)];
   
   // Can split on numeric or missing value
-  double min_val = data[0].at(split_feature);
-  double max_val = min_val;
+  double min_val = 0.0;
+  double max_val = 0.0;
   bool has_missing = false;
+  bool has_valid = false;
   for (const auto& row : data) {
-    if (std::isnan(row.at(split_feature))) {
+    double val = row.at(split_feature);
+    if (std::isnan(val)) {
       has_missing = true;
       continue;
     }
-    min_val = std::min(min_val, row.at(split_feature));
-    max_val = std::max(max_val, row.at(split_feature));
+    if (!has_valid) {
+      min_val = val;
+      max_val = val;
+      has_valid = true;
+    } else {
+      min_val = std::min(min_val, val);
+      max_val = std::max(max_val, val);
+    }
   }
-  
-  if (min_val == max_val && !has_missing) {
+
+  if (!has_valid || (min_val == max_val && !has_missing)) {
     return std::make_unique<IsolationTreeNode>(n);
   }
-  
+
   // TODO: Chance to chose missing is 50/50. Could make less likely. Test
   bool is_missing_split = false;
   double split_value = 0.0;