Fix specparam NaN handling and add flexible z-score normalization options

ncpi/EphysDatasetParser.py

@@ -180,6 +180,8 @@ class ParseConfig:
     # Post-processing
     # ---------------------------
     zscore: bool = False
+    zscore_after_epoch: bool = False  # If True, apply z-score AFTER epoching (per-epoch)
+    exclude_last_epoch: bool = False  # If True, exclude the last epoch from each time series
 
     # Aggregation: aggregate over one or more categorical columns.
     # Examples:
@@ -235,8 +237,8 @@ def parse(self, source: Union[str, Path, Any]) -> "Any":
         rows = self._parse_object(obj, source_file=source_file)
         df = self._rows_to_df(rows)
 
-        # 2) Optional z-scoring (sensor-wise, pre-aggregation)
-        if self.config.zscore:
+        # 2) Optional z-scoring BEFORE epoching (default behavior)
+        if self.config.zscore and not self.config.zscore_after_epoch:
             df = self._apply_zscore(df)
 
         # 3) Aggregate FIRST (e.g. collapse sensors)
@@ -249,6 +251,14 @@ def parse(self, source: Union[str, Path, Any]) -> "Any":
             rows = self._apply_epoching_rows(rows)
             df = self._rows_to_df(rows)
 
+            # 4b) Exclude last epoch if requested
+            if self.config.exclude_last_epoch:
+                df = self._exclude_last_epoch(df)
+
+        # 5) Optional z-scoring AFTER epoching (per-epoch normalization)
+        if self.config.zscore and self.config.zscore_after_epoch:
+            df = self._apply_zscore(df)
+
         return df
 
     # -------------------------
@@ -946,6 +956,28 @@ def z(x):
         df["data"] = df["data"].apply(z)
         return df
 
+    def _exclude_last_epoch(self, df: "Any") -> "Any":
+        """Exclude the last epoch for each unique combination of grouping columns.
+
+        This is useful for discarding the last (potentially incomplete) epoch
+        from each time series when the signal length is not perfectly divisible
+        by the epoch length.
+        """
+        if "epoch" not in df.columns:
+            return df
+
+        # Group by columns that identify unique time series (e.g., subject_id, sensor)
+        group_cols = [c for c in ["subject_id", "sensor"] if c in df.columns]
+
+        if not group_cols:
+            # No grouping columns, just exclude the global max epoch
+            max_epoch = df["epoch"].max()
+            return df[df["epoch"] < max_epoch].reset_index(drop=True)
+
+        # For each group, find the max epoch and exclude it
+        max_epochs = df.groupby(group_cols)["epoch"].transform("max")
+        return df[df["epoch"] < max_epochs].reset_index(drop=True)
+
     def _apply_aggregation(self, df):
         over = list(self.config.aggregate_over or [])
         method = self.config.aggregate_method

ncpi/Features.py

@@ -1343,8 +1343,9 @@ def specparam(
             if select_peak == "all":
                 all_peaks = peaks
             elif select_peak == "max_pw":
-                idx = int(np.nanargmax(pws))
-                selected_peaks = peaks[idx:idx + 1]
+                if not np.all(np.isnan(pws)):
+                    idx = int(np.nanargmax(pws))
+                    selected_peaks = peaks[idx:idx + 1]
             elif select_peak == "max_cf_in_range":
                 mask = (cfs >= fmin) & (cfs <= fmax)
                 if np.any(mask):