--- a
+++ b/datasets/preprocess/tools.py
@@ -0,0 +1,154 @@
+import math
+
+import numpy as np
+import pandas as pd
+
+
+def calculate_data_existing_length(data):
+    res = 0
+    for i in data:
+        if not pd.isna(i):
+            res += 1
+    return res
+
+
+# elements in data are sorted in time ascending order
+def fill_missing_value(data, to_fill_value=0):
+    data_len = len(data)
+    data_exist_len = calculate_data_existing_length(data)
+    if data_len == data_exist_len:
+        return data
+    elif data_exist_len == 0:
+        # data = [to_fill_value for _ in range(data_len)]
+        for i in range(data_len):
+            data[i] = to_fill_value
+        return data
+    if pd.isna(data[0]):
+        # find the first non-nan value's position
+        not_na_pos = 0
+        for i in range(data_len):
+            if not pd.isna(data[i]):
+                not_na_pos = i
+                break
+        # fill element before the first non-nan value with median
+        for i in range(not_na_pos):
+            data[i] = to_fill_value
+    # fill element after the first non-nan value
+    for i in range(1, data_len):
+        if pd.isna(data[i]):
+            data[i] = data[i - 1]
+    return data
+
+
+def forward_fill_pipeline(
+    df: pd.DataFrame,
+    default_fill: pd.DataFrame,
+    demographic_features: list[str],
+    labtest_features: list[str],
+    target_features: list[str],
+    require_impute_features: list[str],
+):
+    grouped = df.groupby("PatientID")
+
+    all_x = []
+    all_y = []
+    all_pid = []
+    all_record_times = []  # List to store record times for each patient
+    all_missing_masks = []
+    
+
+    for name, group in grouped:
+        sorted_group = group.sort_values(by=["RecordTime"], ascending=True)
+        patient_x = []
+        patient_y = []
+        patient_record_times = []  # List to store record times for the current patient
+        patient_missing_masks = pd.isna(sorted_group[labtest_features]).values.astype(int).tolist()
+
+        for f in require_impute_features:
+            # if the f is not in the default_fill, then default to -1
+            if f not in default_fill: # these are normally categorical features
+                to_fill_value = -1
+            else:
+                to_fill_value = default_fill[f]
+            # take median patient as the default to-fill missing value
+            fill_missing_value(sorted_group[f].values, to_fill_value)
+
+        for _, v in sorted_group.iterrows():
+            patient_record_times.append(v['RecordTime'])
+
+            target_values = []
+            for f in target_features:
+                target_values.append(v[f])
+            patient_y.append(target_values)
+            x = []
+            for f in demographic_features + labtest_features:
+                x.append(v[f])
+            patient_x.append(x)
+        all_x.append(patient_x)
+        all_y.append(patient_y)
+        all_pid.append(name)
+        all_record_times.append(patient_record_times)
+        all_missing_masks.append(patient_missing_masks)
+    return all_x, all_y, all_pid, all_record_times, all_missing_masks
+
+
+# outlier processing
+def filter_outlier(element):
+    if np.abs(float(element)) > 1e4:
+        return 0
+    else:
+        return element
+
+def normalize_dataframe(train_df, val_df, test_df, normalize_features, require_norm_later=True):
+    # Calculate the quantiles
+    q_low = train_df[normalize_features].quantile(0.05)
+    q_high = train_df[normalize_features].quantile(0.95)
+
+    # Filter the DataFrame based on the quantiles
+    filtered_df = train_df[(train_df[normalize_features] > q_low) & (
+        train_df[normalize_features] < q_high)]
+
+    # Calculate the mean and standard deviation and median of the filtered data, also the default fill value
+    train_mean = filtered_df[normalize_features].mean()
+    train_std = filtered_df[normalize_features].std()
+    train_median = filtered_df[normalize_features].median()
+
+    # if certain feature's mean/std/median is NaN, then set it as 0. This feature will be filled with 0 in the following steps
+    train_mean = train_mean.fillna(0)
+    train_std = train_std.fillna(0)
+    train_median = train_median.fillna(0)
+
+    if require_norm_later:
+        default_fill: pd.DataFrame = (train_median-train_mean)/(train_std+1e-12)
+        # LOS info
+        los_info = {"los_mean": train_mean["LOS"].item(
+        ), "los_std": train_std["LOS"].item(), "los_median": train_median["LOS"].item()}
+
+        # Calculate large los and threshold (optional, designed for covid-19 benchmark)
+        los_array = train_df.groupby('PatientID')['LOS'].max().values
+        los_p95 = np.percentile(los_array, 95)
+        los_p5 = np.percentile(los_array, 5)
+        filtered_los = los_array[(los_array >= los_p5) & (los_array <= los_p95)]
+        los_info.update({"large_los": los_p95.item(), "threshold": filtered_los.mean().item()*0.5})
+
+
+        # Z-score normalize the train, val, and test sets with train_mean and train_std
+        train_df.loc[:, normalize_features] = (train_df.loc[:, normalize_features] - train_mean) / (train_std+1e-12)
+        val_df.loc[:, normalize_features] = (val_df.loc[:, normalize_features] - train_mean) / (train_std+1e-12)
+        test_df.loc[:, normalize_features] = (test_df.loc[:, normalize_features] - train_mean) / (train_std+1e-12)
+
+        train_df.loc[:, normalize_features] = train_df.loc[:, normalize_features].map(filter_outlier)
+        val_df.loc[:, normalize_features] = val_df.loc[:, normalize_features].map(filter_outlier)
+        test_df.loc[:, normalize_features] = test_df.loc[:, normalize_features].map(filter_outlier)
+
+        return train_df, val_df, test_df, default_fill, los_info, train_mean, train_std
+
+    else:
+        default_fill: pd.DataFrame = train_median
+        return default_fill
+
+def normalize_df_with_statistics(df, normalize_features, train_mean, train_std):
+    df.loc[:, normalize_features] = (df.loc[:, normalize_features] - train_mean) / (train_std+1e-12)
+    df.loc[:, normalize_features] = df.loc[:, normalize_features].map(filter_outlier)
+    return df
+