Source code for benchpots.datasets.physionet_2012

"""
Preprocessing func for the dataset PhysionNet2012.

"""

# Created by Wenjie Du <wenjay.du@gmail.com>
# License: BSD-3-Clause

from typing import Any, Optional, Sequence, Union

import numpy as np
import pandas as pd
import tsdb
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler

from ..utils.logging import logger, print_final_dataset_info
from ..utils.missingness import create_missingness
from ..utils.task_type import convert_processed_dataset_by_task_type


[docs] def preprocess_physionet2012( subset: str, rate: float, pattern: str = "point", features: Optional[list] = None, random_state: Optional[int] = None, task_type: str = "imputation", n_pred_steps: int = 1, forecast_feature_indices: Optional[Union[int, Sequence[int]]] = None, **kwargs: Any, ) -> dict: """Load and preprocess the dataset PhysionNet2012. Parameters ---------- subset: The name of the subset dataset to be loaded. Must be one of ['all', 'set-a', 'set-b', 'set-c']. rate: The missing rate. pattern: The missing pattern to apply to the dataset. Must be one of ['point', 'subseq', 'block']. features: The features to be used in the dataset. If None, all features except the static features will be used. random_state: Controls the randomness of the train/validation/test split. Pass an int for reproducible splits across runs. task_type: Task type for postprocessing. Supported values are ['imputation', 'forecasting', 'classification', 'clustering', 'anomaly_detection']. n_pred_steps: Forecasting horizon. Effective only when task_type is 'forecasting'. forecast_feature_indices: Target feature indices for forecasting labels. If None, all features are used. Returns ------- processed_dataset : A dictionary containing the processed PhysionNet2012. """ def apply_func(df_temp: pd.DataFrame) -> pd.DataFrame: # pad and truncate to set the max length of samples as 48 missing = list(set(range(0, 48)).difference(set(df_temp["Time"]))) missing_part = pd.DataFrame({"Time": missing}) df_temp = pd.concat( [df_temp, missing_part], ignore_index=False, sort=False ) # pad the sample's length to 48 if it doesn't have enough time steps df_temp = df_temp.set_index("Time").sort_index().reset_index() df_temp = df_temp.iloc[:48] # truncate return df_temp all_subsets = ["all", "set-a", "set-b", "set-c"] assert subset.lower() in all_subsets, f"subset should be one of {all_subsets}, but got {subset}" assert 0 <= rate < 1, f"rate must be in [0, 1), but got {rate}" # read the raw data data = tsdb.load("physionet_2012") all_features = set(data["set-a"].columns) data["static_features"].remove("ICUType") # keep ICUType for now if subset != "all": df = data[subset] X = df.reset_index(drop=True) unique_ids = df["RecordID"].unique() y = data[f"outcomes-{subset.split('-')[-1]}"] y = y.loc[unique_ids] else: df = pd.concat([data["set-a"], data["set-b"], data["set-c"]], sort=True) X = df.reset_index(drop=True) unique_ids = df["RecordID"].unique() y = pd.concat([data["outcomes-a"], data["outcomes-b"], data["outcomes-c"]]) y = y.loc[unique_ids] if features is None: # if features are not specified, we use all features except the static features, e.g. age X = X.drop(data["static_features"], axis=1) else: # if features are specified by users, only use the specified features # check if the given features are valid features_set = set(features) if not all_features.issuperset(features_set): intersection_feats = all_features.intersection(features_set) difference = features_set.difference(intersection_feats) raise ValueError(f"Given features contain invalid features that not in the dataset: {difference}") # check if the given features contain necessary features for preprocessing if "RecordID" not in features: features.append("RecordID") if "ICUType" not in features: features.append("ICUType") if "Time" not in features: features.append("Time") # select the specified features finally X = X[features] X = X.groupby("RecordID").apply(apply_func) # pandas versions differ on whether group keys are kept as columns after groupby-apply. X = X.drop(columns=["RecordID"], errors="ignore") X = X.reset_index() ICUType = X[["RecordID", "ICUType"]].set_index("RecordID").dropna() X = X.drop(["level_1", "ICUType"], axis=1) # PhysioNet2012 is an imbalanced dataset, hence, we separate positive and negative samples here for later splitting # This is to ensure positive and negative ratios are similar in train/val/test sets # Cast to numpy array for sklearn compatibility when pandas returns extension arrays (e.g., pyarrow-backed). all_recordID = np.asarray(X["RecordID"].unique()) positive = (y == 1)["In-hospital_death"] positive_sample_IDs = y.loc[positive].index.to_list() negative_sample_IDs = np.setxor1d(all_recordID, positive_sample_IDs) assert len(positive_sample_IDs) + len(negative_sample_IDs) == len(all_recordID) # split the dataset into the train, val, and test sets train_positive_set_ids, test_positive_set_ids = train_test_split( positive_sample_IDs, test_size=0.2, random_state=random_state ) train_positive_set_ids, val_positive_set_ids = train_test_split( train_positive_set_ids, test_size=0.2, random_state=random_state ) train_negative_set_ids, test_negative_set_ids = train_test_split( negative_sample_IDs, test_size=0.2, random_state=random_state ) train_negative_set_ids, val_negative_set_ids = train_test_split( train_negative_set_ids, test_size=0.2, random_state=random_state ) train_set_ids = np.concatenate([train_positive_set_ids, train_negative_set_ids]) val_set_ids = np.concatenate([val_positive_set_ids, val_negative_set_ids]) test_set_ids = np.concatenate([test_positive_set_ids, test_negative_set_ids]) train_set_ids.sort() val_set_ids.sort() test_set_ids.sort() train_set = X[X["RecordID"].isin(train_set_ids)].sort_values(["RecordID", "Time"]) val_set = X[X["RecordID"].isin(val_set_ids)].sort_values(["RecordID", "Time"]) test_set = X[X["RecordID"].isin(test_set_ids)].sort_values(["RecordID", "Time"]) # remove useless columns and turn into numpy arrays train_set = train_set.drop(["RecordID", "Time"], axis=1) val_set = val_set.drop(["RecordID", "Time"], axis=1) test_set = test_set.drop(["RecordID", "Time"], axis=1) train_X, val_X, test_X = ( train_set.to_numpy(), val_set.to_numpy(), test_set.to_numpy(), ) # normalization scaler = StandardScaler() train_X = scaler.fit_transform(train_X) val_X = scaler.transform(val_X) test_X = scaler.transform(test_X) # reshape into time series samples train_X = train_X.reshape(len(train_set_ids), 48, -1) val_X = val_X.reshape(len(val_set_ids), 48, -1) test_X = test_X.reshape(len(test_set_ids), 48, -1) # fetch labels for train/val/test sets train_y = y[y.index.isin(train_set_ids)].sort_index() val_y = y[y.index.isin(val_set_ids)].sort_index() test_y = y[y.index.isin(test_set_ids)].sort_index() train_y, val_y, test_y = train_y.to_numpy(), val_y.to_numpy(), test_y.to_numpy() # fetch ICUType for train/val/test sets train_ICUType = ICUType[ICUType.index.isin(train_set_ids)].sort_index() val_ICUType = ICUType[ICUType.index.isin(val_set_ids)].sort_index() test_ICUType = ICUType[ICUType.index.isin(test_set_ids)].sort_index() train_ICUType, val_ICUType, test_ICUType = ( train_ICUType.to_numpy(), val_ICUType.to_numpy(), test_ICUType.to_numpy(), ) # assemble the final processed data into a dictionary processed_dataset = { # general info "n_classes": 2, "n_steps": 48, "n_features": train_X.shape[-1], "scaler": scaler, # train set "train_X": train_X, "train_y": train_y.flatten(), "train_ICUType": train_ICUType.flatten(), # val set "val_X": val_X, "val_y": val_y.flatten(), "val_ICUType": val_ICUType.flatten(), # test set "test_X": test_X, "test_y": test_y.flatten(), "test_ICUType": test_ICUType.flatten(), } processed_dataset = convert_processed_dataset_by_task_type( processed_dataset, task_type=task_type, n_pred_steps=n_pred_steps, forecast_feature_indices=forecast_feature_indices, ) if rate > 0: logger.warning( "Note that physionet_2012 has sparse observations in the time series, " "hence we don't add additional missing values to the training dataset. " ) # hold out ground truth in the original data for evaluation val_X_ori = processed_dataset["val_X"] test_X_ori = processed_dataset["test_X"] # mask values in the validation set as ground truth val_X = create_missingness(processed_dataset["val_X"], rate, pattern, **kwargs) # mask values in the test set as ground truth test_X = create_missingness(processed_dataset["test_X"], rate, pattern, **kwargs) processed_dataset["val_X"] = val_X processed_dataset["val_X_ori"] = val_X_ori val_X_indicating_mask = np.isnan(val_X_ori) ^ np.isnan(val_X) logger.info( f"{val_X_indicating_mask.sum()} values masked out in the val set as ground truth, " f"take {val_X_indicating_mask.sum() / (~np.isnan(val_X_ori)).sum():.2%} of the original observed values" ) processed_dataset["test_X"] = test_X processed_dataset["test_X_ori"] = test_X_ori test_X_indicating_mask = np.isnan(test_X_ori) ^ np.isnan(test_X) logger.info( f"{test_X_indicating_mask.sum()} values masked out in the test set as ground truth, " f"take {test_X_indicating_mask.sum() / (~np.isnan(test_X_ori)).sum():.2%} of the original observed values" ) else: logger.warning("rate is 0, no missing values are artificially added.") print_final_dataset_info(processed_dataset) return processed_dataset