def clean_epic_notes(EPIC,
EPIC_cc,
notes_cols,
data_path,
save_path=None,
save_index=False):
'''
Fill in missing notes by CC and clean notes to fit the BERT format.
Input : EPIC = [DataFrame] Data containing notes.
EPIC_cc = [DataFrame] Must contain "CC".
notes_cols = [list] Column name of the notes to be cleaned.
data_path = [str] path to the raw EPIC data (must contain notes).
save_path = [str] path to save the cleaned notes.
save_index = [bool] whether to save the index of the cleaned notes.
Output:
EPIC with notes cleaned.
'''
# Loop over each file and write to a csv
print("\nStart cleaning notes ...")
# Clean text
for col in notes_cols:
print("Cleaning {}".format(col))
EPIC.loc[:, col] = list(map(clean_text, EPIC[col]))
# Save data
EPIC.to_csv(save_path + 'EPIC_triage.csv', index=False)
# Load data nonetheless to convert empty notes "" to nan
EPIC = pd.read_csv(save_path + 'EPIC_triage.csv')
# Fill in missing vals
EPIC = fill_missing_text(EPIC, EPIC_cc, notes_cols)
# Save imputed text
EPIC.to_csv(save_path + 'EPIC_triage.csv', index=False)
# Further preprocessing
preprocessor = EPICPreprocess.Preprocess(data_path)
_, _, _, EPIC_arrival = preprocessor.streamline()
# Remove the obvious outliers
EPIC = EPIC.loc[EPIC_arrival.index, :]
# Add time variable
EPIC = pd.concat([EPIC, EPIC_arrival["Arrived"].astype(int)], axis=1)
# Get time span
time_span = EPIC['Arrived'].unique().tolist()
# Save data
if save_path is not None:
pickle.dump(time_span, open(save_path + "time_span", "wb"))
if not save_index:
EPIC.to_csv(save_path + 'EPIC.csv', index=False)
else:
EPIC.to_csv(save_path + 'EPIC.csv', index=True)
return EPIC
drop_cols = [
'Distance.To.Sick.Kids', 'Distance.To.Walkin', 'Distance.To.Hospital',
'Systolic', 'First.ED.Provider', 'Last.ED.Provider',
'ED.Longest.Attending.ED.Provider'
]
drop_cols = ['Name.Of.Walkin', 'Day.of.Arrival']
drop_cols = [
'First.ED.Provider', 'Last.ED.Provider',
'ED.Longest.Attending.ED.Provider', 'Admitting.Provider',
'Name.Of.Hospital'
]
rm_features = ['Systolic', 'Day.of.Arrival_Monday', 'Gender_M']
rm_features = []
preprocessor = EPICPreprocess.Preprocess(DATA_PATH, drop_cols=drop_cols)
EPIC, EPIC_enc, EPIC_CUI, EPIC_arrival = preprocessor.streamline()
# Get numerical columns (for later transformation)
num_cols = preprocessor.which_numerical(EPIC)
num_cols.remove("Primary.Dx")
# Get time span
time_span = EPIC_arrival['Arrived'].unique().tolist()
# ----------------------------------------------------
# ========= 2.a. One-month ahead prediction =========
print("====================================")
print("Dynamically evaluate the model ...\n")
j = 0
# Path set-up
FIG_PATH = "../../results/neural_net/"
DATA_PATH = "../../data/EPIC_DATA/preprocessed_EPIC_with_dates_and_notes.csv"
RAW_DATA_PATH = "../../data/EPIC_DATA/EPIC.csv"
# Create folder if not already exist
if not os.path.exists(FIG_PATH):
os.makedirs(FIG_PATH)
# ----------------------------------------------------
# ========= 1.i. Further preprocessing =========
preprocessor = EPICPreprocess.Preprocess(DATA_PATH)
EPIC, EPIC_enc, EPIC_CUI, EPIC_arrival = preprocessor.streamline()
# Get numerical columns (for later transformation)
num_cols = preprocessor.which_numerical(EPIC)
num_cols.remove("Primary.Dx")
# Get time span
time_span = EPIC_arrival['Arrived'].unique().tolist()
# ----------------------------------------------------
# ========= 1.ii. Append arrival date =========
EPIC_raw = pd.read_csv(RAW_DATA_PATH, encoding = "ISO-8859-1")
date = pd.to_datetime(EPIC_raw["Arrived"]).loc[EPIC_arrival.index]
# Change name to differentiate from Arrived
RAW_TEXT_PATH = "../../data/EPIC_DATA/EPIC.csv"
RAW_SAVE_DIR = FIG_PATH + "Raw_Notes/"
# This is where BERT will look for pre-trained models to load parameters from.
CACHE_DIR = '../../ClinicalBert/pretrained_bert_tf/biobert_pretrain_output_all_notes_150000/'
# ----------------------------------------------------
# Create folder to save raw text data if not exist
if not os.path.exists(RAW_SAVE_DIR):
os.makedirs(RAW_SAVE_DIR)
# ----------------------------------------------------
# Prepare train and test sets
# Load file
EPIC_original = pd.read_csv(RAW_TEXT_PATH, encoding="ISO-8859-1")
preprocessor = EPICPreprocess.Preprocess(path=RAW_TEXT_PATH)
EPIC_original = preprocessor.BinarizeSepsis(EPIC_original)
# Only keep text columns and label
notes_cols = ["Note.Data_ED.Triage.Notes"]
EPIC = EPIC_original[["Primary.Dx"] + notes_cols]
# ----------------------------------------------------
# ========= 1. Further preprocessing =========
# Clean the file if not already done
if not os.path.exists(RAW_SAVE_DIR + "EPIC.csv"):
_ = clean_epic_notes(EPIC=EPIC,
EPIC_cc=EPIC_original,
notes_cols=notes_cols,
data_path=DATA_PATH,
save_path=RAW_SAVE_DIR)
# parser = setup_parser()
# args = parser.parse_args()
# ----------------------------------------------------
# Path to save figures
fig_path = "/".join(
os.getcwd().split("/")[:3]) + "/Pictures/logistic_regression/"
data_path = "/home/xingliu/Documents/ED/data/EPIC_DATA/preprocessed_EPIC_with_dates_and_notes.csv"
# Create folder if not already exist
if not os.path.exists(fig_path):
os.makedirs(fig_path)
# ----------------------------------------------------
# ========= 1. Further preprocessing =========
preprocessor = EPICPreprocess.Preprocess(data_path)
EPIC, EPIC_enc, EPIC_CUI, EPIC_arrival = preprocessor.streamline()
# Get numerical columns (for later transformation)
num_cols = preprocessor.which_numerical(EPIC)
num_cols.remove("Primary.Dx")
num_cols.remove("Will.Return")
# Get time span
time_span = EPIC_arrival['Arrived'].unique().tolist()
# ----------------------------------------------------
# ========= 2.a. One-month ahead prediction =========
print("====================================")
print("Dynamically evaluate the model ...\n")
def prepare_data(data_path,
mode,
random_seed,
validation_size=0,
clean_notes=True):
RANDOM_SEED = random_seed
MODE = mode
# MODE = "a"
VALID_SIZE = validation_size
# Path set-up
# FIG_PATH = "../../../results/stacked_model/"
FIG_PATH = data_path
RAW_TEXT_PATH = "../../../data/EPIC_DATA/EPIC.csv"
DATA_PATH = "../../../data/EPIC_DATA/preprocessed_EPIC_with_dates_and_notes.csv"
FIG_ROOT_PATH = FIG_PATH + f"dynamic_{MODE}/"
RAW_SAVE_DIR = FIG_PATH + "Raw_Notes/"
CLEAN_NOTES = clean_notes
# Create folder if not already exist
if not os.path.exists(FIG_PATH):
os.makedirs(FIG_PATH)
# ----------------------------------------------------
# ========= 1.i. Further preprocessing =========
# Create folder if not already exist
if not os.path.exists(RAW_SAVE_DIR):
os.makedirs(RAW_SAVE_DIR)
preprocessor = EPICPreprocess.Preprocess(DATA_PATH)
EPIC, EPIC_enc, EPIC_CUI, EPIC_arrival = preprocessor.streamline()
# Get numerical columns (for later transformation)
num_cols = preprocessor.which_numerical(EPIC)
num_cols.remove("Primary.Dx")
# Get time span
time_span = EPIC_arrival['Arrived'].unique().tolist()
# ========= 1.ii. Clean text data =========
# Text data
EPIC_original = pd.read_csv(RAW_TEXT_PATH, encoding='ISO-8859-1')
preprocessor = EPICPreprocess.Preprocess(path=RAW_TEXT_PATH)
EPIC_original = preprocessor.BinarizeSepsis(EPIC_original)
# Only keep text columns and label
notes_cols = ['Note.Data_ED.Triage.Notes']
EPIC = EPIC_original[['Primary.Dx'] + notes_cols]
if CLEAN_NOTES:
# Clean texts
EPIC_text = clean_epic_notes(EPIC=EPIC,
EPIC_cc=EPIC_original,
notes_cols=notes_cols,
data_path=DATA_PATH,
save_path=RAW_SAVE_DIR,
save_index=True)
print("Cleaned text saved to {}".format(RAW_SAVE_DIR))
else:
# Load data
print("Loading cleaned text from {}".format(RAW_SAVE_DIR))
EPIC_text = pd.read_csv(RAW_SAVE_DIR + "EPIC.csv")
# Assign index back
EPIC_text.index = EPIC_text.iloc[:, 0]
EPIC_text = EPIC_text.drop(EPIC_text.columns[0], axis=1)
time_span = pickle.load(open(RAW_SAVE_DIR + "time_span", "rb"))
discrepancy = (EPIC_enc.index != EPIC_text.index).sum()
if discrepancy != 0:
raise Warning(
"EPIC numerics and text data do not match! Number of unmatched cases: {}"
.format(discrepancy))
# ========= 1.iii. Prepare train/test/validation sets =========
# Splitting data by month
for j, time in enumerate(time_span[2:-1]):
# Month to be predicted
time_pred = time_span[j + 3]
# Create folder if not already exist
DYNAMIC_PATH = FIG_ROOT_PATH + f"{time_pred}/"
NUMERICS_DATA_PATH = DYNAMIC_PATH + "numerical_data/"
print(NUMERICS_DATA_PATH)
for path in [DYNAMIC_PATH, NUMERICS_DATA_PATH]:
if not os.path.exists(path):
os.makedirs(path)
# Create BERT folder if not already exist
# FIG_ROOT_PATH = FIG_ROOT_PATH + f"{TASK_NAME}/"
OUTPUT_DIR = DYNAMIC_PATH + f'Saved_Checkpoints/'
REPORTS_DIR = DYNAMIC_PATH + "Reports/"
PROCESSED_NOTES_DIR = DYNAMIC_PATH + "Processed_Texts/"
for path in [
DYNAMIC_PATH, OUTPUT_DIR, REPORTS_DIR, PROCESSED_NOTES_DIR
]:
if not os.path.exists(path):
os.makedirs(path)
# Valid set for the first 3 months
if j == 0:
# Prepare train/test/valid sets
# Not prepare validation set if required
if VALID_SIZE == 0:
XTrain, XTest, yTrain, yTest = splitter(EPIC_arrival,
num_cols,
MODE,
time_threshold=time,
test_size=None,
EPIC_CUI=EPIC_CUI,
seed=RANDOM_SEED)
print("Saving data up to {} ...".format(time))
print(
"Train size: {}. Test size: {}. Sepsis cases in [train, test]: [{}, {}]."
.format(yTrain.shape, yTest.shape, yTrain.sum(),
yTest.sum()))
else:
XTrain, XTest, XValid, yTrain, yTest, yValid = splitter(
EPIC_arrival,
num_cols,
MODE,
time_threshold=time,
test_size=None,
valid_size=VALID_SIZE,
EPIC_CUI=EPIC_CUI,
seed=RANDOM_SEED)
print("Saving data up to {} ...".format(time))
print(
"Train size: {}. Test size: {}. Validation size: {}. Sepsis cases in [train, test, valid]: [{}, {}, {}]."
.format(yTrain.shape, yTest.shape, len(yValid),
yTrain.sum(), yTest.sum(), yValid.sum()))
# Get validation index
valid_index = XValid.index
# Get text data
XValidText = EPIC_text.loc[valid_index, :]
valid_bert = create_bert_data(
x_data=XValidText["Note.Data_ED.Triage.Notes"],
y_data=yValid,
save_path=PROCESSED_NOTES_DIR + "valid.tsv")
# Save numerics data
XValid.to_csv(NUMERICS_DATA_PATH + "x_valid.csv", index=False)
yValid.to_csv(NUMERICS_DATA_PATH + "y_valid.csv",
index=False,
header=True)
# Labels for the text set
yTrainText = yTrain
else:
XTrain, XTest, yTrain, yTest = splitter(EPIC_arrival,
num_cols,
MODE,
time_threshold=time,
test_size=None,
EPIC_CUI=EPIC_CUI,
seed=RANDOM_SEED)
print("Saving data up to {} ...".format(time))
print(
"Train size: {}. Test size: {}. Sepsis cases in [train, test]: [{}, {}]."
.format(yTrain.shape, yTest.shape, yTrain.sum(), yTest.sum()))
# Set text train data to the previous month
XTrainText = XTrainTextOld
yTrainText = yTrainTextOld
# Save train and test sets
train_index = XTrain.index
test_index = XTest.index
XTrainText = EPIC_text.loc[train_index, :]
XTestText = EPIC_text.loc[test_index, :]
# Save text data
train_bert = create_bert_data(
x_data=XTrainText["Note.Data_ED.Triage.Notes"],
y_data=yTrainText,
save_path=PROCESSED_NOTES_DIR + "train.tsv")
test_bert = create_bert_data(
x_data=XTestText["Note.Data_ED.Triage.Notes"],
y_data=yTest,
save_path=PROCESSED_NOTES_DIR + "dev.tsv")
# Save numerics data
XTrain.to_csv(NUMERICS_DATA_PATH + "x_train.csv", index=False)
yTrain.to_csv(NUMERICS_DATA_PATH + "y_train.csv",
index=False,
header=True)
XTest.to_csv(NUMERICS_DATA_PATH + "x_test.csv", index=False)
yTest.to_csv(NUMERICS_DATA_PATH + "y_test.csv",
index=False,
header=True)
# Only store text set of the previous month to save time in training the BERT
XTrainTextOld = XTestText
yTrainTextOld = yTest
CLEAN_NOTES = False
# ----------------------------------------------------
# Create folder to save raw text data if not exist
if not os.path.exists(RAW_SAVE_DIR):
os.makedirs(RAW_SAVE_DIR)
# ----------------------------------------------------
# Prepare train and test sets
EPIC_original = pd.read_csv(TEXT_DATA_PATH, encoding = 'ISO-8859-1')
preprocessor = EPICPreprocess.Preprocess(path = TEXT_DATA_PATH)
EPIC_original = preprocessor.BinarizeSepsis(EPIC_original)
# Only keep text columns and label
notesCols = ['Note.Data_ED.Triage.Notes']
EPIC = EPIC_original[['Primary.Dx'] + notesCols]
# ----------------------------------------------------
# ========= 1. Further preprocessing =========
if CLEAN_NOTES:
# Loop over each file and write to a csv
print("\nStart cleaning notes ...")
# Clean text
