Mimic4Dataset / day_intervals_cohort_v22.py
thbndi's picture
Update day_intervals_cohort_v22.py
e2dd7cc
raw
history blame
18.1 kB
import datetime
import os
import sys
import numpy as np
import pandas as pd
from pathlib import Path
from tqdm import tqdm
import importlib
import disease_cohort
importlib.reload(disease_cohort)
import disease_cohort
sys.path.append(os.path.dirname(os.path.abspath(__file__)) + './../..')
if not os.path.exists("./data/cohort"):
os.makedirs("./data/cohort")
if not os.path.exists("./data/summary"):
os.makedirs("./data/summary")
def get_visit_pts(mimic4_path:str, group_col:str, visit_col:str, admit_col:str, disch_col:str, adm_visit_col:str, use_mort:bool, use_los:bool, los:int, use_admn:bool, disease_label:str,use_ICU:bool):
"""Combines the MIMIC-IV core/patients table information with either the icu/icustays or core/admissions data.
Parameters:
mimic4_path: path to mimic-iv folder containing MIMIC-IV data
group_col: patient identifier to group patients (normally subject_id)
visit_col: visit identifier for individual patient visits (normally hadm_id or stay_id)
admit_col: column for visit start date information (normally admittime or intime)
disch_col: column for visit end date information (normally dischtime or outtime)
use_ICU: describes whether to speficially look at ICU visits in icu/icustays OR look at general admissions from core/admissions
"""
visit = None # df containing visit information depending on using ICU or not
if use_ICU:
visit = pd.read_csv(mimic4_path + "icu/icustays.csv.gz", compression='gzip', header=0, index_col=None, parse_dates=[admit_col, disch_col])
if use_admn:
# icustays doesn't have a way to identify if patient died during visit; must
# use core/patients to remove such stay_ids for readmission labels
pts = pd.read_csv(mimic4_path + "hosp/patients.csv.gz", compression='gzip', header=0, index_col=None, usecols=['subject_id', 'dod'], parse_dates=['dod'])
visit = visit.merge(pts, how='inner', left_on='subject_id', right_on='subject_id')
visit = visit.loc[(visit.dod.isna()) | (visit.dod >= visit[disch_col])]
if len(disease_label):
hids=disease_cohort.extract_diag_cohort(visit['hadm_id'],disease_label,mimic4_path)
visit=visit[visit['hadm_id'].isin(hids['hadm_id'])]
print("[ READMISSION DUE TO "+disease_label+" ]")
else:
visit = pd.read_csv(mimic4_path + "hosp/admissions.csv.gz", compression='gzip', header=0, index_col=None, parse_dates=[admit_col, disch_col])
visit['los']=visit[disch_col]-visit[admit_col]
visit[admit_col] = pd.to_datetime(visit[admit_col])
visit[disch_col] = pd.to_datetime(visit[disch_col])
visit['los']=pd.to_timedelta(visit[disch_col]-visit[admit_col],unit='h')
visit['los']=visit['los'].astype(str)
visit[['days', 'dummy','hours']] = visit['los'].str.split(' ', -1, expand=True)
visit['los']=pd.to_numeric(visit['days'])
visit=visit.drop(columns=['days', 'dummy','hours'])
if use_admn:
# remove hospitalizations with a death; impossible for readmission for such visits
visit = visit.loc[visit.hospital_expire_flag == 0]
if len(disease_label):
hids=disease_cohort.extract_diag_cohort(visit['hadm_id'],disease_label,mimic4_path)
visit=visit[visit['hadm_id'].isin(hids['hadm_id'])]
print("[ READMISSION DUE TO "+disease_label+" ]")
pts = pd.read_csv(
mimic4_path + "hosp/patients.csv.gz", compression='gzip', header=0, index_col = None, usecols=[group_col, 'anchor_year', 'anchor_age', 'anchor_year_group', 'dod','gender']
)
pts['yob']= pts['anchor_year'] - pts['anchor_age'] # get yob to ensure a given visit is from an adult
pts['min_valid_year'] = pts['anchor_year'] + (2019 - pts['anchor_year_group'].str.slice(start=-4).astype(int))
# Define anchor_year corresponding to the anchor_year_group 2017-2019. This is later used to prevent consideration
# of visits with prediction windows outside the dataset's time range (2008-2019)
#[[group_col, visit_col, admit_col, disch_col]]
if use_ICU:
visit_pts = visit[[group_col, visit_col, adm_visit_col, admit_col, disch_col,'los']].merge(
pts[[group_col, 'anchor_year', 'anchor_age', 'yob', 'min_valid_year', 'dod','gender']], how='inner', left_on=group_col, right_on=group_col
)
else:
visit_pts = visit[[group_col, visit_col, admit_col, disch_col,'los']].merge(
pts[[group_col, 'anchor_year', 'anchor_age', 'yob', 'min_valid_year', 'dod','gender']], how='inner', left_on=group_col, right_on=group_col
)
# only take adult patients
# visit_pts['Age']=visit_pts[admit_col].dt.year - visit_pts['yob']
# visit_pts = visit_pts.loc[visit_pts['Age'] >= 18]
visit_pts['Age']=visit_pts['anchor_age']
visit_pts = visit_pts.loc[visit_pts['Age'] >= 18]
##Add Demo data
eth = pd.read_csv(mimic4_path + "hosp/admissions.csv.gz", compression='gzip', header=0, usecols=['hadm_id', 'insurance','race'], index_col=None)
visit_pts= visit_pts.merge(eth, how='inner', left_on='hadm_id', right_on='hadm_id')
if use_ICU:
return visit_pts[[group_col, visit_col, adm_visit_col, admit_col, disch_col,'los', 'min_valid_year', 'dod','Age','gender','race', 'insurance']]
else:
return visit_pts.dropna(subset=['min_valid_year'])[[group_col, visit_col, admit_col, disch_col,'los', 'min_valid_year', 'dod','Age','gender','race', 'insurance']]
def validate_row(row, ctrl, invalid, max_year, disch_col, valid_col, gap):
"""Checks if visit's prediction window potentially extends beyond the dataset range (2008-2019).
An 'invalid row' is NOT guaranteed to be outside the range, only potentially outside due to
de-identification of MIMIC-IV being done through 3-year time ranges.
To be invalid, the end of the prediction window's year must both extend beyond the maximum seen year
for a patient AND beyond the year that corresponds to the 2017-2019 anchor year range for a patient"""
print("disch_col",row[disch_col])
print(gap)
pred_year = (row[disch_col] + gap).year
if max_year < pred_year and pred_year > row[valid_col]:
invalid = invalid.append(row)
else:
ctrl = ctrl.append(row)
return ctrl, invalid
def partition_by_los(df:pd.DataFrame, los:int, group_col:str, visit_col:str, admit_col:str, disch_col:str, valid_col:str):
invalid = df.loc[(df[admit_col].isna()) | (df[disch_col].isna()) | (df['los'].isna())]
cohort = df.loc[(~df[admit_col].isna()) & (~df[disch_col].isna()) & (~df['los'].isna())]
#cohort=cohort.fillna(0)
pos_cohort=cohort[cohort['los']>los]
neg_cohort=cohort[cohort['los']<=los]
neg_cohort=neg_cohort.fillna(0)
pos_cohort=pos_cohort.fillna(0)
pos_cohort['label']=1
neg_cohort['label']=0
cohort=pd.concat([pos_cohort,neg_cohort], axis=0)
cohort=cohort.sort_values(by=[group_col,admit_col])
#print("cohort",cohort.shape)
print("[ LOS LABELS FINISHED ]")
return cohort, invalid
def partition_by_readmit(df:pd.DataFrame, gap:datetime.timedelta, group_col:str, visit_col:str, admit_col:str, disch_col:str, valid_col:str):
"""Applies labels to individual visits according to whether or not a readmission has occurred within the specified `gap` days.
For a given visit, another visit must occur within the gap window for a positive readmission label.
The gap window starts from the disch_col time and the admit_col of subsequent visits are considered."""
case = pd.DataFrame() # hadm_ids with readmission within the gap period
ctrl = pd.DataFrame() # hadm_ids without readmission within the gap period
invalid = pd.DataFrame() # hadm_ids that are not considered in the cohort
# Iterate through groupbys based on group_col (subject_id). Data is sorted by subject_id and admit_col (admittime)
# to ensure that the most current hadm_id is last in a group.
#grouped= df[[group_col, visit_col, admit_col, disch_col, valid_col]].sort_values(by=[group_col, admit_col]).groupby(group_col)
grouped= df.sort_values(by=[group_col, admit_col]).groupby(group_col)
for subject, group in tqdm(grouped):
max_year = group.max()[disch_col].year
if group.shape[0] <= 1:
#ctrl, invalid = validate_row(group.iloc[0], ctrl, invalid, max_year, disch_col, valid_col, gap) # A group with 1 row has no readmission; goes to ctrl
ctrl = ctrl.append(group.iloc[0])
else:
for idx in range(group.shape[0]-1):
visit_time = group.iloc[idx][disch_col] # For each index (a unique hadm_id), get its timestamp
if group.loc[
(group[admit_col] > visit_time) & # Readmissions must come AFTER the current timestamp
(group[admit_col] - visit_time <= gap) # Distance between a timestamp and readmission must be within gap
].shape[0] >= 1: # If ANY rows meet above requirements, a readmission has occurred after that visit
case = case.append(group.iloc[idx])
else:
# If no readmission is found, only add to ctrl if prediction window is guaranteed to be within the
# time range of the dataset (2008-2019). Visits with prediction windows existing in potentially out-of-range
# dates (like 2018-2020) are excluded UNLESS the prediction window takes place the same year as the visit,
# in which case it is guaranteed to be within 2008-2019
ctrl = ctrl.append(group.iloc[idx])
#ctrl, invalid = validate_row(group.iloc[-1], ctrl, invalid, max_year, disch_col, valid_col, gap) # The last hadm_id datewise is guaranteed to have no readmission logically
ctrl = ctrl.append(group.iloc[-1])
#print(f"[ {gap.days} DAYS ] {case.shape[0] + ctrl.shape[0]}/{df.shape[0]} {visit_col}s processed")
print("[ READMISSION LABELS FINISHED ]")
return case, ctrl, invalid
def partition_by_mort(df:pd.DataFrame, group_col:str, visit_col:str, admit_col:str, disch_col:str, death_col:str):
"""Applies labels to individual visits according to whether or not a death has occurred within
the times of the specified admit_col and disch_col"""
invalid = df.loc[(df[admit_col].isna()) | (df[disch_col].isna())]
cohort = df.loc[(~df[admit_col].isna()) & (~df[disch_col].isna())]
cohort['label']=0
pos_cohort=cohort[~cohort[death_col].isna()]
neg_cohort=cohort[cohort[death_col].isna()]
neg_cohort=neg_cohort.fillna(0)
pos_cohort=pos_cohort.fillna(0)
pos_cohort[death_col] = pd.to_datetime(pos_cohort[death_col])
pos_cohort['label'] = np.where((pos_cohort[death_col] >= pos_cohort[admit_col]) & (pos_cohort[death_col] <= pos_cohort[disch_col]),1,0)
pos_cohort['label'] = pos_cohort['label'].astype("Int32")
cohort=pd.concat([pos_cohort,neg_cohort], axis=0)
cohort=cohort.sort_values(by=[group_col,admit_col])
#print("cohort",cohort.shape)
print("[ MORTALITY LABELS FINISHED ]")
return cohort, invalid
def get_case_ctrls(df:pd.DataFrame, gap:int, group_col:str, visit_col:str, admit_col:str, disch_col:str, valid_col:str, death_col:str, use_mort=False,use_admn=False,use_los=False) -> pd.DataFrame:
"""Handles logic for creating the labelled cohort based on arguments passed to extract().
Parameters:
df: dataframe with patient data
gap: specified time interval gap for readmissions
group_col: patient identifier to group patients (normally subject_id)
visit_col: visit identifier for individual patient visits (normally hadm_id or stay_id)
admit_col: column for visit start date information (normally admittime or intime)
disch_col: column for visit end date information (normally dischtime or outtime)
valid_col: generated column containing a patient's year that corresponds to the 2017-2019 anchor time range
dod_col: Date of death column
"""
case = None # hadm_ids with readmission within the gap period
ctrl = None # hadm_ids without readmission within the gap period
invalid = None # hadm_ids that are not considered in the cohort
if use_mort:
return partition_by_mort(df, group_col, visit_col, admit_col, disch_col, death_col)
elif use_admn:
gap = datetime.timedelta(days=gap)
# transform gap into a timedelta to compare with datetime columns
case, ctrl, invalid = partition_by_readmit(df, gap, group_col, visit_col, admit_col, disch_col, valid_col)
# case hadm_ids are labelled 1 for readmission, ctrls have a 0 label
case['label'] = np.ones(case.shape[0]).astype(int)
ctrl['label'] = np.zeros(ctrl.shape[0]).astype(int)
return pd.concat([case, ctrl], axis=0), invalid
elif use_los:
return partition_by_los(df, gap, group_col, visit_col, admit_col, disch_col, death_col)
# print(f"[ {gap.days} DAYS ] {invalid.shape[0]} hadm_ids are invalid")
def extract_data(use_ICU:str, label:str, time:int, icd_code:str, root_dir,mimic_path, disease_label, cohort_output=None, summary_output=None):
"""Extracts cohort data and summary from MIMIC-IV data based on provided parameters.
Parameters:
cohort_output: name of labelled cohort output file
summary_output: name of summary output file
use_ICU: state whether to use ICU patient data or not
label: Can either be '{day} day Readmission' or 'Mortality', decides what binary data label signifies"""
print("===========MIMIC-IV v2============")
if not cohort_output:
cohort_output="cohort_" + use_ICU.lower() + "_" + label.lower().replace(" ", "_") + "_" + str(time) + "_" + disease_label
if not summary_output:
summary_output="summary_" + use_ICU.lower() + "_" + label.lower().replace(" ", "_") + "_" + str(time) + "_" + disease_label
if icd_code=="No Disease Filter":
if len(disease_label):
print(f"EXTRACTING FOR: | {use_ICU.upper()} | {label.upper()} DUE TO {disease_label.upper()} | {str(time)} | ")
else:
print(f"EXTRACTING FOR: | {use_ICU.upper()} | {label.upper()} | {str(time)} |")
else:
if len(disease_label):
print(f"EXTRACTING FOR: | {use_ICU.upper()} | {label.upper()} DUE TO {disease_label.upper()} | ADMITTED DUE TO {icd_code.upper()} | {str(time)} |")
else:
print(f"EXTRACTING FOR: | {use_ICU.upper()} | {label.upper()} | ADMITTED DUE TO {icd_code.upper()} | {str(time)} |")
#print(label)
cohort, invalid = None, None # final labelled output and df of invalid records, respectively
pts = None # valid patients generated by get_visit_pts based on use_ICU and label
ICU=use_ICU
group_col, visit_col, admit_col, disch_col, death_col, adm_visit_col = "", "", "", "", "", ""
#print(label)
use_mort = label == "Mortality" # change to boolean value
use_admn=label=='Readmission'
los=0
use_los= label=='Length of Stay'
#print(use_mort)
#print(use_admn)
#print(use_los)
if use_los:
los=time
use_ICU = use_ICU == "ICU" # change to boolean value
use_disease=icd_code!="No Disease Filter"
if use_ICU:
group_col='subject_id'
visit_col='stay_id'
admit_col='intime'
disch_col='outtime'
death_col='dod'
adm_visit_col='hadm_id'
else:
group_col='subject_id'
visit_col='hadm_id'
admit_col='admittime'
disch_col='dischtime'
death_col='dod'
pts = get_visit_pts(
mimic4_path=mimic_path,
group_col=group_col,
visit_col=visit_col,
admit_col=admit_col,
disch_col=disch_col,
adm_visit_col=adm_visit_col,
use_mort=use_mort,
use_los=use_los,
los=los,
use_admn=use_admn,
disease_label=disease_label,
use_ICU=use_ICU
)
# cols to be extracted from get_case_ctrls
cols = [group_col, visit_col, admit_col, disch_col, 'Age','gender','ethnicity','insurance','label']
if use_mort:
cols.append(death_col)
cohort, invalid = get_case_ctrls(pts, None, group_col, visit_col, admit_col, disch_col,'min_valid_year', death_col, use_mort=True,use_admn=False,use_los=False)
elif use_admn:
interval = time
cohort, invalid = get_case_ctrls(pts, interval, group_col, visit_col, admit_col, disch_col,'min_valid_year', death_col, use_mort=False,use_admn=True,use_los=False)
elif use_los:
cohort, invalid = get_case_ctrls(pts, los, group_col, visit_col, admit_col, disch_col,'min_valid_year', death_col, use_mort=False,use_admn=False,use_los=True)
#print(cohort.head())
if use_ICU:
cols.append(adm_visit_col)
if use_disease:
hids=disease_cohort.extract_diag_cohort(cohort['hadm_id'],icd_code,mimic_path)
cohort=cohort[cohort['hadm_id'].isin(hids['hadm_id'])]
cohort_output=cohort_output+"_"+icd_code
summary_output=summary_output+"_"+icd_code
#print(cohort[cols].head())
# save output
cohort=cohort.rename(columns={"race":"ethnicity"})
cohort[cols].to_csv("./data/cohort/"+cohort_output+".csv.gz", index=False, compression='gzip')
print("[ COHORT SUCCESSFULLY SAVED ]")
summary = "\n".join([
f"{label} FOR {ICU} DATA",
f"# Admission Records: {cohort.shape[0]}",
f"# Patients: {cohort[group_col].nunique()}",
f"# Positive cases: {cohort[cohort['label']==1].shape[0]}",
f"# Negative cases: {cohort[cohort['label']==0].shape[0]}"
])
# save basic summary of data
with open(f"./data/cohort/{summary_output}.txt", "w") as f:
f.write(summary)
print("[ SUMMARY SUCCESSFULLY SAVED ]")
print(summary)
return cohort_output