Mimic4Dataset / data_generation_modify.py
thbndi's picture
Upload data_generation_modify.py
d198eb0
raw
history blame
21.7 kB
import numpy as np
import pandas as pd
from tqdm import tqdm
from datetime import datetime
import pickle
import datetime
import os
import sys
from pathlib import Path
sys.path.append(os.path.dirname(os.path.abspath(__file__)) + './../..')
if not os.path.exists("./data/dict"):
os.makedirs("./data/dict")
class Generator():
def __init__(self,cohort_output,if_mort,if_admn,if_los,feat_cond,feat_lab,feat_proc,feat_med,impute,include_time=24,bucket=1,predW=0):
self.impute=impute
self.feat_cond,self.feat_proc,self.feat_med,self.feat_lab = feat_cond,feat_proc,feat_med,feat_lab
self.cohort_output=cohort_output
self.data = self.generate_adm()
print("[ READ COHORT ]")
self.generate_feat()
print("[ READ ALL FEATURES ]")
if if_mort:
print(predW)
self.mortality_length(include_time,predW)
print("[ PROCESSED TIME SERIES TO EQUAL LENGTH ]")
elif if_admn:
self.readmission_length(include_time)
print("[ PROCESSED TIME SERIES TO EQUAL LENGTH ]")
elif if_los:
self.los_length(include_time)
print("[ PROCESSED TIME SERIES TO EQUAL LENGTH ]")
self.smooth_meds(bucket)
#if(self.feat_lab):
# print("[ ======READING LABS ]")
# nhid=len(self.hids)
# for n in range(0,nhids,10000):
# self.generate_labs(self.hids[n,n+10000])
print("[ SUCCESSFULLY SAVED DATA DICTIONARIES ]")
def generate_feat(self):
if(self.feat_cond):
print("[ ======READING DIAGNOSIS ]")
self.generate_cond()
if(self.feat_proc):
print("[ ======READING PROCEDURES ]")
self.generate_proc()
if(self.feat_med):
print("[ ======READING MEDICATIONS ]")
self.generate_meds()
if(self.feat_lab):
print("[ ======READING LABS ]")
self.generate_labs()
def generate_adm(self):
data=pd.read_csv(f"./data/cohort/{self.cohort_output}.csv.gz", compression='gzip', header=0, index_col=None)
data['admittime'] = pd.to_datetime(data['admittime'])
data['dischtime'] = pd.to_datetime(data['dischtime'])
data['los']=pd.to_timedelta(data['dischtime']-data['admittime'],unit='h')
data['los']=data['los'].astype(str)
data[['days', 'dummy','hours']] = data['los'].str.split(' ', -1, expand=True)
data[['hours','min','sec']] = data['hours'].str.split(':', -1, expand=True)
data['los']=pd.to_numeric(data['days'])*24+pd.to_numeric(data['hours'])
data=data.drop(columns=['days', 'dummy','hours','min','sec'])
data=data[data['los']>0]
data['Age']=data['Age'].astype(int)
return data
def generate_cond(self):
cond=pd.read_csv("./data/features/preproc_diag.csv.gz", compression='gzip', header=0, index_col=None)
cond=cond[cond['hadm_id'].isin(self.data['hadm_id'])]
cond_per_adm = cond.groupby('hadm_id').size().max()
self.cond, self.cond_per_adm = cond, cond_per_adm
def generate_proc(self):
proc=pd.read_csv("./data/features/preproc_proc.csv.gz", compression='gzip', header=0, index_col=None)
proc=proc[proc['hadm_id'].isin(self.data['hadm_id'])]
proc[['start_days', 'dummy','start_hours']] = proc['proc_time_from_admit'].str.split(' ', -1, expand=True)
proc[['start_hours','min','sec']] = proc['start_hours'].str.split(':', -1, expand=True)
proc['start_time']=pd.to_numeric(proc['start_days'])*24+pd.to_numeric(proc['start_hours'])
proc=proc.drop(columns=['start_days', 'dummy','start_hours','min','sec'])
proc=proc[proc['start_time']>=0]
###Remove where event time is after discharge time
proc=pd.merge(proc,self.data[['hadm_id','los']],on='hadm_id',how='left')
proc['sanity']=proc['los']-proc['start_time']
proc=proc[proc['sanity']>0]
del proc['sanity']
self.proc=proc
def generate_labs(self):
chunksize = 10000000
final=pd.DataFrame()
for labs in tqdm(pd.read_csv("./data/features/preproc_labs.csv.gz", compression='gzip', header=0, index_col=None,chunksize=chunksize)):
labs=labs[labs['hadm_id'].isin(self.data['hadm_id'])]
labs[['start_days', 'dummy','start_hours']] = labs['lab_time_from_admit'].str.split(' ', -1, expand=True)
labs[['start_hours','min','sec']] = labs['start_hours'].str.split(':', -1, expand=True)
labs['start_time']=pd.to_numeric(labs['start_days'])*24+pd.to_numeric(labs['start_hours'])
labs=labs.drop(columns=['start_days', 'dummy','start_hours','min','sec'])
labs=labs[labs['start_time']>=0]
###Remove where event time is after discharge time
labs=pd.merge(labs,self.data[['hadm_id','los']],on='hadm_id',how='left')
labs['sanity']=labs['los']-labs['start_time']
labs=labs[labs['sanity']>0]
del labs['sanity']
if final.empty:
final=labs
else:
final=final.append(labs, ignore_index=True)
self.labs=final
def generate_meds(self):
meds=pd.read_csv("./data/features/preproc_med.csv.gz", compression='gzip', header=0, index_col=None)
meds[['start_days', 'dummy','start_hours']] = meds['start_hours_from_admit'].str.split(' ', -1, expand=True)
meds[['start_hours','min','sec']] = meds['start_hours'].str.split(':', -1, expand=True)
meds['start_time']=pd.to_numeric(meds['start_days'])*24+pd.to_numeric(meds['start_hours'])
meds[['start_days', 'dummy','start_hours']] = meds['stop_hours_from_admit'].str.split(' ', -1, expand=True)
meds[['start_hours','min','sec']] = meds['start_hours'].str.split(':', -1, expand=True)
meds['stop_time']=pd.to_numeric(meds['start_days'])*24+pd.to_numeric(meds['start_hours'])
meds=meds.drop(columns=['start_days', 'dummy','start_hours','min','sec'])
#####Sanity check
meds['sanity']=meds['stop_time']-meds['start_time']
meds=meds[meds['sanity']>0]
del meds['sanity']
#####Select hadm_id as in main file
meds=meds[meds['hadm_id'].isin(self.data['hadm_id'])]
meds=pd.merge(meds,self.data[['hadm_id','los']],on='hadm_id',how='left')
#####Remove where start time is after end of visit
meds['sanity']=meds['los']-meds['start_time']
meds=meds[meds['sanity']>0]
del meds['sanity']
####Any stop_time after end of visit is set at end of visit
meds.loc[meds['stop_time'] > meds['los'],'stop_time']=meds.loc[meds['stop_time'] > meds['los'],'los']
del meds['los']
meds['dose_val_rx']=meds['dose_val_rx'].apply(pd.to_numeric, errors='coerce')
self.meds=meds
def mortality_length(self,include_time,predW):
self.los=include_time
self.data=self.data[(self.data['los']>=include_time+predW)]
self.hids=self.data['hadm_id'].unique()
if(self.feat_cond):
self.cond=self.cond[self.cond['hadm_id'].isin(self.data['hadm_id'])]
self.data['los']=include_time
###MEDS
if(self.feat_med):
self.meds=self.meds[self.meds['hadm_id'].isin(self.data['hadm_id'])]
self.meds=self.meds[self.meds['start_time']<=include_time]
self.meds.loc[self.meds.stop_time >include_time, 'stop_time']=include_time
###PROCS
if(self.feat_proc):
self.proc=self.proc[self.proc['hadm_id'].isin(self.data['hadm_id'])]
self.proc=self.proc[self.proc['start_time']<=include_time]
###LAB
if(self.feat_lab):
self.labs=self.labs[self.labs['hadm_id'].isin(self.data['hadm_id'])]
self.labs=self.labs[self.labs['start_time']<=include_time]
self.los=include_time
def los_length(self,include_time):
self.los=include_time
self.data=self.data[(self.data['los']>=include_time)]
self.hids=self.data['hadm_id'].unique()
if(self.feat_cond):
self.cond=self.cond[self.cond['hadm_id'].isin(self.data['hadm_id'])]
self.data['los']=include_time
###MEDS
if(self.feat_med):
self.meds=self.meds[self.meds['hadm_id'].isin(self.data['hadm_id'])]
self.meds=self.meds[self.meds['start_time']<=include_time]
self.meds.loc[self.meds.stop_time >include_time, 'stop_time']=include_time
###PROCS
if(self.feat_proc):
self.proc=self.proc[self.proc['hadm_id'].isin(self.data['hadm_id'])]
self.proc=self.proc[self.proc['start_time']<=include_time]
###LAB
if(self.feat_lab):
self.labs=self.labs[self.labs['hadm_id'].isin(self.data['hadm_id'])]
self.labs=self.labs[self.labs['start_time']<=include_time]
#self.los=include_time
def readmission_length(self,include_time):
self.los=include_time
self.data=self.data[(self.data['los']>=include_time)]
self.hids=self.data['hadm_id'].unique()
if(self.feat_cond):
self.cond=self.cond[self.cond['hadm_id'].isin(self.data['hadm_id'])]
self.data['select_time']=self.data['los']-include_time
self.data['los']=include_time
####Make equal length input time series and remove data for pred window if needed
###MEDS
if(self.feat_med):
self.meds=self.meds[self.meds['hadm_id'].isin(self.data['hadm_id'])]
self.meds=pd.merge(self.meds,self.data[['hadm_id','select_time']],on='hadm_id',how='left')
self.meds['stop_time']=self.meds['stop_time']-self.meds['select_time']
self.meds['start_time']=self.meds['start_time']-self.meds['select_time']
self.meds=self.meds[self.meds['stop_time']>=0]
self.meds.loc[self.meds.start_time <0, 'start_time']=0
###PROCS
if(self.feat_proc):
self.proc=self.proc[self.proc['hadm_id'].isin(self.data['hadm_id'])]
self.proc=pd.merge(self.proc,self.data[['hadm_id','select_time']],on='hadm_id',how='left')
self.proc['start_time']=self.proc['start_time']-self.proc['select_time']
self.proc=self.proc[self.proc['start_time']>=0]
###LABS
if(self.feat_lab):
self.labs=self.labs[self.labs['hadm_id'].isin(self.data['hadm_id'])]
self.labs=pd.merge(self.labs,self.data[['hadm_id','select_time']],on='hadm_id',how='left')
self.labs['start_time']=self.labs['start_time']-self.labs['select_time']
self.labs=self.labs[self.labs['start_time']>=0]
def smooth_meds(self,bucket):
final_meds=pd.DataFrame()
final_proc=pd.DataFrame()
final_labs=pd.DataFrame()
if(self.feat_med):
self.meds=self.meds.sort_values(by=['start_time'])
if(self.feat_proc):
self.proc=self.proc.sort_values(by=['start_time'])
t=0
for i in tqdm(range(0,self.los,bucket)):
###MEDS
if(self.feat_med):
sub_meds=self.meds[(self.meds['start_time']>=i) & (self.meds['start_time']<i+bucket)].groupby(['hadm_id','drug_name']).agg({'stop_time':'max','subject_id':'max','dose_val_rx':np.nanmean})
sub_meds=sub_meds.reset_index()
sub_meds['start_time']=t
sub_meds['stop_time']=sub_meds['stop_time']/bucket
if final_meds.empty:
final_meds=sub_meds
else:
final_meds=final_meds.append(sub_meds)
###PROC
if(self.feat_proc):
sub_proc=self.proc[(self.proc['start_time']>=i) & (self.proc['start_time']<i+bucket)].groupby(['hadm_id','icd_code']).agg({'subject_id':'max'})
sub_proc=sub_proc.reset_index()
sub_proc['start_time']=t
if final_proc.empty:
final_proc=sub_proc
else:
final_proc=final_proc.append(sub_proc)
###LABS
if(self.feat_lab):
sub_labs=self.labs[(self.labs['start_time']>=i) & (self.labs['start_time']<i+bucket)].groupby(['hadm_id','itemid']).agg({'subject_id':'max','valuenum':np.nanmean})
sub_labs=sub_labs.reset_index()
sub_labs['start_time']=t
if final_labs.empty:
final_labs=sub_labs
else:
final_labs=final_labs.append(sub_labs)
t=t+1
los=int(self.los/bucket)
###MEDS
if(self.feat_med):
f2_meds=final_meds.groupby(['hadm_id','drug_name']).size()
self.med_per_adm=f2_meds.groupby('hadm_id').sum().reset_index()[0].max()
self.medlength_per_adm=final_meds.groupby('hadm_id').size().max()
###PROC
if(self.feat_proc):
f2_proc=final_proc.groupby(['hadm_id','icd_code']).size()
self.proc_per_adm=f2_proc.groupby('hadm_id').sum().reset_index()[0].max()
self.proclength_per_adm=final_proc.groupby('hadm_id').size().max()
###LABS
if(self.feat_lab):
f2_labs=final_labs.groupby(['hadm_id','itemid']).size()
self.labs_per_adm=f2_labs.groupby('hadm_id').sum().reset_index()[0].max()
self.labslength_per_adm=final_labs.groupby('hadm_id').size().max()
###CREATE DICT
print("[ PROCESSED TIME SERIES TO EQUAL TIME INTERVAL ]")
self.create_Dict(final_meds,final_proc,final_labs,los)
def create_Dict(self,meds,proc,labs,los):
print("[ CREATING DATA DICTIONARIES ]")
dataDic={}
labels_csv=pd.DataFrame(columns=['hadm_id','label'])
labels_csv['hadm_id']=pd.Series(self.hids)
labels_csv['label']=0
for hid in self.hids:
grp=self.data[self.data['hadm_id']==hid]
dataDic[hid]={'Cond':{},'Proc':{},'Med':{},'Lab':{},'ethnicity':grp['ethnicity'].iloc[0],'age':int(grp['Age']),'gender':grp['gender'].iloc[0],'label':int(grp['label'])}
for hid in tqdm(self.hids):
grp=self.data[self.data['hadm_id']==hid]
###MEDS
if(self.feat_med):
feat=meds['drug_name'].unique()
df2=meds[meds['hadm_id']==hid]
if df2.shape[0]==0:
val=pd.DataFrame(np.zeros([los,len(feat)]),columns=feat)
val=val.fillna(0)
val.columns=pd.MultiIndex.from_product([["MEDS"], val.columns])
else:
val=df2.pivot_table(index='start_time',columns='drug_name',values='dose_val_rx')
df2=df2.pivot_table(index='start_time',columns='drug_name',values='stop_time')
#print(df2.shape)
add_indices = pd.Index(range(los)).difference(df2.index)
add_df = pd.DataFrame(index=add_indices, columns=df2.columns).fillna(np.nan)
df2=pd.concat([df2, add_df])
df2=df2.sort_index()
df2=df2.ffill()
df2=df2.fillna(0)
val=pd.concat([val, add_df])
val=val.sort_index()
val=val.ffill()
val=val.fillna(-1)
#print(df2.head())
df2.iloc[:,0:]=df2.iloc[:,0:].sub(df2.index,0)
df2[df2>0]=1
df2[df2<0]=0
val.iloc[:,0:]=df2.iloc[:,0:]*val.iloc[:,0:]
#print(df2.head())
dataDic[hid]['Med']['signal']=df2.iloc[:,0:].to_dict(orient="list")
dataDic[hid]['Med']['val']=val.iloc[:,0:].to_dict(orient="list")
###PROCS
if(self.feat_proc):
feat=proc['icd_code'].unique()
df2=proc[proc['hadm_id']==hid]
if df2.shape[0]==0:
df2=pd.DataFrame(np.zeros([los,len(feat)]),columns=feat)
df2=df2.fillna(0)
df2.columns=pd.MultiIndex.from_product([["PROC"], df2.columns])
else:
df2['val']=1
df2=df2.pivot_table(index='start_time',columns='icd_code',values='val')
#print(df2.shape)
add_indices = pd.Index(range(los)).difference(df2.index)
add_df = pd.DataFrame(index=add_indices, columns=df2.columns).fillna(np.nan)
df2=pd.concat([df2, add_df])
df2=df2.sort_index()
df2=df2.fillna(0)
df2[df2>0]=1
#print(df2.head())
dataDic[hid]['Proc']=df2.to_dict(orient="list")
###LABS
if(self.feat_lab):
feat=labs['itemid'].unique()
df2=labs[labs['hadm_id']==hid]
if df2.shape[0]==0:
val=pd.DataFrame(np.zeros([los,len(feat)]),columns=feat)
val=val.fillna(0)
val.columns=pd.MultiIndex.from_product([["LAB"], val.columns])
else:
val=df2.pivot_table(index='start_time',columns='itemid',values='valuenum')
df2['val']=1
df2=df2.pivot_table(index='start_time',columns='itemid',values='val')
#print(df2.shape)
add_indices = pd.Index(range(los)).difference(df2.index)
add_df = pd.DataFrame(index=add_indices, columns=df2.columns).fillna(np.nan)
df2=pd.concat([df2, add_df])
df2=df2.sort_index()
df2=df2.fillna(0)
val=pd.concat([val, add_df])
val=val.sort_index()
if self.impute=='Mean':
val=val.ffill()
val=val.bfill()
val=val.fillna(val.mean())
elif self.impute=='Median':
val=val.ffill()
val=val.bfill()
val=val.fillna(val.median())
val=val.fillna(0)
df2[df2>0]=1
df2[df2<0]=0
#print(df2.head())
dataDic[hid]['Lab']['signal']=df2.iloc[:,0:].to_dict(orient="list")
dataDic[hid]['Lab']['val']=val.iloc[:,0:].to_dict(orient="list")
##########COND#########
if(self.feat_cond):
feat=self.cond['new_icd_code'].unique()
grp=self.cond[self.cond['hadm_id']==hid]
if(grp.shape[0]==0):
dataDic[hid]['Cond']={'fids':list(['<PAD>'])}
else:
dataDic[hid]['Cond']={'fids':list(grp['new_icd_code'])}
######SAVE DICTIONARIES##############
metaDic={'Cond':{},'Proc':{},'Med':{},'Lab':{},'LOS':{}}
metaDic['LOS']=los
with open("./data/dict/dataDic", 'wb') as fp:
pickle.dump(dataDic, fp)
with open("./data/dict/hadmDic", 'wb') as fp:
pickle.dump(self.hids, fp)
with open("./data/dict/ethVocab", 'wb') as fp:
pickle.dump(list(self.data['ethnicity'].unique()), fp)
self.eth_vocab = self.data['ethnicity'].nunique()
with open("./data/dict/ageVocab", 'wb') as fp:
pickle.dump(list(self.data['Age'].unique()), fp)
self.age_vocab = self.data['Age'].nunique()
with open("./data/dict/insVocab", 'wb') as fp:
pickle.dump(list(self.data['insurance'].unique()), fp)
self.ins_vocab = self.data['insurance'].nunique()
if(self.feat_med):
with open("./data/dict/medVocab", 'wb') as fp:
pickle.dump(list(meds['drug_name'].unique()), fp)
self.med_vocab = meds['drug_name'].nunique()
metaDic['Med']=self.med_per_adm
if(self.feat_cond):
with open("./data/dict/condVocab", 'wb') as fp:
pickle.dump(list(self.cond['new_icd_code'].unique()), fp)
self.cond_vocab = self.cond['new_icd_code'].nunique()
metaDic['Cond']=self.cond_per_adm
if(self.feat_proc):
with open("./data/dict/procVocab", 'wb') as fp:
pickle.dump(list(proc['icd_code'].unique()), fp)
self.proc_vocab = proc['icd_code'].unique()
metaDic['Proc']=self.proc_per_adm
if(self.feat_lab):
with open("./data/dict/labsVocab", 'wb') as fp:
pickle.dump(list(labs['itemid'].unique()), fp)
self.lab_vocab = labs['itemid'].unique()
metaDic['Lab']=self.labs_per_adm
with open("./data/dict/metaDic", 'wb') as fp:
pickle.dump(metaDic, fp)