import os import pandas as pd import datasets import sys import pickle import subprocess import shutil from urllib.request import urlretrieve from sklearn.model_selection import train_test_split from sklearn.preprocessing import LabelEncoder import yaml import numpy as np from .dataset_utils import vocab, concat_data, generate_deep, generate_ml, generate_text from .task_cohort import create_cohort ################################################################################ ################################################################################ ## ## ## MIMIC IV DATASET GENERATION SCRIPT ## ## ## ################################################################################ ################################################################################ _DESCRIPTION = """\ Dataset for mimic4 data, by default for the Mortality task. Available tasks are: Mortality, Length of Stay, Readmission, Phenotype. The data is extracted from the mimic4 database using this pipeline: 'https://github.com/healthylaife/MIMIC-IV-Data-Pipeline/tree/main' mimic path should have this form : "path/to/mimic4data/from/username/mimiciv/2.2" If you choose a Custom task provide a configuration file for the Time series. Currently working with Mimic-IV ICU Data. """ _BASE_URL = "https://huggingface.co/datasets/thbndi/Mimic4Dataset/resolve/main" _HOMEPAGE = "https://huggingface.co/datasets/thbndi/Mimic4Dataset" _CITATION = "https://proceedings.mlr.press/v193/gupta22a.html" _GIT_URL = "https://github.com/healthylaife/MIMIC-IV-Data-Pipeline" _DATA_GEN = f"{_BASE_URL}/data_generation_icu_modify.py" _DATA_GEN_HOSP= f"{_BASE_URL}/data_generation_modify.py" _DAY_INT= f"{_BASE_URL}/day_intervals_cohort_v22.py" _CONFIG_URLS = {'los' : f"{_BASE_URL}/config/los.config", 'mortality' : f"{_BASE_URL}/config/mortality.config", 'phenotype' : f"{_BASE_URL}/config/phenotype.config", 'readmission' : f"{_BASE_URL}/config/readmission.config" } class Mimic4DatasetConfig(datasets.BuilderConfig): """BuilderConfig for Mimic4Dataset.""" def __init__( self, **kwargs, ): super().__init__(**kwargs) class Mimic4Dataset(datasets.GeneratorBasedBuilder): """Create Mimic4Dataset dataset from Mimic-IV data stored in user machine.""" VERSION = datasets.Version("1.0.0") def __init__(self, **kwargs): self.mimic_path = kwargs.pop("mimic_path", None) self.encoding = kwargs.pop("encoding",'concat') self.config_path = kwargs.pop("config_path",None) self.test_size = kwargs.pop("test_size",0.2) self.val_size = kwargs.pop("val_size",0.1) self.generate_cohort = kwargs.pop("generate_cohort",True) if self.encoding == 'concat': self.concat = True else: self.concat = False super().__init__(**kwargs) BUILDER_CONFIGS = [ Mimic4DatasetConfig( name="Phenotype", version=VERSION, description="Dataset for mimic4 Phenotype task" ), Mimic4DatasetConfig( name="Readmission", version=VERSION, description="Dataset for mimic4 Readmission task" ), Mimic4DatasetConfig( name="Length of Stay", version=VERSION, description="Dataset for mimic4 Length of Stay task" ), Mimic4DatasetConfig( name="Mortality", version=VERSION, description="Dataset for mimic4 Mortality task" ), ] DEFAULT_CONFIG_NAME = "Mortality" def init_cohort(self): if self.config_path==None: if self.config.name == 'Phenotype' : self.config_path = _CONFIG_URLS['phenotype'] if self.config.name == 'Readmission' : self.config_path = _CONFIG_URLS['readmission'] if self.config.name == 'Length of Stay' : self.config_path = _CONFIG_URLS['los'] if self.config.name == 'Mortality' : self.config_path = _CONFIG_URLS['mortality'] version = self.mimic_path.split('/')[-1] mimic_folder= self.mimic_path.split('/')[-2] mimic_complete_path='/'+mimic_folder+'/'+version current_directory = os.getcwd() if os.path.exists(os.path.dirname(current_directory)+'/MIMIC-IV-Data-Pipeline-main'): dir =os.path.dirname(current_directory) os.chdir(dir) else: #move to parent directory of mimic data dir = self.mimic_path.replace(mimic_complete_path,'') print('dir : ',dir) if dir[-1]!='/': dir=dir+'/' elif dir=='': dir="./" parent_dir = os.path.dirname(self.mimic_path) os.chdir(parent_dir) #####################clone git repo if doesnt exists repo_url='https://github.com/healthylaife/MIMIC-IV-Data-Pipeline' if os.path.exists('MIMIC-IV-Data-Pipeline-main'): path_bench = './MIMIC-IV-Data-Pipeline-main' else: path_bench ='./MIMIC-IV-Data-Pipeline-main' subprocess.run(["git", "clone", repo_url, path_bench]) os.makedirs(path_bench+'/'+'mimic-iv') shutil.move(version,path_bench+'/'+'mimic-iv') os.chdir(path_bench) self.mimic_path = './'+'mimic-iv'+'/'+version ####################Get configurations param #download config file if not custom if self.config_path[0:4] == 'http': c = self.config_path.split('/')[-1] file_path, head = urlretrieve(self.config_path,c) else : file_path = self.config_path if not os.path.exists('./config'): os.makedirs('config') #save config file in config folder self.conf='./config/'+file_path.split('/')[-1] if not os.path.exists(self.conf): shutil.move(file_path,'./config') with open(self.conf) as f: config = yaml.safe_load(f) #get config parameters for time series and features timeW = config['timeWindow'] self.timeW=int(timeW.split()[1]) self.bucket = config['timebucket'] self.predW = config['predW'] self.data_icu = config['icu_no_icu']=='ICU' if self.data_icu: self.feat_cond, self.feat_chart, self.feat_proc, self.feat_meds, self.feat_out, self.feat_lab = config['diagnosis'], config['chart'], config['proc'], config['meds'], config['output'], False else: self.feat_cond, self.feat_lab, self.feat_proc, self.feat_meds, self.feat_chart, self.feat_out = config['diagnosis'], config['lab'], config['proc'], config['meds'], False, False #####################downloads modules from hub if not os.path.exists('./model/data_generation_icu_modify.py'): file_path, head = urlretrieve(_DATA_GEN, "data_generation_icu_modify.py") shutil.move(file_path, './model') if not os.path.exists('./model/data_generation_modify.py'): file_path, head = urlretrieve(_DATA_GEN_HOSP, "data_generation_modify.py") shutil.move(file_path, './model') if not os.path.exists('./preprocessing/day_intervals_preproc/day_intervals_cohort_v22.py'): file_path, head = urlretrieve(_DAY_INT, "day_intervals_cohort_v22.py") shutil.move(file_path, './preprocessing/day_intervals_preproc') data_dir = "./data/dict/"+self.config.name.replace(" ","_")+"/dataDic" sys.path.append(path_bench) config = self.config_path.split('/')[-1] #####################create task cohort if self.generate_cohort: create_cohort(self.config.name.replace(" ","_"),self.mimic_path,config) #####################Split data into train, test and val with open(data_dir, 'rb') as fp: dataDic = pickle.load(fp) data = pd.DataFrame.from_dict(dataDic) dict_dir = "./data/dict/"+self.config.name.replace(" ","_") data=data.T train_data, test_data = train_test_split(data, test_size=self.test_size, random_state=42) if self.val_size > 0 : train_data, val_data = train_test_split(train_data, test_size=self.val_size, random_state=42) val_dic = val_data.to_dict('index') val_path = dict_dir+'/val_data.pkl' with open(val_path, 'wb') as f: pickle.dump(val_dic, f) train_dic = train_data.to_dict('index') test_dic = test_data.to_dict('index') train_path = dict_dir+'/train_data.pkl' test_path = dict_dir+'/test_data.pkl' with open(train_path, 'wb') as f: pickle.dump(train_dic, f) with open(test_path, 'wb') as f: pickle.dump(test_dic, f) return dict_dir #verify if the dimension of the tensors corresponds to the time window def verif_dim_tensor(self, proc, out, chart, meds, lab,interv): verif=True if self.feat_proc: if (len(proc)!= interv): verif=False if self.feat_out: if (len(out)!=interv): verif=False if self.feat_chart: if (len(chart)!=interv): verif=False if self.feat_meds: if (len(meds)!=interv): verif=False if self.feat_lab: if (len(lab)!=interv): verif=False return verif def save_features(self,concat_cols,dyn_df,cond_df,demo): #create csv with the description of each feature for analysis purpose df_feats = pd.DataFrame(columns=['feature','description']) icd = pd.read_csv(self.mimic_path+'/hosp/d_icd_diagnoses.csv.gz',compression='gzip', header=0) items= pd.read_csv(self.mimic_path+'/icu/d_items.csv.gz',compression='gzip', header=0) if self.encoding == 'concat': feats = concat_cols.copy() df_feats['feature'] = feats for _, data in df_feats.iterrows(): txt=(items[items['itemid'] == int(data['feature'].split('_')[0])]['label']).to_string(index=False) data['description']=txt+' at interval '+data['feature'].split('_')[1] else: feats = list(dyn_df.columns.droplevel(0)) for _, data in df_feats.iterrows(): data['description']=(items[items['itemid'] == int(data['feature'])]['label']).to_string(index=False) for diag in list(cond_df.columns): df_feats.loc[len(df_feats)] = [diag,icd[icd['icd_code'] == diag]['long_title'].to_string(index=False)] df_feats.loc[len(df_feats)]='Age' df_feats.loc[len(df_feats)]='gender' df_feats.loc[len(df_feats)]='ethnicity' df_feats.loc[len(df_feats)]='insurance' feats.extend(list(cond_df.columns)) feats.extend(list(demo.columns)) path = './data/dict/'+self.config.name.replace(" ","_")+'/features_description_'+self.encoding+'.csv' df_feats.to_csv(path,index=False) feat_tocsv=False return feat_tocsv, feats ###########################################################RAW################################################################## def _info_raw(self): features = datasets.Features( { "label": datasets.ClassLabel(num_classes=2,names=["0", "1"]), "gender": datasets.Value("string"), "ethnicity": datasets.Value("string"), "insurance": datasets.Value("string"), "age": datasets.Value("int32"), "COND": datasets.Sequence(datasets.Value("string")), "MEDS": { "signal": { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) } , "rate": { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) } , "amount": { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) } }, "PROC": { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) }, "CHART/LAB": { "signal" : { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) }, "val" : { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) }, }, "OUT": { "id": datasets.Sequence(datasets.Value("int32")), "value": datasets.Sequence(datasets.Sequence(datasets.Value("float32"))) }, } ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, citation=_CITATION, ) def _generate_examples_raw(self, filepath): with open(filepath, 'rb') as fp: dataDic = pickle.load(fp) for hid, data in dataDic.items(): proc_features = data['Proc'] meds_features = data['Med'] out_features = data['Out'] cond_features = data['Cond']['fids'] eth= data['ethnicity'] age = data['age'] gender = data['gender'] label = data['label'] insurance=data['insurance'] items = list(proc_features.keys()) values =[proc_features[i] for i in items ] procs = {"id" : items, "value": values} items_outs = list(out_features.keys()) values_outs =[out_features[i] for i in items_outs ] outs = {"id" : items_outs, "value": values_outs} if self.data_icu: chart_features = data['Chart'] else: chart_features = data['Lab'] #chart signal if ('signal' in chart_features): items_chart_sig = list(chart_features['signal'].keys()) values_chart_sig =[chart_features['signal'][i] for i in items_chart_sig ] chart_sig = {"id" : items_chart_sig, "value": values_chart_sig} else: chart_sig = {"id" : [], "value": []} #chart val if ('val' in chart_features): items_chart_val = list(chart_features['val'].keys()) values_chart_val =[chart_features['val'][i] for i in items_chart_val ] chart_val = {"id" : items_chart_val, "value": values_chart_val} else: chart_val = {"id" : [], "value": []} charts = {"signal" : chart_sig, "val" : chart_val} #meds signal if ('signal' in meds_features): items_meds_sig = list(meds_features['signal'].keys()) values_meds_sig =[meds_features['signal'][i] for i in items_meds_sig ] meds_sig = {"id" : items_meds_sig, "value": values_meds_sig} else: meds_sig = {"id" : [], "value": []} #meds rate if ('rate' in meds_features): items_meds_rate = list(meds_features['rate'].keys()) values_meds_rate =[meds_features['rate'][i] for i in items_meds_rate ] meds_rate = {"id" : items_meds_rate, "value": values_meds_rate} else: meds_rate = {"id" : [], "value": []} #meds amount if ('amount' in meds_features): items_meds_amount = list(meds_features['amount'].keys()) values_meds_amount =[meds_features['amount'][i] for i in items_meds_amount ] meds_amount = {"id" : items_meds_amount, "value": values_meds_amount} else: meds_amount = {"id" : [], "value": []} meds = {"signal" : meds_sig, "rate" : meds_rate, "amount" : meds_amount} yield int(hid), { "label" : label, "gender" : gender, "ethnicity" : eth, "insurance" : insurance, "age" : age, "COND" : cond_features, "PROC" : procs, "CHART/LAB" : charts, "OUT" : outs, "MEDS" : meds } ###########################################################ENCODED################################################################## def _info_encoded(self): features = datasets.Features( { "label": datasets.ClassLabel(num_classes=2,names=["0", "1"]), "features" : datasets.Sequence(datasets.Value("float32")), "features_names" : datasets.Sequence(datasets.Value("string")), } ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, citation=_CITATION, ) def _generate_examples_encoded(self, filepath): path= './data/dict/'+self.config.name.replace(" ","_")+'/ethVocab' with open(path, 'rb') as fp: ethVocab = pickle.load(fp) path= './data/dict/'+self.config.name.replace(" ","_")+'/insVocab' with open(path, 'rb') as fp: insVocab = pickle.load(fp) genVocab = ['', 'M', 'F'] gen_encoder = LabelEncoder() eth_encoder = LabelEncoder() ins_encoder = LabelEncoder() gen_encoder.fit(genVocab) eth_encoder.fit(ethVocab) ins_encoder.fit(insVocab) with open(filepath, 'rb') as fp: dico = pickle.load(fp) df = pd.DataFrame.from_dict(dico, orient='index') feat_tocsv=True for i, data in df.iterrows(): dyn_df,cond_df,demo=concat_data(data,self.interval,self.feat_cond,self.feat_proc,self.feat_out, self.feat_chart, self.feat_meds,self.feat_lab,self.condDict, self.procDict, self.outDict, self.chartDict, self.medDict) dyn=dyn_df.copy() dyn.columns=dyn.columns.droplevel(0) concat_cols = [f"{col}_{t}" for t in range(dyn.shape[0]) for col in dyn.columns] demo['gender']=gen_encoder.transform(demo['gender']) demo['ethnicity']=eth_encoder.transform(demo['ethnicity']) demo['insurance']=ins_encoder.transform(demo['insurance']) label = data['label'] demo=demo.drop(['label'],axis=1) if feat_tocsv: feat_tocsv, feats = self.save_features(concat_cols,dyn_df,cond_df,demo) X= generate_ml(dyn_df,cond_df,demo,concat_cols,self.concat) X=X.values[0] size_concat = self.size_cond+ self.size_proc * self.interval + self.size_meds * self.interval+ self.size_out * self.interval+ self.size_chart *self.interval+ self.size_lab * self.interval + 4 size_aggreg = self.size_cond+ self.size_proc + self.size_meds+ self.size_out+ self.size_chart+ self.size_lab + 4 if ((self.concat and len(X)==size_concat) or ((not self.concat) and len(X)==size_aggreg)): yield int(i), { "label": label, "features": X, "features_names" : feats } ######################################################DEEP############################################################### def _info_deep(self): features = datasets.Features( { "label": datasets.ClassLabel(num_classes=2,names=["0", "1"]), "DEMO": datasets.Sequence(datasets.Value("int64")), "COND" : datasets.Sequence(datasets.Value("int64")), "MEDS" : datasets.Array2D(shape=(None, self.size_meds), dtype='int64') , "PROC" : datasets.Array2D(shape=(None, self.size_proc), dtype='int64') , "CHART/LAB" : datasets.Array2D(shape=(None, self.size_chart), dtype='int64') , "OUT" : datasets.Array2D(shape=(None, self.size_out), dtype='int64') , } ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, citation=_CITATION, ) def _generate_examples_deep(self, filepath): with open(filepath, 'rb') as fp: dico = pickle.load(fp) for key, data in dico.items(): stat, demo, meds, chart, out, proc, lab, y = generate_deep(data,self.interval, self.config.name.replace(" ","_"), self.feat_cond, self.feat_proc, self.feat_out, self.feat_chart, self.feat_meds,self.feat_lab,self.condDict, self.procDict, self.outDict, self.chartDict, self.medDict, self.eth_vocab,self.gender_vocab,self.age_vocab,self.ins_vocab) if self.verif_dim_tensor(proc, out, chart, meds, lab, self.interval): if self.data_icu: yield int(key), { 'label': y, 'DEMO': demo, 'COND': stat, 'MEDS': meds, 'PROC': proc, 'CHART/LAB': chart, 'OUT': out, } else: yield int(key), { 'label': y, 'DEMO': demo, 'COND': stat, 'MEDS': meds, 'PROC': proc, 'CHART/LAB': lab, 'OUT': out, } ######################################################text############################################################## def _info_text(self): features = datasets.Features( { "label": datasets.ClassLabel(num_classes=2,names=["0", "1"]), "text" : datasets.Value(dtype='string', id=None), } ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, homepage=_HOMEPAGE, citation=_CITATION, ) def _generate_examples_text(self, filepath): icd = pd.read_csv(self.mimic_path+'/hosp/d_icd_diagnoses.csv.gz',compression='gzip', header=0) items= pd.read_csv(self.mimic_path+'/icu/d_items.csv.gz',compression='gzip', header=0) with open(filepath, 'rb') as fp: dico = pickle.load(fp) for key, data in dico.items(): cond_text,chart_text,meds_text,proc_text,out_text = generate_text(data,icd,items, self.feat_cond, self.feat_chart, self.feat_meds, self.feat_proc, self.feat_out) text= cond_text+chart_text+meds_text+proc_text+out_text yield int(key),{ 'label' : data['label'], 'text': text } ############################################################################################################################# def _info(self): self.path = self.init_cohort() self.interval = (self.timeW//self.bucket) self.size_cond, self.size_proc, self.size_meds, self.size_out, self.size_chart, self.size_lab, self.eth_vocab,self.gender_vocab,self.age_vocab,self.ins_vocab,self.condDict,self.procDict,self.medDict,self.outDict,self.chartDict,self.labDict=vocab(self.config.name.replace(" ","_"),self.feat_cond,self.feat_proc,self.feat_out,self.feat_chart,self.feat_meds,self.feat_lab) if (self.encoding == 'concat' or self.encoding =='aggreg'): return self._info_encoded() elif self.encoding == 'tensor' : return self._info_deep() elif self.encoding == 'text' : return self._info_text() else: return self._info_raw() def _split_generators(self, dl_manager): data_dir = "./data/dict/"+self.config.name.replace(" ","_") if self.val_size > 0 : return [ datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": data_dir+'/train_data.pkl'}), datasets.SplitGenerator(name=datasets.Split.VALIDATION, gen_kwargs={"filepath": data_dir+'/val_data.pkl'}), datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"filepath": data_dir+'/test_data.pkl'}), ] else : return [ datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": data_dir+'/train_data.pkl'}), datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"filepath": data_dir+'/test_data.pkl'}), ] def _generate_examples(self, filepath): if (self.encoding == 'concat' or self.encoding == 'aggreg'): yield from self._generate_examples_encoded(filepath) elif self.encoding == 'tensor' : yield from self._generate_examples_deep(filepath) elif self.encoding == 'text' : yield from self._generate_examples_text(filepath) else : yield from self._generate_examples_raw(filepath)