cestwc/heritage-health-prize-release-3

Dataset Card for Heritage Health Prize

It is often believed that this piece of data can be found at here and here, although we have not yet figured out what this piece of data is really used for.

To save time, we directly follow the preprocessing script here. More specifically, we used the following script to produce this Hugging Face dataset.

"""
Preprocessing based on: https://github.com/truongkhanhduy95/Heritage-Health-Prize
"""
import zipfile
from os import path
from urllib import request

import numpy as np
import pandas as pd


column_names = ['MemberID', 'ProviderID', 'Sex', 'AgeAtFirstClaim']
claims_cat_names = ['PrimaryConditionGroup', 'Specialty', 'ProcedureGroup', 'PlaceSvc']


def preprocess_claims(df_claims):
    df_claims.loc[df_claims['PayDelay'] == '162+', 'PayDelay'] = 162
    df_claims['PayDelay'] = df_claims['PayDelay'].astype(int)

    df_claims.loc[df_claims['DSFS'] == '0- 1 month', 'DSFS'] = 1
    df_claims.loc[df_claims['DSFS'] == '1- 2 months', 'DSFS'] = 2
    df_claims.loc[df_claims['DSFS'] == '2- 3 months', 'DSFS'] = 3
    df_claims.loc[df_claims['DSFS'] == '3- 4 months', 'DSFS'] = 4
    df_claims.loc[df_claims['DSFS'] == '4- 5 months', 'DSFS'] = 5
    df_claims.loc[df_claims['DSFS'] == '5- 6 months', 'DSFS'] = 6
    df_claims.loc[df_claims['DSFS'] == '6- 7 months', 'DSFS'] = 7
    df_claims.loc[df_claims['DSFS'] == '7- 8 months', 'DSFS'] = 8
    df_claims.loc[df_claims['DSFS'] == '8- 9 months', 'DSFS'] = 9
    df_claims.loc[df_claims['DSFS'] == '9-10 months', 'DSFS'] = 10
    df_claims.loc[df_claims['DSFS'] == '10-11 months', 'DSFS'] = 11
    df_claims.loc[df_claims['DSFS'] == '11-12 months', 'DSFS'] = 12

    df_claims.loc[df_claims['CharlsonIndex'] == '0', 'CharlsonIndex'] = 0
    df_claims.loc[df_claims['CharlsonIndex'] == '1-2', 'CharlsonIndex'] = 1
    df_claims.loc[df_claims['CharlsonIndex'] == '3-4', 'CharlsonIndex'] = 2
    df_claims.loc[df_claims['CharlsonIndex'] == '5+', 'CharlsonIndex'] = 3

    df_claims.loc[df_claims['LengthOfStay'] == '1 day', 'LengthOfStay'] = 1
    df_claims.loc[df_claims['LengthOfStay'] == '2 days', 'LengthOfStay'] = 2
    df_claims.loc[df_claims['LengthOfStay'] == '3 days', 'LengthOfStay'] = 3
    df_claims.loc[df_claims['LengthOfStay'] == '4 days', 'LengthOfStay'] = 4
    df_claims.loc[df_claims['LengthOfStay'] == '5 days', 'LengthOfStay'] = 5
    df_claims.loc[df_claims['LengthOfStay'] == '6 days', 'LengthOfStay'] = 6
    df_claims.loc[df_claims['LengthOfStay'] == '1- 2 weeks', 'LengthOfStay'] = 11
    df_claims.loc[df_claims['LengthOfStay'] == '2- 4 weeks', 'LengthOfStay'] = 21
    df_claims.loc[df_claims['LengthOfStay'] == '4- 8 weeks', 'LengthOfStay'] = 42
    df_claims.loc[df_claims['LengthOfStay'] == '26+ weeks', 'LengthOfStay'] = 180
    df_claims['LengthOfStay'].fillna('?', inplace=True)
    # df_claims['LengthOfStay'] = df_claims['LengthOfStay'].astype(int)

    for cat_name in claims_cat_names:
        df_claims[cat_name].fillna('?', inplace=True)
    # df_claims = pd.get_dummies(df_claims, columns=claims_cat_names, prefix_sep='=')

    oh = [col for col in df_claims if '=' in col]

    agg = {
        'ProviderID': ['count', 'nunique'],
        'Vendor': 'nunique',
        'PCP': 'nunique',
        'CharlsonIndex': 'max',
        # 'PlaceSvc': 'nunique',
        # 'Specialty': 'nunique',
        # 'PrimaryConditionGroup': 'nunique',
        # 'ProcedureGroup': 'nunique',
        'PayDelay': ['sum', 'max', 'min']
    }
    for col in oh:
        agg[col] = 'sum'

    df_group = df_claims.groupby(['Year', 'MemberID'])
    df_claims = df_group.agg(agg).reset_index()
    df_claims.columns = [
                            'Year', 'MemberID', 'no_Claims', 'no_Providers', 'no_Vendors', 'no_PCPs',
                            'max_CharlsonIndex', 'PayDelay_total', 'PayDelay_max', 'PayDelay_min'
                        ] + oh

    return df_claims

def preprocess_drugs(df_drugs):
    df_drugs.drop(columns=['DSFS'], inplace=True)
    # df_drugs['DSFS'] = df_drugs['DSFS'].apply(lambda x: int(x.split('-')[0])+1)
    df_drugs['DrugCount'] = df_drugs['DrugCount'].apply(lambda x: int(x.replace('+', '')))
    df_drugs = df_drugs.groupby(['Year', 'MemberID']).agg({'DrugCount': ['sum', 'count']}).reset_index()
    df_drugs.columns = ['Year', 'MemberID', 'DrugCount_total', 'DrugCount_months']
    print('df_drugs.shape = ', df_drugs.shape)
    return df_drugs

def preprocess_labs(df_labs):
    df_labs.drop(columns=['DSFS'], inplace=True)
    # df_labs['DSFS'] = df_labs['DSFS'].apply(lambda x: int(x.split('-')[0])+1)
    df_labs['LabCount'] = df_labs['LabCount'].apply(lambda x: int(x.replace('+', '')))
    df_labs = df_labs.groupby(['Year', 'MemberID']).agg({'LabCount': ['sum', 'count']}).reset_index()
    df_labs.columns = ['Year', 'MemberID', 'LabCount_total', 'LabCount_months']
    print('df_labs.shape = ', df_labs.shape)
    return df_labs

def preprocess_members(df_members):
    df_members['AgeAtFirstClaim'].fillna('?', inplace=True)
    df_members['Sex'].fillna('?', inplace=True)
    # df_members = pd.get_dummies(
    #     df_members, columns=['AgeAtFirstClaim', 'Sex'], prefix_sep='='
    # )
    print('df_members.shape = ', df_members.shape)
    return df_members

# request.urlretrieve('https://foreverdata.org/1015/content/HHP_release3.zip', 'HHP_release3.zip')

zf = zipfile.ZipFile('HHP_release3.zip')

df_claims = preprocess_claims(pd.read_csv(zf.open('Claims.csv'), sep=','))
df_drugs = preprocess_drugs(pd.read_csv(zf.open('DrugCount.csv'), sep=','))
df_labs = preprocess_labs(pd.read_csv(zf.open('LabCount.csv'), sep=','))
df_members = preprocess_members(pd.read_csv(zf.open('Members.csv'), sep=','))

df_labs_drugs = pd.merge(df_labs, df_drugs, on=['MemberID', 'Year'], how='outer')
df_labs_drugs_claims = pd.merge(df_labs_drugs, df_claims, on=['MemberID', 'Year'], how='outer')
df_health = pd.merge(df_labs_drugs_claims, df_members, on=['MemberID'], how='outer')

# df_health.drop(['Year', 'MemberID'], axis=1, inplace=True)
# df_health.fillna(0, inplace=True)


# Convert continuous columns to float
continuous_columns = ["MemberID", "LabCount_total", 'LabCount_months',
 'DrugCount_total',
 'DrugCount_months',
 'no_Claims',
 'no_Providers',
 'no_Vendors',
 'no_PCPs',
 'max_CharlsonIndex',
 'PayDelay_total',
 'PayDelay_max',
 'PayDelay_min']

for col in continuous_columns:
    df_health[col] = pd.to_numeric(df_health[col], errors='coerce')


# Define categorical columns
categorical_columns = [
    'Year', 'AgeAtFirstClaim',
 'Sex'
]

# Dictionary to store category mappings
category_mappings = {}

for col in categorical_columns:
    # Convert train column to category and extract categories
    df_health[col] = df_health[col].astype("category")
    category_mappings[col] = df_health[col].cat.categories.to_list()  # Store category order

    # Convert to integer codes
    df_health[col] = df_health[col].cat.codes
    

from datasets import Dataset, DatasetDict, Features, Value, ClassLabel

# Define Hugging Face dataset schema
hf_features = Features({
    "Year": ClassLabel(names=category_mappings["Year"]),
    "AgeAtFirstClaim": ClassLabel(names=category_mappings["AgeAtFirstClaim"]),
    "Sex": ClassLabel(names=category_mappings["Sex"]),    
    **{col: Value("int64") for col in continuous_columns}
})

hf_features = Features({
    col: Value("int64") if col in continuous_columns else
         ClassLabel(names=category_mappings[col])
    for col in df_health.columns
})

from datasets import DatasetDict

# Store in a dataset dictionary
hf_dataset = DatasetDict({"train": Dataset.from_pandas(df_health, features=hf_features)})

# Print dataset structure
print(hf_dataset)

The printed output could look like

DatasetDict({
    train: Dataset({
        features: ['Year', 'MemberID', 'LabCount_total', 'LabCount_months', 'DrugCount_total', 'DrugCount_months', 'no_Claims', 'no_Providers', 'no_Vendors', 'no_PCPs', 'max_CharlsonIndex', 'PayDelay_total', 'PayDelay_max', 'PayDelay_min', 'AgeAtFirstClaim', 'Sex'],
        num_rows: 218415
    })
})

We may still update this dataset, once we figure out more.