文章大綱
翻譯： season
美國的一部分醫療數據是通過HIPPA 脫密後在 https://www.hcup-us.ahrq.gov/ 網站上對研究者開放進行探索的。但是由於她給出的數據格式爲asc 的不常見格式，我們需要轉化成csv 後才能正常使用spark 等大數據分析組件進行分析。
還好2015年，有人用python 寫了一個調用SAS 解析hcup 數據的開源庫，那麼今天我們就一起來探索一下，如何用python 對hcup 的asc 數據進行解析並使用。
環境搭建
python 及jupyter 環境
# 設置環境變量
export PATH="/root/anaconda2/bin/:$PATH"
source ~/.bashrc
jupyter notebook --no-browser --port 8888 --ip=0.0.0.0 --allow-root
jupyter notebook --generate-config
在~/home 或者c盤usrs administrators 下找到文件夾 .jupyter 修改jupyter_application_config.py 文件。
# c.NotebookApp.notebook_dir = '' 去掉註釋
conda 虛環境
conda create -n iz_pyhcup --copy -y -q python=2.7 ipykernel pandas numpy
source activate iz_pyhcup
echo "y" |pip install PyHCUP
echo "y" |pip install sqlalchemy
source deactivate
About
PyHCUP is a Python library for parsing and importing data obtained from the United States Healthcare Cost and Utilization Program (http://hcup-us.ahrq.gov).
Data from HCUP come as a text file, with each column a specific width. However, the widths of these columns, and their names, are elsewhere. HCUP provide this meta data as either SAS or SPSS data loading programs.
PyHCUP is built to extract meta data from the SAS loading programs, then use that meta data to parse the actual data in the fixed-width text files. You’ll still need to acquire the actual data through HCUP.
A more verbose set of instructions is available in a series of posts on the author’s blog at
http://bielism.blogspot.com/2013/12/hcup-and-python-pt-i-background.html.
Example Usage
Load a datafile/loadfile combination.
import pyhcup
# specify where your data and loadfiles live
datafile = 'D:\\Users\\hcup\\sid\\NY_SID_2009_CORE.asc'
loadfile = 'D:\\Users\\hcup\\sid\\sasload\\NY_SID_2009_CORE.sas'
# pull basic meta from SAS loadfile
meta_df = pyhcup.meta_from_sas(loadfile)
# use meta knowledge to parse datafile into a pandas DataFrame
df = pyhcup.read(datafile, meta_df)
# that's it. use df from here.
Deal with very large files that cannot be held in memory in two ways.
- To import a subset of rows, such as for preliminary work or troubleshooting, specify nrows to read and/or skiprows to skip using sas.df_from_sas().
# optionally specify nrows and/or skiprows to handle larger files
df = pyhcup.read(datafile, meta_df, nrows=500000, skiprows=1000000)
- To iterate through chunks of rows, such as for importing into a database, first use the metadata to build lists of column names and widths. Next, pass a chunksize to the read() function above to create a generator yielding manageable-sized chunks.
chunk_size = 500000
reader = pyhcup.read(datafile, meta_df, chunksize=chunk_size)
for df in reader:
# do your business
# such as replacing sentinel values (below)
# or inserting into a database with another Python library
Whether you are pulling in all records or just a chunk of records, you can also replace all those pesky missing/invalid data placeholders from HCUP (this is less useful for generically parsing missing values for non-HCUP files).
::
# fyi, this bulldozes through all values in all columns with no per-column control
replaced = pyhcup.replace_sentinels(df)
樣例程序
上文提供了兩種加載大數據文件的辦法(原始文件一般非常大，一次性加載到pandas 中肯定會報錯)，一種是迭代，一種是直接定位到某些行，進行子數據集的分析，下面給出一段樣例分析代碼，將hcup 數據集中的asc 文件轉化成標準csv
#### save NY_SASD_2016_CORE.asc
filename = "NY_SASD_2016_CORE.asc"
data_path = filename
load_path = 'NY_SASD_2016_CORE.sas'
#build a pandas DataFrame object from meta data
meta_df = pyhcup.sas.meta_from_sas(load_path)
chunk_size = 500000
reader = pyhcup.read(data_path, meta_df, chunksize=chunk_size)
index = 1
for df in reader:
if index==1:
#首先讀一次，去掉前兩行，生成文件
index = index +1
df[2:].to_csv('NY_SASD_2016_CORE.csv', index=None)
else:
#後面不帶header，追加文件
index = index +1
df.to_csv('NY_SASD_2016_CORE.csv', mode='a', header=False,index=None)
print(index)
寫了兩個封裝的函數，對應的status 類的asc 文件進行csv 文件的導出
##################### 批量寫入 ####################################
def write_hcupAsc_to_csv(file_name_for_status_And_Year):
filename = file_name_for_status_And_Year + ".asc"
load_path = file_name_for_status_And_Year + ".sas"
save_name = file_name_for_status_And_Year + ".csv"
meta_df = pyhcup.sas.meta_from_sas(load_path)
chunk_size = 500000
reader = pyhcup.read(filename, meta_df, chunksize=chunk_size)
index = 1
for df in reader:
if index==1:
#首先讀一次，去掉前兩行，生成文件
index = index +1
df[2:].to_csv(save_name, index=None)
print(type(df['KEY'].dtype))
else:
#後面不帶header，追加文件
index = index +1
df.to_csv(save_name, mode='a', header=False,index=None)
print(index)
########################### 測試寫入 從開頭第二行開始寫 nrows 行 ################################
def write_Test_hcupAsc_to_csv(file_name_for_status_And_Year,save_name,nrows):
filename = file_name_for_status_And_Year + ".asc"
load_path = file_name_for_status_And_Year + ".sas"
save_name = save_name + ".csv"
meta_df = pyhcup.sas.meta_from_sas(load_path)
df = pyhcup.read(filename, meta_df, nrows=nrows, skiprows=2)
df.to_csv(save_name,index=None)
還有一種讀取的方法，我們沒有用常用的chunksize，而是每次計算從特定位置開始讀取
#第二種方式，不用chunksize
filename = "NY_SID_2016_CORE.asc"
load_path = 'NY_SID_2016_CORE.sas'
save_name = 'NY_SID_2016_CORE.csv'
#build a pandas DataFrame object from meta data
meta_df = pyhcup.sas.meta_from_sas(load_path)
#獲取文件行數
length = len([ "" for line in open(filename,"r")])
print(length)
chunk_size = 500000
step = int(length /chunk_size)
df = pyhcup.read(filename, meta_df, nrows=nrows, skiprows=2)
df.to_csv(save_name,index=None)
for i in range(1,step):
reader = pyhcup.read(filename, meta_df, nrows=chunk_size, skiprows=2+i*chunk_size)
df.to_csv(save_name, mode='a', header=False,index=None)
Shortcut to loadfiles (meta data)
The SAS loading program files provided by HCUP for the State Inpatient Database (SID), State Ambulatory Surgery Database (SASD), and State Emergency Department Database (SEDD) are bundled in this package for easy access. You can retrieve the meta data for these directly, without having to specify a loadfile path as described above.
Acquire meta in this way using the get_meta() function. You must pass a state abbreviation as the first argument and a year as the second arugment, like so.
meta_df = pyhcup.get_meta('NY', 2009)
By default, get_meta() acquires SID CORE data. Other meta can be acquired with the optional keyword arguments datafile (‘SID’, ‘SEDD’, or ‘SASD’) and category (‘CORE’, ‘CHGS’, ‘SEVERITY’, ‘DX_PR_GRPS’, or ‘AHAL’).
# California emergency department charges meta for 2010
ca_2010_emergency_charges_meta = pyhcup.get_meta('CA', 2010, datafile='SEDD', category='CHGS')
# Arizona outpatient surgery DRG records meta for 2004
az_2004_surg_groups_meta = pyhcup.get_meta('AZ', 2004, datafile='SASD', category='DX_PR_GRPS'
# etc.
參考文獻
http://bielism.blogspot.com/2013/12/hcup-and-python-pt-5-nulls-and-pre.html