Medical information extraction using Python

I am a nurse and I know python but I am not an expert, just used it to process DNA sequences

We got hospital records written in hum开发者_JS百科an languages and I am supposed to insert these data into a database or csv file but they are more than 5000 lines and this can be so hard. All the data are written in a consistent format let me show you an example

11/11/2010 - 09:00am : He got nausea, vomiting and died 4 hours later

I should get the following data

Sex: Male
Symptoms: Nausea
    Vomiting
Death: True
Death Time: 11/11/2010 - 01:00pm

Another example

11/11/2010 - 09:00am : She got heart burn, vomiting of blood and died 1 hours later in the operation room

And I get

Sex: Female
Symptoms: Heart burn
    Vomiting of blood
Death: True
Death Time: 11/11/2010 - 10:00am

the order is not consistent by when I say in ....... so in is a keyword and all the text after is a place until i find another keyword

At the beginnning He or She determine sex, got ........ whatever follows is a group of symptoms that i should split according to the separator which can be a comma, hypen or whatever but it's consistent for the same line

died ..... hours later also should get how many hours, sometimes the patient is stil alive and discharged ....etc

That's to say we have a lot of conventions and I think if i can tokenize the text with keywords and patterns i can get the job done. So please if you know a useful function/modules/tutorial/tool for doing that preferably in python (if not python so a gui tool would be nice)

Some few information:

there are a lot of rules to express various medical data but here are few examples
- Start with the same date/time format followed by a space followd by a colon followed by a space followed by He/She followed space followed by rules separated by and
- Rules:
    * got <symptoms>,<symptoms>,....
    * investigations were done <investigation>,<investigation>,<investigation>,......
    * received <drug or procedure>,<drug or procedure>,.....
    * discharged <digit> (hour|hours) later
    * kept under observation
    * died <digit> (hour|hours) later
    * died <digit> (hour|hours) later in <place>
other rules do exist but they follow the same idea

This uses dateutil to parse the date (e.g. '11/11/2010 - 09:00am'), and parsedatetime to parse the relative time (e.g. '4 hours later'):

import dateutil.parser as dparser
import parsedatetime.parsedatetime as pdt
import parsedatetime.parsedatetime_consts as pdc
import time
import datetime
import re
import pprint
pdt_parser = pdt.Calendar(pdc.Constants())   
record_time_pat=re.compile(r'^(.+)\s+:')
sex_pat=re.compile(r'\b(he|she)\b',re.IGNORECASE)
death_time_pat=re.compile(r'died\s+(.+hours later).*$',re.IGNORECASE)
symptom_pat=re.compile(r'[,-]')

def parse_record(astr):    
    match=record_time_pat.match(astr)
    if match:
        record_time=dparser.parse(match.group(1))
        astr,_=record_time_pat.subn('',astr,1)
    else: sys.exit('Can not find record time')
    match=sex_pat.search(astr)    
    if match:
        sex=match.group(1)
        sex='Female' if sex.lower().startswith('s') else 'Male'
        astr,_=sex_pat.subn('',astr,1)
    else: sys.exit('Can not find sex')
    match=death_time_pat.search(astr)
    if match:
        death_time,date_type=pdt_parser.parse(match.group(1),record_time)
        if date_type==2:
            death_time=datetime.datetime.fromtimestamp(
                time.mktime(death_time))
        astr,_=death_time_pat.subn('',astr,1)
        is_dead=True
    else:
        death_time=None
        is_dead=False
    astr=astr.replace('and','')    
    symptoms=[s.strip() for s in symptom_pat.split(astr)]
    return {'Record Time': record_time,
            'Sex': sex,
            'Death Time':death_time,
            'Symptoms': symptoms,
            'Death':is_dead}


if __name__=='__main__':
    tests=[('11/11/2010 - 09:00am : He got nausea, vomiting and died 4 hours later',
            {'Sex':'Male',
             'Symptoms':['got nausea', 'vomiting'],
             'Death':True,
             'Death Time':datetime.datetime(2010, 11, 11, 13, 0),
             'Record Time':datetime.datetime(2010, 11, 11, 9, 0)}),
           ('11/11/2010 - 09:00am : She got heart burn, vomiting of blood and died 1 hours later in the operation room',
           {'Sex':'Female',
             'Symptoms':['got heart burn', 'vomiting of blood'],
             'Death':True,
             'Death Time':datetime.datetime(2010, 11, 11, 10, 0),
             'Record Time':datetime.datetime(2010, 11, 11, 9, 0)})
           ]

    for record,answer in tests:
        result=parse_record(record)
        pprint.pprint(result)
        assert result==answer
        print

yields:

{'Death': True,
 'Death Time': datetime.datetime(2010, 11, 11, 13, 0),
 'Record Time': datetime.datetime(2010, 11, 11, 9, 0),
 'Sex': 'Male',
 'Symptoms': ['got nausea', 'vomiting']}

{'Death': True,
 'Death Time': datetime.datetime(2010, 11, 11, 10, 0),
 'Record Time': datetime.datetime(2010, 11, 11, 9, 0),
 'Sex': 'Female',
 'Symptoms': ['got heart burn', 'vomiting of blood']}

Note: Be careful parsing dates. Does '8/9/2010' mean August 9th, or September 8th? Do all the record keepers use the same convention? If you choose to use dateutil (and I really think that's the best option if the date string is not rigidly structured) be sure to read the section on "Format precedence" in the dateutil documentation so you can (hopefully) resolve '8/9/2010' properly. If you can't guarantee that all the record keepers use the same convention for specifying dates, then the results of this script would have be checked manually. That might be wise in any case.

Here are some possible way you can solve this -

1. Using Regular Expressions - Define them according to the patterns in your text. Match the expressions, extract pattern and you repeat for all records. This approach needs good understanding of the format in which the data is & of course regular expressions :)
2. String Manipulation - This approach is relatively simpler. Again one needs a good understanding of the format in which the data is. This is what I have done below.
3. Machine Learning - You could define all you rules & train a model on these rules. After this the model tries to extract data using the rules you provided. This is a lot more generic approach than the first two. Also the toughest to implement.

See if this work for you. Might need some adjustments.

new_file = open('parsed_file', 'w')
for rec in open("your_csv_file"):
    tmp = rec.split(' : ')
    date = tmp[0]
    reason = tmp[1]

    if reason[:2] == 'He':
        sex = 'Male'
        symptoms = reason.split(' and ')[0].split('He got ')[1]
    else:
        sex = 'Female'
        symptoms = reason.split(' and ')[0].split('She got ')[1]
    symptoms = [i.strip() for i in symptoms.split(',')]
    symptoms = '\n'.join(symptoms)
    if 'died' in rec:
        died = 'True'
    else:
        died = 'False'
    new_file.write("Sex: %s\nSymptoms: %s\nDeath: %s\nDeath Time: %s\n\n" % (sex, symptoms, died, date))

Ech record is newline separated \n & since you did not mention one patient record is 2 newlines separated \n\n from the other.

LATER: @Nurse what did you end up doing? Just curious.

Maybe this can help you too , it's not tested

import collections
import datetime
import re

retrieved_data = []

Data = collections.namedtuple('Patient', 'Sex, Symptoms, Death, Death_Time')
dict_data = {'Death':'',
             'Death_Time':'',
             'Sex' :'',
             'Symptoms':''}


with open('data.txt') as f:
     for line in iter(f.readline, ""):

         date, text = line.split(" : ")
         if 'died' in text:
             dict_data['Death'] = True
             dict_data['Death_Time'] = datetime.datetime.strptime(date, 
                                                                 '%d/%m/%Y - %I:%M%p')
             hours = re.findall('[\d]+', datetime.text)
             if hours:
                 dict_data['Death_Time'] += datetime.timedelta(hours=int(hours[0]))
         if 'she' in text:
            dict_data['Sex'] = 'Female'
         else:
            dict_data['Sex'] = 'Male'

         symptoms = text[text.index('got'):text.index('and')].split(',')

         dict_data['Symptoms'] = '\n'.join(symptoms) 

         retrieved_data.append(Data(**dict_data))

         # EDIT : Reset the data dictionary.
         dict_data = {'Death':'',
             'Death_Time':'',
             'Sex' :'',
             'Symptoms':''}

It would be relatively easy to do most of the processing with regards to sex, date/time, etc., as those before you have shown, since you can really just define a set of keywords that would indicate these things and use those keywords.

However, the matter of processing symptoms is a bit different, as a definitive list of keywords representing symptoms would be difficult and most likely impossible.

Here's the choice you have to make: does processing this data really represent enough work to spend days writing a program to do it for me? If that's the case, then you should look into natural language processing (or machine learning, as someone before me said). I've heard pretty good things about nltk, a natural language toolkit for Python. If the format is as consistent as you say it is, the natural language processing might not be too difficult.

But, if you're not willing to expend the time and effort to tackle a truly difficult CS problem (and believe me, natural language processing is), then you ought to do most of the processing in Python by parsing dates, gender-specific pronouns, etc. and enter in the tougher parts by hand (e.g. symptoms).

Again, it depends on whether or not you think the programmatic or the manual solution will take less time in the long run.