Convert Ordered ICD DataFrame to Indicator

convert_ordered_to_indicator.py

import numpy as np
import pandas as pd
import argparse

def convert_ordered_dx_df_to_indicator_df(
        ordered_dx_df,
        list_of_possible_icd_code_values=None,
        list_of_possible_icd_code_names=None):
    ''' Convert rank-ordered diagnosis codes into indicator format

    Args
    ----
    ordered_dx_df : Pandas Dataframe
        One column for each possible rank 
        E.g. 1st column is most important dx, 2nd col is next important, ...

    Returns
    -------
    indicator_df : Pandas Dataframe
        One column for each possible unique diagnosis code
    '''
    assert ordered_dx_df.values.dtype == object

    if list_of_possible_icd_code_values is None:
        list_of_possible_icd_code_values = np.unique(ordered_dx_df.values)

    list_of_possible_icd_code_values = [
        x for x in list_of_possible_icd_code_values if len(x) > 0]
    list_of_possible_icd_code_values = np.asarray(list_of_possible_icd_code_values, dtype=unicode)


    if list_of_possible_icd_code_names is None:
        list_of_possible_icd_code_names = [
            'icd_%s' % (code) for code in list_of_possible_icd_code_values]
    list_of_possible_icd_code_names = np.asarray(list_of_possible_icd_code_names, dtype=unicode)

    sort_ids = np.argsort(list_of_possible_icd_code_values)
    list_of_possible_icd_code_values = list_of_possible_icd_code_values[sort_ids]
    list_of_possible_icd_code_names = list_of_possible_icd_code_names[sort_ids]

    vec_list = list()
    for row in range(ordered_dx_df.shape[0]):
        cur_vec = np.zeros(len(list_of_possible_icd_code_values), dtype=np.int32)
        for col in range(ordered_dx_df.shape[1]):
            cur_code = ordered_dx_df.iloc[row, col]
            if cur_code == '':
                continue
            loc = np.searchsorted(list_of_possible_icd_code_values, cur_code)
            cur_vec[loc] = 1
        vec_list.append(cur_vec)

    indicator_df = pd.DataFrame(
        vec_list,
        columns=list_of_possible_icd_code_names)
    return indicator_df


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--ordered_dx_csv_file', type=str,
        default='ordered_dx_simple.csv')
    parser.add_argument('--indicator_dx_csv_file', type=str,
        default='indicator_dx_simple.csv')
    parser.add_argument('--n_icd_codes', type=int, default=10)
    args = parser.parse_args()

    ordered_dx_df = pd.read_csv(args.ordered_dx_csv_file, dtype=object)
    ordered_dx_df.fillna('', inplace=True)

    # Create estimate of indicators
    # Using known list of possible codes
    est_ind_df = convert_ordered_dx_df_to_indicator_df(
        ordered_dx_df, map(str, range(args.n_icd_codes)))

    # Create estimate of indicators version 2
    # Detect possible codes by what's in the file
    est2_ind_df = convert_ordered_dx_df_to_indicator_df(
        ordered_dx_df)

    true_ind_df = pd.read_csv(args.indicator_dx_csv_file, dtype=np.int32)

    print("True indicator_df has %d columns" % true_ind_df.shape[1])
    print("True indicator_df: first 5 rows")
    print(true_ind_df.head())

    print("Estimated indicator_df has %d columns" % est_ind_df.shape[1])
    print("Estimated indicator_df: first 5 rows")
    print(est_ind_df.head())

    print("Estimated version 2:")
    print(est2_ind_df.head())

create_random_icd_dataset.py

import numpy as np
import pandas as pd
import argparse

from collections import OrderedDict

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--n_subjects', type=int, default=5)
    parser.add_argument('--n_icd_codes', type=int, default=10)
    parser.add_argument('--min_codes_per_subject', type=int, default=1)
    parser.add_argument('--max_codes_per_subject', type=int, default=4)
    parser.add_argument('--seed', type=int, default=54321)
    parser.add_argument('--nickname', type=str, default='simple')
    args = parser.parse_args()

    n_subjects = args.n_subjects
    n_icd_codes = args.n_icd_codes
    min_codes_per_subject = args.min_codes_per_subject
    max_codes_per_subject = args.max_codes_per_subject

    prng = np.random.RandomState(args.seed)

    codes_per_subject = dict()

    ordered_dx_dicts = list()
    indicator_vectors = list()

    # Loop over each subject, generate a random set of icd codes
    for subj_id in range(n_subjects):
        # How many distinct codes to draw for this subj
        n_codes = prng.randint(low=min_codes_per_subject, high=max_codes_per_subject)

        # Select a list of codes uniformly from all available codes without replacement
        code_list = prng.choice(n_icd_codes, n_codes, replace=False)

        codes_per_subject[subj_id] = code_list.tolist()

        # Store representation with ordered priority of dx
        dx_dict = OrderedDict()
        for ii, code in enumerate(code_list):
            dx_dict['dx_%02d' % ii] = code
        ordered_dx_dicts.append(dx_dict)

        # Store representation as an indicator vector
        vec = np.zeros(n_icd_codes, dtype=np.int32)
        vec[code_list] = 1
        indicator_vectors.append(vec)

    ordered_dx_df = pd.DataFrame(ordered_dx_dicts)
    ordered_dx_df.to_csv(
        'ordered_dx_%s.csv' % args.nickname,
        float_format='%.0f',
        index=False)

    indicator_dx_df = pd.DataFrame(
        np.vstack(indicator_vectors),
        columns=['icd_%s' % i for i in range(n_icd_codes)])
    indicator_dx_df.to_csv(
        'indicator_dx_%s.csv' % args.nickname,
        float_format='%.0f',
        index=False)

Expected output:

True indicator_df has 10 columns
True indicator_df: first 5 rows
   icd_0  icd_1  icd_2  icd_3  icd_4  icd_5  icd_6  icd_7  icd_8  icd_9
0      0      1      0      0      0      0      0      0      1      0
1      0      0      0      0      0      0      0      0      1      0
2      0      1      0      0      0      1      0      0      0      1
3      0      0      1      0      0      0      0      1      1      0
4      0      0      0      0      0      1      0      0      0      0
Estimated indicator_df has 10 columns
Estimated indicator_df: first 5 rows
   icd_0  icd_1  icd_2  icd_3  icd_4  icd_5  icd_6  icd_7  icd_8  icd_9
0      0      1      0      0      0      0      0      0      1      0
1      0      0      0      0      0      0      0      0      1      0
2      0      1      0      0      0      1      0      0      0      1
3      0      0      1      0      0      0      0      1      1      0
4      0      0      0      0      0      1      0      0      0      0
Estimated version 2:
   icd_1  icd_2  icd_5  icd_7  icd_8  icd_9
0      1      0      0      0      1      0
1      0      0      0      0      1      0
2      1      0      1      0      0      1
3      0      1      0      1      1      0
4      0      0      1      0      0      0