shamir-lab / abxappropriatenessml

Prediction of ABX appropriateness in ICU inpatients

Predicting appropriateness of antibiotic treatment among ICU patients with hospital acquired infection

Public repository containing code for model predicting antibiotic treatment appropriateness, as described in the manuscript "Predicting appropriateness of antibiotic treatment among ICU patients with hospital acquired infection"

Authors:

Ella Goldschmidt*, Ella Rannon*, Daniel Bernstein, Asaf Wasserman, Dan Coster, Ron Shamir

Modules

The repository contains the following modules:

• data_processing Code for preprocessing the data, including time-series feature creation, imputation, outlier removal, etc.
• feature_engineering Code for the creation of the features used by the model.
• feature_selection Code for feature selection methods and filtration of correlated and redundant features.
• feature_stats Code for conducting statistical tests on the features.
• models Code for the BalancedDataEnsemble model and a generic class for ABX appropriateness prediction model.

Data:

The data used in our study is from MIMIC-III. This section describes the data format used for the code. The 'target' variable represents whether the antibiotic treatment administered to the patient was appropriate, given the antibiogram results of a specific culture sample. A value of 1 indicates inappropriate treatment, while a value of 0 signifies appropriate treatment. The determination of appropriateness is based on the sensitivity of the identified bacteria to the antibiotics administered to the patient.

Our data-specific parser generates 6 main pandas dataframes:

• Patient_df Contains demographics (static features).
• Drug_df Drugs entries.
• Culture_df Culture entries.
• Culture_antibiotic_df Cultures paired with antibiotic checked on it.
• Procedures_df Procedures entries.
• Labs_df Lab tests and vital signs (longitudinal features).

Patient dataframe format (Patient_df):

This dataset encompasses all demographic details of the patients, with each individual assigned two specific identifiers: 'subject_id', a unique identifier that remains constant throughout a patient's lifetime, and 'hadm_id', which is unique to each hospital stay of a patient. We introduced the 'identifier' column, which combines 'subject_id' and 'hadm_id' using a hyphen (i.e. 'subject_id-hadm_id'). Essential information about each patient, including age, weight, height, and ethnicity, is captured. The 'target' variable serves as the label indicating the appropriateness of the treatment, which is essential only during the model training phase. There is no requirement for this label when performing inference. Furthermore, we calculate the time elapsed from hospital admission to the target time (the time of the relevant culture sample), as well as the duration from ICU admission to the target time.

Columns	Data type
identifier	String
subject_id	int
hadm_id	int
target	int
gender	String
age	int
admission_weight	float64
admission_height	float64
hours_from_admittime_to_target_time	float64
hours_from_icutime_to_target_time	float64
ethnicity	String

Drug dataframe format (Drug_df):

All the drugs administered in the hospital were aggregated into 11 categories and for each category. Here 'itemid' is the id of a specific drug and we calculate the time from administration to target time. This table will be used in our model to calculate the total amount of drugs administered from each category and used as a feature for the model.

Columns	Data type
identifier	String
subject_id	int
hadm_id	int
itemid	int
hours_from_charttime_to_target_time	float64

Culture dataframe format (Culture_df):

The cultures table contains two key attributes that later will serve as features: the tissue in which the sample was taken from, labeled as 'spec_type_desc', and the name of the organism detected within, referred to as 'org_name'. Additionally, we compute the duration from the time the culture was taken to the target time.

Columns	Data type
identifier	String
subject_id	int
hadm_id	int
spec_type_desc	String
org_name	String
hours_from_charttime_to_target_time	float64

Culture Antibiotics dataframe format (Culture_antibiotic_df):

The table includes outcomes of cultures obtained from patients, specifically focusing on the antibiotics tested against each culture, denoted as 'ab_name', and whether the bacteria were sensitive (indicated by "S") or resistant ("R") to the antibiotic, as recorded under 'interpretation'. Additionally, we compute the duration from the time the culture was taken to the target time.

Columns	Data type
identifier	String
subject_id	int
hadm_id	int
org_name	String
ab_name	String
hours_from_charttime_to_target_time	float64
interpretation	String

Procedures dataframe format (Procedures_df):

For all the procedures conducted in the hospital, our code generates features for several invasive procedures, categorizing them into four types: Arterial Line, Catheter, Ventilation, and Tubes. Only procedures with labels falling under one of these four categories should be included in the table (see double_names_mimic.json). Additionally, we compute the duration from the time of the procedure to the target time.

Columns	Data type
identifier	String
subject_id	int
hadm_id	int
hours_from_charttime_to_target_time	float64
label	String

Lab and Vital signs dataframe format (Lab_df):

This table stores lab measurements and vital signs, where 'itemid' represents the hospital-assigned ID for a specific measurement. The 'label' refers to the name of the measurement conducted, and 'valuenum' is the recorded numerical value. Additionally, we compute the duration from the time of measurement to the target time.

Columns	Data type
identifier	String
subject_id	int
hadm_id	int
itemid	int
label	String
valuenum	float64
hours_from_charttime_to_target_time	float64