The dataset comes from MIMIC project (https://mimic.physionet.org/). MIMIC-III (Medical Information Mart for Intensive Care III) is a large, freely-available database comprising deidentified health-related data associated with over forty thousand patients who stayed in critical care units of the Beth Israel Deaconess Medical Center between 2001 and 2012.
Each row of mimic_train_X.csv correponds to one ICU stay (hadm_id+icustay_id) of one patient (subject_id). In mimic_train_y.csv HOSPITAL_EXPIRE_FLAG is the indicator of death (=1) as a result of the current hospital stay; this is the outcome to predict in our modelling exercise. The columns of mimic_train_X.csv correspond to vitals of each patient (when entering the ICU), plus some general characteristics (age, gender, etc.), and their explanation can be found at mimic_patient_metadata.csv.
Note that the main cause/disease of patient condition is embedded as a code at ICD9_diagnosis column. The meaning of this code can be found at MIMIC_metadata_diagnose.csv. But this is only the main one; a patient can have co-occurrent diseases (comorbidities). These secondary codes can be found at extra_data/MIMIC_diagnoses.csv.
There is an extra test dataset, mimic_test_X.csv. Apply your final model to this extra dataset and produce a prediction .csv file in same format as mimic_kaggle_death_sample_submission.csv.
In this project, you have to predict the probability of death of a patient that is entering an ICU (Intensive Care Unit), using the machine learning models we have covered in class.
The dataset comes from MIMIC project (https://mimic.physionet.org/). MIMIC-III (Medical Information Mart for Intensive Care III) is a large, freely-available database comprising deidentified health-related data associated with over forty thousand patients who stayed in critical care units of the Beth Israel Deaconess Medical Center between 2001 and 2012.
Each row of mimic_train_X.csv correponds to one ICU stay (hadm_id+icustay_id) of one patient (subject_id). In mimic_train_y.csv HOSPITAL_EXPIRE_FLAG is the indicator of death (=1) as a result of the current hospital stay; this is the outcome to predict in our modelling exercise. The columns of mimic_train_X.csv correspond to vitals of each patient (when entering the ICU), plus some general characteristics (age, gender, etc.), and their explanation can be found at mimic_patient_metadata.csv.
Note that the main cause/disease of patient condition is embedded as a code at ICD9_diagnosis column. The meaning of this code can be found at MIMIC_metadata_diagnose.csv. But this is only the main one; a patient can have co-occurrent diseases (comorbidities). These secondary codes can be found at extra__data/MIMIC_diagnoses.csv.
As performance metric, you can use AUC for the binary classification case, but feel free to report as well any other metric if you can justify that is particularly suitable for this case.
Main tasks are:
-
Using mimic_train.csv file build a predictive model for HOSPITAL_EXPIRE_FLAG*.
-
For this analysis there is an extra test dataset, mimic_test_death.csv. Apply your final model to this extra dataset and produce a prediction csv file in same format as mimic_kaggle_death_sample_submission.csv.
The evaluation metric for this competition is ROC AUC (Area Under the Curve). The AUC, commonly used in binary classification models, measures the area under a curve that is obtained by varying the threshold for binary classification (0.5 by default) and computing True Positive Rates and False Positive Rates.
For every patient in the dataset, submission files should contain two columns: icustayid (this defines the individual prediction, and it’s extracted from the test dataset) and HOSPITAL_EXPIRE_FLAG (float number between 0 and 1, the probability of death).
The file should contain a header and have the following format:
icustay_id,HOSPITAL_EXPIRE_FLAG
1346,0.67
You can check the sample submission file mimic_kaggle_death_sample_submission.csv.