Future Label Engineering (2h, 4h, 6, 8h) + window splits

Hyperparameter tuning for XGboost model

Data Partitioning at Patient Level (70-15-15)

Initial cleaning and imputation for static features

To sort and imputation of Vital Signs in 'Per-patient loop' structure

Imputation of Lab variables in 'Per-patient loop' structure

ADDED XXX_MISSING FOR 26 LAB VARIABLES EACH

Feature Engineering

Created 3 hours lag (History) for all vital and lab markers

Lag1 = t - 1

Lag2 = t - 2

Lag3 = t - 3

Created 3 delta (Sudden changes) for all vital and lab markers

Delta1 = t - (t-1)

Delta2 = t - (t-2)

Delta3 = t - (t-3)

Created Rolling averages (Trends) for all vital and lab markers with mean values and resolve missing values for incomplete windows.

Window type: backward

Window length: 3

Time Encoding Features

Sanity Check Temporal Features (Lag and Delta correctness for one patient)

Groupby patient with ICULOS max.

Prefer a patient with ICULOS >50 to check temporal features

Lag1 = t - 1

Lag2 = t - 2

Lag3 = t - 3

Delta1 = t - (t-1)

Delta2 = t - (t-2)

Delta3 = t - (t-3)

Mean/MA = [ t + (t-1), + (t-2) / 3 ]

Since Physionet 2019 dataset, Sepsislabel as 1 indicated that sepsis will occur within the next 6 hours.

Created 2h, 4h, 6h, and 8h windows to challenge if the model is able to predict sepsis earlier than 6 hrs before clinical recognition.

SepsisLabel_2h, Sepsis will occur within the next 8 hours

SepsisLabel_4h, Sepsis will occur within the next 10 hours

SepsisLabel_6h ,Sepsis will occur within the next 12 hours

SepsisLabel_8h, Sepsis will occur within the next 14 hours

Remove patient data that has less than 9 rows of data.

Reason: The node crashes if the table has fewer rows than the window size. Since sepsislabel_8h is window length = 9

It need at least 9 rows of per patient data for the node to function without errors.

328 patients has less than 9 rows of data will be excluded.

total patient strength will be 40008 instead of 40336

Sanity Check Temporal Features (Future label)

Training set (70%)

Validation set (15%)

Test set (15%)

Sepsislabel window

Sepsislabel_2h window

Sepsislabel_8h window

Sepsislabel_6h window

Sepsislabel_4h window

Class imbalance of Sepsislabel

Class imbalance of Sepsislabel_2h

Class imbalance of Sepsislabel_4h

Class imbalance of Sepsislabel_6h

Class imbalance of Sepsislabel_8h

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Nested Hyperparameter tuning for Logistic Regression model

Evaluation for Logistic Regression Model (Test Set)

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for Logistic Regression Model (Test Set)

Hyperparameter tuning for Logistic Regression model

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for Logistic Regression Model (Test Set)

Hyperparameter tuning for Logistic Regression model

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for Logistic Regression Model (Test Set)

Hyperparameter tuning for Logistic Regression model

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for Logistic Regression Model (Test Set)

Hyperparameter tuning for Logistic Regression model

Evaluation for XGBoost Model (Test Set)

Evaluation for XGBoost Model (Test Set)

MICU

SICU

Sepsislabel_8h window

Class imbalance of Sepsislabel_8h

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Data Partitioning at Patient Level (70-15-15)

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Class imbalance of Sepsislabel_6h

Sepsislabel_4h window

Class imbalance of Sepsislabel_4h

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Sepsislabel_2h window

Class imbalance of Sepsislabel_2h

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Evaluation for XGBoost Model (Test Set)

Training set (70%)

Validation set (15%)

Test set (15%)

Class imbalance of Sepsislabel

Sepsislabel_8h window

Class imbalance of Sepsislabel_8h

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Data Partitioning at Patient Level (70-15-15)

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Hyperparameter tuning for XGboost model

Class imbalance of Sepsislabel_6h

Sepsislabel_4h window

Class imbalance of Sepsislabel_4h

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Sepsislabel_2h window

Class imbalance of Sepsislabel_2h

Hyperparameter tuning for XGboost model

Training set (70%)

Validation set (15%)

Test set (15%)

Remove patient_ID to prevent the models to memorize it during model training which cause data leakage

Evaluation for XGBoost Model (Test Set)

Evaluation for XGBoost Model (Test Set)

Training set (70%)

Validation set (15%)

Test set (15%)

Class imbalance of Sepsislabel