ICU SURVIVAL PREDICTION

https://icusurvivalapp.herokuapp.com/

Table of Contents

• Business problem
• Quick glance at the results
• Limitation and what can be improved
• Run Locally
• Repository structure

1. Business problem

The challenge is to create a model that uses data from the early hours of intensive care unit admission to predict patient survival. In clinical practice, estimates of mortality risk can be useful in triage and resource allocation, to determine appropriate levels of care, to prepare discussions with patients and their families around expected outcomes and help policymakers identify and make better health policies.

1.1 Objective

Objective of this project is to create a model predict patient survival with better prediction probability than the current apache, model that uses minimize apache features, transparent i.e easy to explain, and generalizability with Less complexity. As current systems often lack generalizability beyond the patients on whom the models were developed, and The models are often proprietary, costly to use (APACHE scoring system...), and suffer from opaque algorithms.

1.2 Data Source

MIT's GOSSIS community initiative, with privacy certification from the Harvard Privacy Lab, has provided a dataset of more than 90000 hospital Intensive Care Unit (ICU) visits from patients, spanning a one-year time frame. This data is part of a growing global effort and consortium spanning Argentina, Australia, New Zealand, Sri Lanka, Brazil, and more than 200 hospitals in the United States.

2. Quick glance at the results

Exploratory analysis

1. Percentage Representation of hospital_death

• The target feature (hospital_death) is highly imbalance with with a 91% survival rate and 9% non survival

2. Body mass index category

• From the dataset most of the patient are overweight and above

• 27% of the the patient in the observation has a normal BMI while 4% are underweight

3. Distribution of Height, weight, BMI and Age

• From the above box plot Age seems to a little left skewed with few outliers.
• BMI (body max index) and weight are right skewed with alot of outliers
• Height has normally distribution.
• The mean age, bmi, height and weight are 62.48, 29.15, 169.66cm and 83.98kg respectively. 4. Age group

• 63% patient from the observation are aged above 60 years old adults

Model performance

1. ROC_AUC Metrics

2. Best 3 models The Top 3 with the best performance are

Model	ROC_AUC score	Recall score
Logistic Regression	0.706	65%
Gradient boosting	0.778	78%
Light Gradient boosting	0.778	78%

3. Confusion matrix and ROC_AUC of LightGB classifier

• Faster traning time.
• Lower memory usage.
• Support parrallisation on distributed systems.
• Better prediction probability.

2.1 Metrics Used: Recall and ROC_AUC

Why choose Recall and ROC_AUC as metrics:

• Since the objective of this problem is to predict patient survival in the early hour of icu admision, Recall will be a better metric as inaccurate prediction of low survival (false negative) will have dire consequences.
• ROC_AUC so we can be able to set the best threashold to capture low survival(True positives).
• Imbalance between dataset.

2.2 Limitations And What Can Be Improved

• Minimum domain knowledge.
• Highly imbalanced dataset.
• Plenty of missing variables.
• Hyperparameter tuning: I used RandomSearchCV to save time but could be improved by couple of % with GridSearchCV.
• More data: Alot of factor contrribute to a patient survival in the ICU. Factor such, Qaulity of hospital, experience of doctor etc are not captured in the dataset.

3. Run Locally

# clone the project.
git clone https://github.com/Sachimugu/ICU-survival.git

# Create a conda virtual environment called icuprediction and install all the packages.
conda create --name icuprediction pandas sklearn Django lightgbm

# Activate the conda environment.
conda activate icuprediction

# enter the Script  directory.
cd ./site/django-web-app

# run server.
python manage.py runserver

Go to http://127.0.0.1:8000/ on web-brower

Repository structure

├── assets
│   ├── age.png
│   ├── bmi.jpg
│   ├── Bmi.png
│   ├── cm.png
│   ├── data.png
│   ├── gcs.jpg
│   ├── gcs.png
│   ├── GCS_Subscales.png
│   ├── roc.png
│   └── sc.png
├── Datasets
│   ├── clean_dataset
│   ├── clean_dataset.csv
│   ├── Data Dictionary.csv
│   ├── dataset.csv
│   └── model_metrics.csv
├── Model
│   └── pickel_lgb_model.pkl
├── Notebook
│   ├── 1_Data_Wranglin_and_Analysis.ipynb
│   ├── 2_Preprocessing_and_ML_Model.ipynb
│   └── US_States_ID.json
├── Presentation
│   ├── report.pdf
│   └── reportp.pptx
└── README.md

Contribution

Pull requests are welcome! For major changes.

Contact