Hospitalization Risk in Chronic Patients Model
Michael Leshchinsky, ML engineer
Background
This project was created as as a part of Clalit's effort to provide to primary care physicians a set of preventive tools that can help them to find high risk patients and intervene at the right time for the right patient.
This project combines machine learning model with clinical knowledge to create a list of chronic patients with high risk of non-ambulatory hospitalization within one year.
Physician or nurse can see this list in the web application and contact a patient to decide what preventive actions can be done.
Prediction Model
We used a Gradient Boosting model (via LightGBM) with hand-picked by clinical experts features to predict the hospitalization risk among patients with chronic diseases.
For data preprocessing, model training, hyper-parameters optimization, evaluation and versioning we used PanPredictor (internal MLOPS tool).
Model performance on the test population:
Model Performance on Test Population
Application:
We create five groups by risk, and provide to the physicians explanations on patient-level with a guideline about urgency of the intervention based on the risk group.
Lists of patients with info about their risk group and explained causes can be seen within C-PI web application. List is being updated automatically on monthly basis, including new patients and receiving new predictions based on updated data.
