Skip to content

Latest commit

 

History

History
90 lines (49 loc) · 4.38 KB

README.md

File metadata and controls

90 lines (49 loc) · 4.38 KB

Senti-Mentor

Helping hospices improve patient and family satisfaction

Insight Data Science Project by Alexander Fiksdal

This project predicts family recommendation rates and estimated mean ratings at the hospice facility level and provides insights into improving those satisfaction measures.

Project Overview

Hospice care eases end of life for patients and families while avoiding unnecessary hospital admissions and saving money. Hospice providers are required to submit a variety of data to the Centers for Medicare & Medicaid Services (CMS), which are then made available to the public. Moreover, these data are used in Medicare's Hospice Compare tool, which allows patients and families to compare providers across a variety of performance and service dimensions. However, provider-submitted data is decentralized and no tools exist to help providers analyze what informs measures of patient/family satisfaction. Senti-Mentor addresses these needs by combining multiple sources of hospice provider data and analyzing them using regression and gradient-boosted regression tree models. Specifically, it helps hospice providers address the following questions:

  • Are our patients and families satisfied?
    • Visualizations of observed satisfaction measures in the context of state and national distributions
    • For facilities that lack satisfaction data: XGBoost model predictions based on available features
  • What factors predict satisfaction?
    • Summaries of regression models predicting family recommendation rates and average ratings from actionable measures of service quality, services delivered, and facility characteristics.
  • What factors should take priority in future interventions?
    • Targeted recommendations constructed by taking into account both the magnitude of regression coefficients and specific provider feature values.

Step 1: Merge CMS Data Sources

Data Sources

Public Use File

Consumer Assessment of Healthcare Providers & Systems (CAHPS)

General Hospice Provider Information

Hospice Quality of Patient Care Data

Raw data files were saved in a folder hidden from github with the following path: data/raw

Python scripts used to clean/combine data are located in the following folder: hospice_project/data/

Step 2: Feature Engineering

The following features were created using the hospice_project/features/build_features.py script:

  • Estimated Ratings
  • Service Quality (Usually + Always %)
  • Probably + Definitely Would Recommend %
  • Facilities within 30/60/90 miles
  • Distance to nearest facility
  • Per-Beneficiary Measures of Services Delivered and Charges

Step 3: Train Models

Two models were trained: Insight Model (for interpretation and recommendations) and Prediction Model (for facilities with missing performance measures). The scripts used for comparing and training final models is located in hospice_project/models/

Scaling and imputation of missing values used transformer classes located in hospice_project/data/transformers.py

Other fixed values (variable lists, custom dictionaries, etc.) defined in definitions.py file.

Step 4: Implementing Everything

Executing the following scripts in this order will produce the files required for the streamlit app to function:

hospice_project/data/clean_data.py

hospice_project/features/build_features.py

hospice_project/models/insight_model_compare.py

hospice_project/models/prediction_model_compare.py

hospice_project/models/train_insight.py

hospice_project/models/train_prediction.py

Step 5: streamlit app

Final tool built as streamlit application. To run locally, type this in the terminal:

streamlit run sentimentor.py

Directory Structure

Directory structure initialized using cookiecutter:

https://github.com/MisterVladimir/cookiecutter-data-science/tree/v0.1.0

Presentation

Google slides summarizing the project can be found here.