Understanding Healthcare (Anthem)

Click here to download the Understanding Healthcare Data PDF

Click here to access the project team's poster session (June 12, 2020, 1pm-2pm PDT)

Abstract

Compared to other wealthy countries, the U.S spends a disproportionate amount on healthcare with the gap widening every year.

We utilized machine learning tools to create models that predict a patient’s future health conditions given their previous medical history.

Anthem, one of the largest healthcare companies in the U.S, is exploring the use of these models in order to provide better care for the patient and at a reduced cost for the care provider.

Key Terms

Synthea: Creates statistically accurate synthetic medical  records curated by doctors for the purpose of research.

One-Hot Encoding (OHE): In a month, if a medical code was in a patient’s history, the entry is 1 or 0 otherwise.

Docker: a platform to package up applications and code for the purpose of reliably running on different machines

Python libraries

Numpy: Contains numerical tools for computations.

Pandas: Helps organize data and perform calculations.

Pytorch: Creates, trains, and tests machine learning models.

Overview

The goal of our project is to use machine learning to predict the factors that lead to the diagnosis of a patient with congestive heart failure. Anthem can use these predictions to preemptively form a treatment plan and inhibit the progression of a patient’s disease.

Machine Learning Technology

Recurrent Neural Network (RNN)

Uses predictions from previous inputs to influence the current input’s prediction.

Convolutional Neural Network (CNN)

An image classification technique used to find unique patterns in different images.

Word Embeddings (Word2Vec)

Vectors of numerical values assigned to words or individual components in a given text.

Approach

Because access to actual patient data is highly restricted, we used Synthea to generate artificial medical records. These records are used to create sparse One-Hot Encoded matrices (OHE) where each row represents one month of a patient's medical history and each column represents a unique medical code. 

Using OHE matrices we trained a CNN and an RNN. We also created word embeddings for each of the medical codes present in a patient's history to train a CNN.

After developing successful models, we packaged our models in a docker container and sent it to Anthem to train on real anonymized patient data.

Results and Conclusion

For the OHE matrices built from the synthetic data, we were able to create a CNN that predicted outcomes at an accuracy of 96.36%, and an RNN that predicted at an accuracy of 82.61%.

In addition to OHE matrices, we also used the word embedding approach, which resulted in an accuracy of 79.34% for the CNN and 71.53% for the RNN.

Anthem then trained our models on the word embeddings processed from real data, which achieved an accuracy of 86.33% on the CNN model and 83.72% on RNN model.