Recommender System for a German Healthcare Provider

Project overview

At QUIBIQ GmbH, I worked in Research and Development on a recommender system for a German healthcare provider. The system recommends GOP service codes based on a patient’s ICD diagnosis codes.

Approach

1. Patient graph construction. Built a patient graph connecting patients with shared or similar diagnosis patterns. Patient representations included demographic and diagnosis information; one representation used an autoencoder to reduce the dimensionality of ICD diagnosis vectors.
2. Recommend procedures. Trained a graph neural network to predict GOP service codes from the patient graph. The paper evaluated the models on a synthetic patient database and reported an average 6.48% F1-score improvement over baseline models.
3. Analyze patient clusters. Used the graph models to identify groups of patients associated with distinct treatment patterns.
4. Interpret and present results. Applied saliency maps to examine model recommendations, and built a Dash dashboard to display the results.

For more details, see the paper.