Apache spark streaming analytic engine for predicting patient deterioration using physiological data

Engine consists of two parts

• Prediction model generator and evaluator (module model)
• Spark engine itself (module engine)

At this early stage real deterioration data is not provided, but there are many physiological realtime datasets on the web, so for a proof-of-concept I selected The University of Queensland Vital Signs Dataset.

I chose 6 features (heart rate, blood ox. saturation, blood pressure etc.). The task is to predict end-tidal sevoflurane concentration that patient will have next second.

Dataset processing, model training and exporting is done in Jupyter Notebook (model/notebooks/predict-etSEV.ipynb). For this task I used linear regression with Tensorflow. Achieved accuracy is >99.99%, so linear methods are more than sufficient for this problem.

For sake of brevity, engine operates on data received through socket and prints predictions to stdout.

• launch a server

nc -l 9999

• launch the engine

./run_clinical_engine.sh

• feed some data (to netcat)

0.73469388 1.0 0.80851064 0.77011494 0.45238095 0.5952381 0.24137931 0.37037037

• add separate script for training and exporting model
• add real datasource and datasink (e.g. Kafka topic and RDBMS)
• serve exported model with Tensorflow Serving
• wrap engine and TF Serving in Docker images and add a deploy script
• fault tolerance
• tests