Deep Learning to Predict Obstructive Sleep Apnea from Overnight Recordings of Oxygen Saturation

This project aims to develop a low-cost solution, based on deep learning of oxygen saturation data, to predict obstructive sleep apnea.
Master
Sleep apnea

Sleep apnea causes breathing to repeatedly stop and start during sleep. There are several types of sleep apnea, but the most common is obstructive sleep apnea. This type of apnea occurs when your throat muscles intermittently relax and block your airway during sleep. Over 1 billion people in the world have undiagnosed obstructive sleep apnea according to a recent study [1].

During periods of apnea, people received less air, which results in decreased oxygen delivery to the body. The oxygen levels in the blood may fall repeatedly. This oxygen decrease is called an oxygen desaturation. It often drops by 3 or 4 percent (and sometimes much more) in association with sleep apnea events.

The objective of this project is to develop deep learning models to predict sleep apnea events in overnight recordings of oxygen saturation. The models will be based on data from 8000 overnight polysomnography recordings made available to us from the Sleep Heart Health Study (sleepdata.org/datasets/shhs). The student will preprocess the data to obtain training and test sets for deep learning. He/she will need to implement different deep learning models for time series data such as ResNet [2] and LSTMs [3] and compare their performance.

If time permits the student will implement a simple smartwatch app for a smartwatch with a pulse oximeter to test the models in a real-world setting with sleeping subjects.

Goal

1) Pre-process data and creating training and test sets to predict obstructive sleep apnea from oxygen saturation data.
2) Train, test, and compare deep learning models to predict obstructive sleep apnea from oxygen saturation data
3) Implement model with high accuracy as a smartwatch app
4) Write a masters thesis and publish a scientific article communicating the results

    Learning outcome

    1) The student will use machine learning for time series data to address an impactful societal challenge
    2) The student will learn to design a machine learning experiment and bring it all the way to product development
    3) The student will be exposed to inter-disciplinary research by interacting with sleep experts at LHL
    4) The student will also understand the business side of things in our effort to commercialise sleep apnea diagnosis

      Qualifications

        1) Strong mathematical foundations necessary to understand deep learning algorithms
        2) Good skills in written and spoken English 
        3) Intermediate to Advanced Python programming skills
        4) Knowledge of machine learning libraries such as PyTorch and TensorFlow is a plus
        5) The student needs to a Bachelor in any science discipline involving data analysis. We are flexible in the background as long as the student is self-taught in the above and has a great interest in the subject.

        Supervisors

        • Sagar Sen
        • Pierre Bernabet, PhD Student, Simula Research Laboratory (deep learning expert)
        • Nils Henrik Holmedahl, LHL Sykehuset, Gardemoen (medical advisor)
        • Håvard Bjor and Leopold Gasteen (Business development)

        Collaboration partners

        LHL-sykehuset, Gardermoen

        References

        References

        1. Benjafield, A., et al. "Global prevalence of obstructive sleep apnea in adults: estimation using currently available data." B67. RISK AND PREVALENCE OF SLEEP DISORDERED BREATHING. American Thoracic Society, 2018. A3962-A3962.
        2. Fawaz, Hassan Ismail, et al. "Deep learning for time series classification: a review." Data Mining and Knowledge Discovery 33.4 (2019): 917-963.
        3. Van Steenkiste, Tom, et al. "Automated Sleep Apnea Detection in Raw Respiratory Signals using Long Short-Term Memory Neural Networks." IEEE journal of biomedical and health informatics (2018).

        Contact person