/breath-sound-detection

Mini 'hackathon' style project for UCL CDT in AI-enabled Healthcare end of year symposium 2021

Primary LanguageJupyter Notebook

breath-sound-detection

This was a short 'hackathon' style project for the UKRI UCL Centre for Doctoral Training in AI-enabled Healthcare Systems end of year symposium 2021.


👩‍💻 Team: Chris Tomlinson, Joe Farrington, Vicky Smith

🎯 Task: Free reign to envisage/prototype an AI-enabled solution to data captured from a digital stethoscope 🩺📈

🌆 Commercial partner: Start-up operating in 'stealth mode' 🕵️‍♂️

⏰ Time scale: ~2 hours before delivering a 10 minute presentation 👨‍🏫


💡 Our solution


We constructed a Random Forest Classifier able to predict periods of inspiration (breathing in) from a .wav file from a digital stethoscope.

  1. We used librosa to import the .wav file, produce a spectrogram and store this as a numpy array
  2. We applied a High Pass Filter at 50 Hz and a Low Pass Filter at 1.5 kHz by cropping the numpy array to focus on the frequencies of interest, producing a feature matrix X

ROC curve

  1. An outcome vector y was created by manually listening to the .wav file and noting timestamps at which inspiration began and ended. These were then converted to a vector, of length X, where y=1 during inspiration, else y=0
  2. We partitioned the .wav file into ~80/20% train/test split, manually adjusting the breakpoint to leave two complete breaths for test evaluation
  3. We trained a RandomForestClassifier() and XGBClassifier() using the sklearn package. We selected these algorithms for their well-documented high-performance
  4. Evaluation on the held-out test set gave an AUC = 0.84 for the Random Forest classifier, as seen on the ROC curve below

ROC curve

  • Despite achieving a remarkably good performance out-of-the-box it's important to note this model was trained and evaluated on the same .wav file (and patient), which does not a represent a real-world scenario
  • Further work could involve:
    • Hyperparameter optimisation
    • Utilising additional data, from multiple patients, for both testing and training
    • Use of a 'sliding-window' style approach and exploring deep learning solutions

Overall this was a really enjoyable morning. Having never taken part in a 'hackathon' or worked in the tech industry it was great a taste for the excitement and creativity that a sprint-style of working can bring, not to mention that after a year of COVID-19 enforced remote learning, collaborating with, and learning from, colleagues was a particular highlight!

Many thanks to the CDT team: Prof Paul Taylor, Dr Ken Li and Craig Smith, for their hard work in organising the event, as well as the commercial sponsors for their time, ideas and mentorship.

I look forward to next year!