ML Algorithm to Detect Rare Clinical Events

Talk held at PyCon JP 2019. Authors: Haque Ishfaq; Samara Sharmeen; Hassan Sami Adnan; Hassan Saad Ifti; Atia Amin

Transcript

1. Haque Ishfaq PhD Researcher Montreal Institute for Learning Algorithms (MILA); School of Computer Science McGill University, Canada @haqueishfaq Samara Sharmeen Resident Surgeon VIVANTES Netzwerk für Gesundheit GmbH RWTH Aachen University, Germany @ssharmeen Hassan Sami Adnan Medical Researcher Faculty of Health, Medicine and Life Sciences, Maastricht University World Health Organization (WHO) @hsami Hassan Saad Ifti DPhil Researcher Department of Engineering Science University of Oxford, United Kingdom @saadifti Atia Amin Researcher Department of Biology University of South Dakota, United States @atiaamin1 ML Algorithm to Detect Rare Clinical Events
2. Success of Deep Learning Source: Stanford CS 231N
3. Step 1: Create a massive labelled dataset
4. Step 2: Train a large deep neural network Source: Stanford CS 231N
5. Step 3: Evaluate Performance
6. But is that all? What if our training set training data
7. And then during test time ... test data
8. Not just a toy example! Perils of biased
9. Not just a toy example!
10. TVAE: Triplet Based Variational - Triplet Loss and Metric Learning - Generative Model - Variational Autoencoder
11. What is Metric Learning? training data
12. But is that all? What if our training set training data Representation
13. What is Metric Learning? Positive Anchor Negative
14. What is Metric Learning? Positive Anchor Negative Learning
15. What is Metric Learning? Positive Anchor Negative Positive Anchor Negative Learning
16. Variational Autoenconder? Source: http://kvfrans.com/variational-autoencoders-explained/
17. Triplet Based Variational Autoencoder
18. MNIST Dataset
19. Triplet Based Variational Autoencoder
20. 0 2000 4000 6000 8000 1999 2001 2003 2005 2007 2009 2011 2013 2015 2017 2019 Digital Health Deep Learning Artificial Intelligence Number of publications on PubMed, on three different topics Emerging trend in medicine
21. Problems of classification - Difficulty in defining clinical classifications - Risk of misdiagnosis - Delay in screening or treatment - Missed opportunity for detecting rare events
22. Benefits of representation - More relaxed paradigm - Physicians in the loop to deal with red flags - Early detection of rare events (e.g. microcalcification in breast cancer)
23. tomorrow’s Talk… Simulation Model for 3D-printed Drug Development
24. PySTEMPlus @PySTEM Haque Ishfaq PhD Researcher Montreal Institute for Learning Algorithms (MILA); School of Computer Science, McGill University, Canada @haqueishfaq Samara Sharmeen Resident Surgeon VIVANTES Netzwerk für Gesundheit GmbH; RWTH Aachen University, Germany @ssharmeen Hassan Sami Adnan Medical Researcher Faculty of Health, Medicine and Life Sciences, Maastricht University; World Health Organization (WHO) @hsami Hassan Saad Ifti DPhil Researcher Department of Engineering Science; University of Oxford, United Kingdom @saadifti Atia Amin Researcher Department of Biology University of South Dakota, United States @atiaamin1 Outreach platform for underprivileged children PyHEALTH Bridging IT and Healthcare Follow us on Twitter for latest updates on our research. Our next event Hosting the first ever Python coding camp for children in Mym, Bangladesh!