Machine Learning for Arrhythmia Detection

Dec. 20, 2017

Dr. Gari Clifford, DPhil has been studying artificial intelligence (AI) and its utility in healthcare for two decades. He holds several prestigious positions in academia and is an Associate Professor of Biomedical Informatics at Emory University and an Associate Professor of Biomedical Engineering at Georgia Institute of Technology. We met him at the San Francisco Data Institute Conference in October where he chaired sessions on Machine Learning and Health.

Gari recently held a competition challenging data scientists to develop predictive algorithms for the early detection of Atrial Fibrillation, using mobile ECG machines. He shares insight into the complexity of using AI to diagnose health conditions and offers a glimpse into the future of healthcare and medical information.

Here’s the outline of this interview with Gari Clifford:

[00:01:07] The road to machine learning and mobile health.

[00:01:27] Lionel Tarassenko: neural networks and artificial intelligence.

[00:03:36] San Francisco Data Institute Conference.

[00:03:54] Jeremy Howard at fast.ai.

[00:04:17] Director of Data Institute David Uminsky.

[00:05:05] Dr. Roger Mark, Computing in Cardiology PhysioNet Challenges.

[00:05:23] 2017 Challenge: Detecting atrial fibrillation in electrocardiograms.

[00:05:44] Atrial Fibrillation.

[00:06:08] KardiaMobile EKG monitor by AliveCor.

[00:06:33] Random forests, support vector machines, heuristics, deep learning.

[00:07:23] Experts don't always agree.

[00:08:33] Labeling ECGs: AF, normal sinus rhythm, another rhythm, or noisy.

[00:09:07] 20-30 experts are required to discern a stable diagnosis.

[00:09:40] Podcast: Arrhythmias in Endurance Athletes, with Peter Backx, PhD.

[00:11:17] Applying additional algorithm on top of all final algorithms: improved score from 83% to 87% accuracy.

[00:11:38] Kaggle for machine learning competitions.

[00:13:44] Overfitting an algorithm increases complexity, decreases utility.

[00:15:01] 10,000 ECGs are not enough.

[00:16:24] Podcast: How to Teach Machines That Can Learn with Dr. Pedro Domingos.

[00:16:50] XGBoost.

[00:19:18] Mechanical Turk.

[00:20:08] QRS onset and T-wave offset.

[00:21:31] Galaxy Zoo.

[00:24:00] Podcast: Jason Moore of Elite HRV.

[00:24:34] Andrew Ng. Paper: Rajpurkar, Pranav, et al. "Cardiologist-level arrhythmia detection with convolutional neural networks." arXiv preprint arXiv:1707.01836 (2017).

[00:28:44] Detecting arrhythmias using other biomarkers.

[00:30:41] Algorithms trained on specific patient populations not accurate for other populations.

[00:31:24] Propensity matching.

[00:31:55] Should we be sharing our medical data?

[00:32:15] Privacy concerns associated with sharing medical data.

[00:32:44] Mass scale research: possible with high-quality data across a large population.

[00:33:04] Selling social media data in exchange for useful or entertaining software.

[00:33:42] Who touched my medical data and why?

[00:36:31] Siloing data, perhaps to protect the current industries.

[00:37:03] Health Insurance Portability and Privacy Act (HIPPA).

[00:37:34] Fast Healthcare Interoperability Resources (FHIR) protocol.

[00:37:48] Microsoft HealthVault and Google Health.

[00:38:46] Blockchain and 3blue1brown.

[00:39:28] Where to go to learn more about Gari Clifford.

[00:39:53] Presentation: Machine learning for FDA-approved consumer level point of care diagnostics – the wisdom of algorithm crowds: (the PhysioNet Computing in Cardiology Challenge 2017).