The Fifth Elephant 2019

Abstract

The field of Deep Learning is making huge inroads in almost all spheres. What caught the world by a storm, surpassing human level performance with image classification, has today matured into a powerful tool to solve real-world problems. Today, Deep Learning is not just a research area limited to academics but a powerful tool utilized and improved by different companies/labs/institutions across the world.

Medical domain is no exception when it comes to utilizing Deep Learning and Machine Learning algorithms to solve issues. Some of the recent research work involves using such techniques to surpass doctors in identifying heart failures, identifying tumours, bone fractures, etc.

Medical domain has a ton of peripheral issues which are important and need to be addressed before tools like Deep Learning can be leveraged in the real world. There are many issues like privacy, efficacy, correctness, completeness/bias and so on. The issues are both technical as well as social in nature. Deep Learning models have largely been black-boxes which JUST work. It isn’t magic, but the theory behind their success has been fuzzy, until now. These models are so complex that it makes them difficult to understand. Thus, impacting their utilization in real-world medical scenarios.

In this talk, we would showcase how we utilize a deep learning model and overcome some of the limitations. We address the most important factor, the interpretability of our deep learning models. The research into interpretability in the recent year has made some real good progress. We particularly delve into the interpretability of attention based models without impacting the performance.

Key Takeaways from this talk

• Learn about the issues associated with Deep Learning Models in real-world
• Understand how attention based models work towards interpretability
• Understand how a real use case utilized this framework to predict future diagnosis

Target Audience

Data Scientists, Engineers, Managers, AI Enthusiasts

Outline

The focus of this session is to present an interpretable deep learning model for disease prediction. The session explains how this model leverages attention to provide insights into prediction of future diagnosis. We also discuss how the model provides capabilities into identification of contributing events and factors. We present how operational and other medical concerns were addressed and our future steps on the same.

Section 1 (Introduction): Deep Learning/Machine Learning in Medical Domain

• Present different use cases addressed in the medical domain using AI

Section 2 (Use Case): Interpretable Patient Diagnosis Framework

• Problem statement
• Brief introduction about research into attention based interpretable models
• Implementation of attention based deep learning model for disease prediction
• Inference Interpretation and its impact

Section 3 : Future Steps and Conclusion

• Improving model performance without impacting interpretability
• Impact of such models in practice/real-world

Requirements

Participants should have a fair understanding of Machine Learning and Deep Learning(especially). Basics of deep learning would be helpful in appreciating the advanced concepts of attention, etc.

Speaker bio

Raghav Bali

Raghav Bali is a Senior Data Scientist at one the world’s largest health care organizations. His work involves research & development of enterprise level solutions based on Machine Learning, Deep Learning and Natural Language Processing for Healthcare & Insurance related use cases. Raghav has also authored multiple books with leading publishers, the recent one on latest in advancements in Transfer Learning research.

Links

• NLP Workshop: https://www.analyticsvidhya.com/datahack-summit-2018/speakers/raghav-bali/

Slides

https://www.slideshare.net/BaliRaghav/deep-diagnosis-150157675