Deep Learning Steals The Show At Singularity’s Exponential Medicine Conference

image

Last week, Singularity University held its annual xMed conference, an invite-only, four-day event held in San Diego that draws a who’s who of entrepreneurs and researchers working at the intersection of computer science and medicine. The conference offers industry leaders an opportunity to update the audience on advances in medicine made across the fields of artificial intelligence, nanotechnology, 3D printing, smartphone-based clinical trial management, genome sequencing, virtual reality, and the ongoing effort to incorporate patient-generated health data into care delivery in a meaningful way.

Among the presenters that shared the stage during the kickoff ceremony was Jeremy Howard, a machine learning expert brought on stage to discuss the migration from basic machine learning algorithms to deep learning and artificial neural networks. His accomplishments in the field are extensive, and include positions as the president and chief scientist of renowned data science community Kaggle, as well as holding data science teaching positions at the University of San Francisco and Singularity University, where he sits as the youngest faculty on staff. He co-developed the Perl programming language, and successfully launched and sold two big data-based startups. Now, he is the CEO of Enlitic, a startup using deep learning algorithms to analyze medical images in search of small, easily overlooked tumors. Notably, however, Howard is completely new to medicine.

During his session, Howard discussed the early works of researchers focused on implementing machine learning in radiology and pathology. He recounts the findings from a Stanford University team that manually compiled a list of 7,000 unique features that a tumor could have, and then wrote a machine-learning algorithm to pour through patient radiology reports to calculate how frequently each feature turns up, and determine its relationship with overall five-year mortality rates. The algorithm then used this information to formulate prognosis predictions for cancer patients based on the unique features of their tumor. Those predictions turned out to be more accurate than clinicians were able to make.

Standing on the shoulders of these earlier successes, Howard explains how in 2012, a team of data scientists created a drug discovery algorithm during a two-week innovation competition that outperformed Merck’s own in-house algorithms. Those data scientists, who also had no background in medicine, used a new branch of machine learning called artificial neural networks, or deep learning, that speeds up development because it does not need any initial search criteria to run. This eliminates a massive amount of upfront effort, as in the case of the Stanford team that had to manually compile those 7,000 tumor features. With deep-learning algorithms, all that is needed is raw data and a stated objective.

Howard next explains that image analysis is the sub-sector of deep learning that is accelerating at a faster pace than any other, and that image analysis algorithms are now 10,000 times faster and 10-times more accurate than they were five years ago. He also notes that each year, accuracy approximately doubles. From here, Howard outlines the vision and technology behind his own recently launched startup, Enlitic.

Enlitic is building a user interface over an easily programmable deep-learning engine that will highlight key areas of a medical image, and then search a database of millions of patient records to find images with near identical properties, consolidating the diagnosis, treatment, and outcomes from those patients and presenting a contextualized summary of the image for the clinical team. They tested the algorithm by having it analyze a batch of lung CT scans, some of which included cancer tumors. The algorithm competed against four radiologists that manually reviewed each image. The results were impressive; the algorithm had a lower false negative and false positive rate than the radiologists.

As EHR data, patient-generated data, and even genetic data make their way into healthcare, Enlitic hopes to grow with the data, introducing the algorithms that will make sense of the data and present it to clinicians in a way that offers insight. It’s a lofty challenge, but one that is already needed, and despite his lack of medical experience, Howard’s extensive understanding of deep learning networks makes him an ideal candidate.


Enjoy HIStalk Connect? Sign up for update alerts, or follow us at @HIStalkConnect.

↑ Back to top

Founding Sponsors

Platinum Sponsors