Urban Hacks
Hacking Health fosters collaborative innovation by engaging key groups of stakeholders, such as healthcare professionals, developers, designers, entrepreneurs, and patients, to create human-centric solutions to improve the health of our communities. Our weekend hackathons are fun, intense, hands-on events where small teams tackle tough problems in a supportive community of peers and mentors.
@HHHamOnt #HHH2019
November 8 - November 10 2019
Hacking Health Hamilton
This was my first Hacking Health hackathon, and much more open-ended than I had expected. It was also rather more business-focussed than my first hackathon. In retrospect, I'd approach the event with an idea in advance, rather than waiting to hear the pitches and join on with that team. Communication about expectations and theming (there was none, just health) was a bit lacking, and would have been good coming into the event.
By the end of the first evening, I joined on to a team that was interested in delivering AI-diagnosis for pneumonia from chest X-rays. After quite a bit of discussion, the team modified the plan somewhat to focus on making AI image analysis tools in medical imaging more transparent to the radiologists who increasingly want to employ them but aren't sure if they can trust these black-boxes.
The promise of AI was to reduce trivial work and allow doctors to focus on difficult critical problems. The $2 billion of current annual spending on healthcare AI, projected at 10-36 billion by 2025, certainly demonstrates our desire. But repeated studies show that, while computer analyses can be more accurate than humans, doctors are resistant to using them.
We believed the problem to be a lack of trust in the AI, because the doctor doesn’t see the AI’s decision-making process, the outcome of which will ultimately be their responsibility. We’ve seen the same hurdle in autonomous vehicles – we still need a human hand on (or at least near) the wheel while AI earns our trust.
Our AI-assisted solution was to combine existing technologies such as GAN machine learning and RIS/PACS workflows:
1. to keep the radiologist in the loop,
2. to reduce their cognitive load, and
3. to focus their attention on critical problem solving.
This is not a replacement for radiologists, it is an augmentation to radiologists.
The RadiAssist workflow presents the queue of cases, prioritised along two lines: human-triaged urgency and the AI’s certainty of disease presence. Lower certainties mean that the radiologist will need to spend more time and energy on this diagnosis, and it is placed at the top of their queue to be reviewed while they are still fresh. High certainty images are likely to be “Aunt Minnies” (my Aunt Minnie could tell you there’s a problem there), and can be safely reviewed when the radiologist is more fatigued.
Selecting a patient from the queue, the radiologist can see the AI’s diagnostic opinion is added to the patient record, and the patient imaging is in the radiologist’s preferred layout. RadiAssist also addresses the AI black-box issue. It provides the radiologist with not only a decision, but a likelihood of disease, and a clear visual rationale for that decision. The GAN is trained to produce convincing fake images of healthy imaging (chest X-rays for our purposes, that weekend), and indicates what changes (density increases and decreases) would be required in a patient image to make it a healthy image.
There are also benefits in the education of radiology residents - who can make their own diagnosis and then be directed by the AI to specific points of interest - and in communication with patients - who are reassured that a real doctor is addressing their personal illness and can see what the cause for concern is.
With AI matching or outperforming human diagnoses, how do we make the most of it without worrying and insulting radiologists and patients alike?
We build the means to let AI tools safely earn our trust
We build communication bridges – between AI and physician, between doctor and patient
RadiAssist, as a module for existing workstations, integrated into local workflows, can be a platform for the wider deployment of healthcare AI and its benefits.
And here are the other great team efforts from that weekend!
FINAL RESULTS
No prizes, but a better understanding of GAN machine learning!