Medical applications of AI are replete with promise, but stymied by opacity: with lives on the line, concerns over AI models’ often-inscrutable reasoning – and as a result, possible biases embedded in those models – largely prevent scaled applications of AI for medical treatment, no matter how promising the underlying research. Recently, researchers from Mederrata Research (a nonprofit aiming to use data-driven techniques to preempt medical errors), Sound Prediction (a digital health informatics company aiming to create transparent AI models) and the NIH leveraged supercomputing at the Pittsburgh Supercomputing Center (PSC) to design a method for recreating the benefits of AI models in medicine with more explicability.

The root of the team’s approach is multilevel modeling (MLM, not to be confused with multilevel marketing). Through MLM, groups of similar cases are bundled and differential equations are used to identify a limited set of controlling factors for each case, allowing for easier – and more consistent – identification of the model’s reasoning compared to post-hoc analyses of more opaque models.

The researchers designed and applied the AI toward predicting – and explaining – readmission and death among Medicare patients following a hospital visit, training the model on three years of data (2009-2011) and testing it on a fourth (2012). The researchers reported that compared to other methods, the MLM performed well – and, importantly, produced understandable, medically sound explanations of its conclusions (post-hoc explainers for comparison models did not).

For instance, the researchers reported, the MLM was able to understand that patients with positive outcomes were discharged to their homes because they were less sick – not the other way around. The overall result: a simpler, more explicable, medically accurate model.

The differential equations that power the researchers’ MLM approach are best served by GPUs – and for those, the researchers leveraged the Bridges-2 system at the Pittsburgh Supercomputing Center (pictured in the header). Bridges-2’s 24 GPU nodes each have dual Intel Xeon “Cascade Lake” CPUs and eight Nvidia Tesla V100 GPUs. Both those nodes and Bridges-2’s four extreme memory nodes (each with 4TB of memory) came in handy when developing and running the model, along with the system’s Nvidia InfiniBand 200Gb/s interconnect.

“What we really like about Bridges is [that] the I/O speeds are very fast,” said Josh Chang, co-founder of Sound Prediction and a director at Mederrata Research, in an interview with PSC’s Ken Chiacchia. “The scheduling [of the computations] also works very well. Things ran much faster at PSC than on equivalent [commercial] systems.”

To learn more about this research, read the reporting from Ken Chiacchia here.