A prelude to the upcoming AIMed ICU (Intensive Care Unit) virtual event taking place on 22 September, AIMed hosted its latest webinar under the theme: key trends in critical care AI (artificial intelligence) on 21 July.

Two of the virtual event guest speakers – Dr. Leo Anthony Celi, Associate Professor of Medicine (part-time) at Harvard Medical School and Principal Research Scientists at the Massachusetts Institute of Technology and Dr. Robert David Stevens, Associate Professor of Anesthesiology and Critical Care Medicine at Johns Hopkins University, were invited to speak with Dr. Anthony Chang, AIMed founder and Chief Artificial Intelligence Officer at Children’s Hospital of Orange County (CHOC) on the essential bits of adopting new technologies into their practice as well as their frontline experiences in combating against the novel coronavirus.

Here are some of the main messages.

Leveraging unsupervised learning techniques During the webinar, Dr. Chang mentioned unsupervised learning and deep reinforcement learning as the two commonly discussed methodologies in the AI World recently and he wondered if any of the guests noticed a similar trend in the AI and clinical medicine World too. With that, Dr. Celi said his research team is continuing with some of their previous work but adopted a pandemic approach. They are using unsupervised learning to single out COVID-19 phenotypes and if patients’ responses to certain treatments are dependent on these phenotypes. As for Dr. Stevens, his group is doing a fair amount of unsupervised clustering on non-COVID data sets but he did not elaborate on the details.

Nevertheless, Dr. Stevens thought the implementation of unsupervised learning techniques is valuable especially during a global health crisis when scientists are faced with an extremely complex and heterogeneous disorder with many presumed biological subtypes. He believes unsupervised learning techniques have the potential to bring us closer to specific genetic clusters that are associated with higher risks of disease; disease progression and criticality. “I think those are important domains; a subject that should be actively pursuit,” Dr. Stevens says.

Always question an AI model

At the same time, Dr. Celi cautioned the use of mortality prediction models. He explained these models were often built with an assumption that they will forecast when a patient is likely to pass away but in reality, what they project is whether treatments are going to be discontinued for the patient. This is especially so in the ICU when mortality tend to signify cessation of treatments. He brought the issue up as he feels this kind of model is picking up on popularity; regarding it as straightforward without giving a thought on possible biases.

“CDC (US Centers for Disease Control and Prevention) advised on the use of SOFA (Sequential Organ Failure Assessment) on intubated patients and I believe it’s greater than 12 is when we have consider removing patients from mechanical ventilation. I felt unease while reading at the recommendation because you could easily magnify disparities in end-of-life care by using such data to build your model. So, my key message is be very careful with mortality prediction model, it’s not as straightforward” Dr. Celi cites.

Do not undermine group effort

Both speakers agreed researchers and institutions are more willing to share data during the pandemic. However, they feel that’s not enough, in an ideal situation, all codes and repository by which analyses are being performed should also be shared. “I think this is crucial because it allows other people to replicate your investigation and to proof that you did not crunch those numbers to get the conclusions that you have wanted. One is held more accountable,” Dr. Celi explains.

Speakers also noted on the noise surrounding COVID-19 research. “We were drowning in small studies that are not well-controlled and observational studies that are clearly confounded with so many factors had made it to the media,” Dr. Celi says. Yet, on the positive side, the phenomenal amount of research work also ignited active public peer-review and demonstrated some consensus in terms of what constitute as good data analyses and data sharing. This, hopefully, will allow the community to learn as a whole in the long run.

Dr. Stevens agreed the part on group effort and told a similar story. In February, Johns Hopkin established a crisis command center and converted a number of non-ICU areas such as the recovery rooms into ICU. Specialists from different medical sub-specialties were also repurposed into ICU residents. For example, he met a neurologist who became his resident during his ICU round. “I think all these taught us, health system can adopt very rapidly because a lot of good will and that people are willing to contribute even at the risk of their health and lives. We learnt a great deal as far as what we are going to do for the future,” Dr. Stevens adds.

The webinar is now available for re-visit here. Register your interest and obtain a copy of the AIMed ICU virtual event agenda here.


Author Bio

Hazel Tang A science writer with data background and an interest in the current affair, culture, and arts; a no-med from an (almost) all-med family. Follow on Twitter.