AIMed hosted AIMed Intensive Care Unit (ICU) virtual conference yesterday (22 September). Close to 20 field experts presented on topics around data; machine and deep learning for decision support; natural language processing (NLP) and workflow, and general issues in artificial intelligence (AI) and ICU during the half-day long livestreamed webinar. This also marks the first time AIMed conducts a real-time application of AI session to provide a step by step guide on the thought processes needed to solve an actual clinical problem and to host a journal club which reviewed top three AI in ICU paper published last year.

The unique ICU challenges

Dr. Robert D. Stevens, Director, Laboratory of Computational ICU, Johns Hopkins University School of Medicine opened the event by highlighting reasons for an AI-ICU meeting. He said intensivists are often confronted with a large amount of simultaneously streamed, granular data. Yet, it’s unsure if they are able to incorporate the information to make the right decisions. In ICU, the time-scale for decisions is compressed as the progression of diseases are rapid and catastrophic. Many times, a sound judgement has to be formed within minutes or even seconds.

Besides, ICU staff are expected to shoulder multiple responsibilities at once. They are not only detecting and diagnosing but also instituting therapies. Heterogeneity, progression, and outcomes of diseases and knowledge of patients’ medical history are creating a lot of uncertainties in the realm too and only with time do intensivists have a better idea of the current and future situations. In the subsequent session, Dr. Stevens also expressed his interest in using AI to overcome complex and novel conditions such as COVID-19 and if it is better than traditional approaches to help us understand its pathology.

Opportunity 1: We have the tools but not the infrastructure

With that, Dr. Randall C. Wetzel, The Anne O’M Wilson Professor of Critical Care Medicine at ​Children’s Hospital Los Angeles​ and Professor of Pediatrics and Anesthesiology at USC Keck School of Medicine thought it would be interesting to have unsupervised learning or other machine learning and classification methods to capture and analyze patients’ data in near real-time, so that they know the concentration of patients arriving in the ICU and look specifically into those that require additional attention and understand the complexity of the disease.

Dr. Wetzel also mentioned his team had accumulated a list of over 1500 comments and case discussions from intensivists in the early days of the pandemic. They saw the emergence of the Kawasaki-like features very early one when others were talking about it but they were unable to use NLP and classify the information. Dr. Wetzel believes if only good NLP techniques were present to categorize the details and connect people with similar interests, that would “super-charge” our abilities to deal with what’s about to come. “So, we have the tools but we don’t have the infrastructure,” he concludes.

Opportunity 2: Democratization and standardization of data

Following that, Dr. David Ledbetter, Principal Data Scientist and Manager at the Virtual Pediatric Intensive Care Unit of ​Children’s Hospital Los Angeles said we are probably lacking a critical component and that is democratization of data. “it’s difficult enough for us to share retrospective datasets, the prospect of being able to share the amount of data necessarily on the scale… is well beyond the scope of what we are prepared to do as a group of institutions”. Dr. Bart Geerts, Consultant Cardiac-Anesthetist and Principal Investigator at AMC Amsterdam added the need to standardize data. “We are disregarding a lot of data… we lack a lot of infrastructure to capture all that data… we will have solved a lot of issues if we can put all that data together and trained models on that”.

Dr. Gilles Clermont, Professor of Critical Care Medicine, Mathematics, Clinical and Translational Science, and Industrial Engineering at the ​University of Pittsburgh thought it would also be worthwhile to mine real-world data or data that are not typically found in the ICU setting. While Dr. Mjaye Mazwi, Director of Translational Engineering, Critical Care Medicine at ​The Hospital for Sick Children brought in the concept of interoperable dataset. However, he felt that the science is far ahead of the present ethical and regulatory framework, so institutions need to be equally forward-looking in order to make data-sharing possible.

On the other hand, Dr. Wetzel expressed the quality of our data especially those found in the electronic health records (EHRs) may not be as good as we will like it to be. Most of the time, data going into the EHRs are not meticulously time-check and not as detailed because the person who is keeping the record is not rewarded by the knowledge that every data point they put on the EHRs would be analyzed and fed to bedside care. Only if that can be done, it may change people’s thought about data.

Opportunity 3: Increase our acceptance of AI

Dr. Clermont commented at the turn of the year, a company in Toronto had already raised the issue that this infection is going to be a pandemic. Other alert systems also gave similar warnings thereafter but no action was ever taken. Dr. Geerts echoed the comment, he felt that the acceptance of AI was clearly not there for some early solutions that we saw. For example, he cited a professor who came up with a number of solutions predicting hospital length of stay and other critical factors to plan ahead during this global health crisis was disregarded because those solutions were not politically tying into what the public would want. Dr. Stevens believes we were taken unaware and if AI has been part of us, we might be thinking about implementing more pro-active data-driven approaches to improve our capability to classify and predict the infection and not waiting for things to happen at the beginning.

Opportunity 4: Bring clinicians to the 21st century

Dr. Wetzel said the pandemic has definitely reflected our shortcomings. “We pride ourselves as clinicians and we are able to fit a symptom and fit it into something we recognize and develop ideas on how to treat it but we failed here. We were faced with a crisis but our attention was to find enough ventilators which we didn’t actually need”. Moreover, Dr. Wetzel said he was involved in data surveillance and pediatric ICU in the past and was still a shock to know that the CDC (US Centers for Disease Control and Prevention) is still waiting for people to fill in cards and mail them for epidemic tracking.

Dr. Mazwi added he found it incredibly frustrating to rely on small, anecdotal experiences that dramatically changed the way that people thought about things and in a way, that’s not doctor-driven but “classic anecdote driven medicine and extraordinarily sort of 19th century”. Dr. Stevens thought back in March, the main way of learning the COVID-19 syndrome was to engage in virtual calls with the laboratory colleagues and asking them what this disease is based on their experiences. “It’s not very data-driven, I think we were misled in many ways”.

Opportunity 5: A new way of sharing AI findings  

Dr. Wetzel recognized the need to convince fellow intensivists to adopt AI through journals. Although there are more and more AI and machine learning based materials showing up in the ICU literature, they are not always complete with details on where the data came from, how the data was cleaned or managed, the computation standard, how to define missing data, performance evaluation of AI techniques and so on. Dr. Mazwi wondered if it’s time for a change.

“I am less convinced that traditional publication is the right way of sharing information about models. Frankly, things like publishing your codes online and having an online repository of de-identified data you used. People can actually test your conclusion against the sample data you used in many trials on their own,” Dr. Mazwi says. Dr. Geerts agreed for a change too. He said when a 30-pages long paper was submitted, it is likely to be cut down into 20 pages, in some ways, it helps to remove nuisance but at the same time, insights or discussions that clinicians can learn from may also be lost.

On a positive note, Dr. Stevens added some top journals are very keen on publishing studies that leverage data science and AI and they will appoint section editors who have some background or expertise in the areas. Some of these papers also showed some forms of consensus standards. For example, a paper he came across recently provided some guidance on how to structure papers, report data and AI methods. “I think a lot more is needed: guidance and standards. I think overall if we just look at the publishing in terms of ICU, we are in the good direction,” Dr. Stevens says.

The virtual event is available on demand 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.