The COVID-19 pandemic is running its apocalyptic course around the world. One thing is certain in the midst of total uncertainty and utter chaos: SARS-CoV-2 is a very daunting enemy and we remain totally subjugated by these viral overlords. It is easy, therefore, for we humans to lose patience and become careless as well as to decrease capacity for resilience. We need to regain our composure and to be defiant in order to for us to prevail this human vs virus struggle.

A major part of this resilience and defiance is continue our work in adoption and deployment of artificial intelligence concepts and projects in clinical medicine. Here are key takeaways from today’s Clinician Series – AIMed ICU virtual meeting in collaboration with the Society of Critical Care Medicine:

My top takeaways:

1. The COVID-19 pandemic has emphasized the urgency of data science in the ICU setting and the importance of data sharing amongst institutions.

2. The AI in ICU agenda starts with a focus on data and entire process of data munging/curation. Decisions made during data munging and imputation have effect on performance assessment.

3. Recurrent neural network has a basic structure of a feedback loop with a temporal dimension for ICU data science (vs convolutional neural networks more designed for medical imaging).

4. Feature selection is important but perhaps not as critical as some may think as machine learning can overcome higher numbers of features.

5. The black box perception can be partly neutralized with tools that are available (but these tools lack UI/UX for clinicians)

6. The number of cases needed for an ICU data science project can be as few as 100 minority class examples, and k-fold validation and transfer learning strategies can also help.

7. Organizational focus on data and management is key for data to be the rich resource that it can be for data science.

8. Constraints in prognostication in neurointensive care needs data science to improve the overall accuracy of prognostication.

9. Biological heterogeneity in gene expression as well as recovery and outcome needs to be supported by data science for additional knowledge and insights.

10. Physiological time series can add to accuracy in predicting outcome by pooling data across international centers (data-intensive collaborative efforts underway).

11. Data science represents the key to leveraging these datasets to design more effective neurointensive care clinical trials.

12. Lead time in predicting an adverse event such as a cardiac arrest is key in the prediction models (sooner better than later).

13. Complexity theory and chaos are important (albeit not well understood) aspects of data science in the ICU setting.

14. Artificial intelligence can lead to new endo types and phenotypes in the ICU setting to help with both diagnosis and therapy.

15. In reinforcement learning, the physician is the agent (an entity that interacts with the environment) who gets a reward for an action with a desired outcome.

16. The challenges in reinforcement learning include high risk environment, limited training data, and difficult trial-and-error situations.

17. It is important to bring the algorithm to the bedside for prospective safe validation and randomized controlled trials as well as medical device development and certification.

18. There is over a million physiological waves per patient per day in the ICU and most of the data is not leveraged for data science.

19. Innovative use of convolutional neural network (CNN) to analyze physiological waveforms as a medical image (in ICU).

20. Prediction for outcome can be enhanced with EKG data analyzed with CNN for an overall hybrid clinical stability score.

21. Clinicians can be at risk at being deskilled in decision making in the ICU setting as more AI tools mature in the ICU setting (a form of automation bias).

22. Steps of a data science project: Define the question> Get the data> Clean the data> Enrich the data> Find insights> Deploy model> Iterate.

23. There can be disconnect between data source like laboratory data and diagnosis so there is a need for multimodal machine learning incorporating multiple data sources.

24. There is discrepancy between natural language processing derived and other data sourced outcomes and this will need to be reconciled.

25. Advanced bidirectional NLP models like BERT or GPT-3 may not be superior to more basic NLP models like Word2Vec (neural network model to learn word associations).

26. The major challenge to deal with the COVID-19 pandemic remains issues with data: consistency, sharing, accuracy, etc. while the pandemic is complex and even chaotic.

27. The pandemic has helped to accelerate the process of data integration with breaking down of traditional silos of data and lack of collaboration amongst centers; this forced disruption is also seen with efforts in innovation as well.

28. AI in medicine will be perceived as successful when we no longer talk about AI in medicine as if it is a separate part of medicine but more that AI is embedded in clinical medicine and healthcare in many ways.

29. Syndromic surveillance and dynamic phenotyping will be key for ICU diagnosis and management in the future that can leverage existing AI methodologies.

30. Immediate relevance and actionable insight are necessary ingredients for wider adoption of artificial intelligence in the future.

31. We should share our AI adoption of real-time ICU decision support process and the issues with these adoption experiences.

32. It is our accountability and responsibility to continue to support AI as it is unethical not to capture the data from our patients for analytics.

33. Bias in our ICU data can be partly managed by algorithms built into the data collection process so that we minimize bias being perpetuated into the models.

Thank you faculty and attendees for your knowledge and expertise as we all learned a great deal at AIMed ICU/Anesthesiology!

Anthony Chang, MD, MBA, MPH, MS
Founder, AIMed
Chief Intelligence and Innovation Officer
Medical Director, The Sharon Disney Lund
Medical Intelligence and Innovation Institute (mi3)
Children’s Hospital of Orange County