Collaborative intelligence for intensive care units
Critical Care volume 25, Article number: 426 (2021)
The intensive care unit (ICU) is a rich and complex data environment, well-suited for artificial intelligence (AI) and machine learning techniques. Numerous AI applications are being developed for the management of critically ill patients , both before and during the COVID-19 pandemic . However, it remains unclear how intensive care clinicians can benefit from AI. Learning from the business literature, the concept of collaborative intelligence can help to clarify how humans can synergize with AI . For the ICU, three domains can be identified where AI can augment the human clinician, and vice versa (Table 1).
The first domain is related to accountability and risk mitigation, where AI amplifies human cognition while humans sustain AI . The speed and consistency of digital systems allow AI to help humans perform continuous and rapid multi-channel monitoring, data harvesting, organization and analysis. Such abilities are particularly useful in the ICU, where critically ill patients quickly amass large quantities of clinical data, and clinicians are at risk of monitoring fatigue. Furthermore, AI-driven decisions can help corroborate human decisions. In return, humans can provide real-world stress testing of AI systems, including simulated adversarial attacks, which are intentional contamination of data aimed at causing AI malfunction. Additionally, humans can audit AI algorithms for accuracy and bias, enhancing confidence and trust in AI.
The second domain is related to sense-making, where AI interacts with humans in intelligible ways (i.e., explainable AI ) while humans help explain AI . Rather than merely providing an output that substantiates human answers, AI can produce lists of salient features and probabilities for various diagnoses, predictions and actions, helping humans prioritize and justify decisions. To avoid the “black-box” effect, human clinicians can augment AI outputs by helping interpret these to lay-persons.
The third and final domain is performance augmentation, where AI embodies human skills while humans train AI . Real-time AI-powered ultrasound systems to guide novices in image acquisition and interpretation are commercially available, e.g., Caption AI (Caption Health, Brisbane, CA). At the cognitive level, just like how human players train using computer chess engines, human clinicians can learn from AI-generated decisions and data summaries. Graph data science methods using multivariate time series can reveal novel visual relationships among patient characteristics, treatments and clinical evolution . In turn, AI methods like reinforcement learning depend on real-life data and decision-making. Ultimately, implementation and scaling of AI solutions require human support for digital resources.
Availability of data and materials
Intensive care unit
Gutierrez G. Artificial intelligence in the intensive care unit. Crit Care. 2020;24(1):101.
Chen J, See KC. Artificial intelligence for COVID-19: rapid review. J Med Internet Res. 2020;22(10):e21476.
Wilson HJ, Daugherty PR. Collaborative intelligence: humans and AI are joining forces. Harv Bus Rev. 2018;96(4):114–23.
Amann J, Blasimme A, Vayena E, Frey D, Madai VI, Precise QC. Explainability for artificial intelligence in healthcare: a multidisciplinary perspective. BMC Med Inform Decis Mak. 2020;20(1):310.
Martinez-Aguero S, Marques AG, Mora-Jimenez I, Alvarez-Rodriguez J, Soguero-Ruiz C. Data and network analytics for COVID-19 ICU patients: a case study for a Spanish Hospital. IEEE J Biomed Health Inform. 2021;25(2):4340–53.
No funding was required for this study.
Ethics approval and consent to participate
Consent for publication
KCS has received honoraria and travel support from Medtronic and GE Healthcare.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
See, K.C. Collaborative intelligence for intensive care units. Crit Care 25, 426 (2021). https://doi.org/10.1186/s13054-021-03852-7