Causal inference using observational intensive care unit data: a scoping review and recommendations for future...

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Causal Inference

Background:

• Actionable artificial intelligence (AI) requires reliable models for causal inference to support clinical decision-making.
• Intensive care units (ICUs) present complex scenarios with time-varying treatments, necessitating advanced causal inference methods.

Purpose of the Study:

• To conduct a scoping review of studies using causal inference models in adult ICUs.
• To evaluate the reporting quality of these studies regarding target trial components and causal assumptions.
• To provide recommendations for enhancing future research practices.

Main Methods:

• Systematic search across multiple scientific databases (e.g., Embase, MEDLINE, Web of Science, arXiv).
• Inclusion of studies on causal inference models for time-varying treatments in adult ICUs.
• Data extraction on study settings, methodologies (G methods, reinforcement learning), treatment regimes (static, dynamic), target trial components, and causal assumptions.

Main Results:

• 79 studies met the inclusion criteria from 2184 titles.
• G methods (61%) and reinforcement learning (39%) were the primary methodologies.
• Significant gaps in reporting were identified: only 38% reported all target trial components, and 9% mentioned all causal assumptions.

Conclusions:

• Current reporting standards for causal inference models in ICU research are insufficient for developing actionable AI.
• Recommendations include clearly defining the causal question as a target trial emulation, employing appropriate methods, and rigorously assessing causal assumptions.
• Improving reporting quality is crucial for advancing AI in critical care settings.