The complexity of clinical trials in critical illness remains a significant challenge. A discussion at the ISICEM Congress in Brussels focused on why trials often fail, the role of heterogeneity in patient responses, and emerging methodologies to improve our understanding of pathways to outcomes.
For over 30 years, researchers have debated why clinical trials in critical illness struggle to deliver clear results. Roger Bone's early work highlighted the interplay of inflammation and anti-inflammatory processes, which fluctuate in real time and are difficult to control for at the bedside. A decade later, John Marshall underscored the variability in responses to interventions, particularly in sepsis, where outcomes are influenced by underlying heterogeneity in patient populations.
Heterogeneity extends beyond patient demographics to differential responses to treatment. For instance, in acute respiratory distress syndrome (ARDS) and sepsis, subphenotypes—such as hyperinflammatory and hypoinflammatory responses—have been identified and are now being tested prospectively to determine their influence on treatment efficacy.
A key concept in understanding these complexities is computational irreducibility, as proposed by physicist Stephen Wolfram. This principle suggests that predicting the behaviour of a complex system requires following each step in its evolution, as there is no shortcut to forecasting patient outcomes in critical care settings.
The discussion of two major trials on acid suppression in the ICU—the SUP-ICU and REVISE trials—illustrates how patient severity affects treatment outcomes. While proton pump inhibitors (PPIs) reduced gastrointestinal bleeding, their impact on mortality varied by risk level: lower-risk patients showed improved outcomes, whereas higher-risk patients had potentially worse survival rates. A meta-analysis confirmed these findings, raising questions about the mechanisms behind these trends and whether certain preventive measures might inadvertently increase mortality.
Clinical trials have long sought to distinguish between the cause and mode of death, recognising that fatal outcomes typically involve multiple intermediate steps. Similarly, studies in ARDS confirm that most deaths result not from refractory respiratory failure but from complications leading to withdrawal of care.
Recent research suggests that pre-existing inflammatory phenotypes influence the mode of death. Hypoinflammatory patients often succumb to respiratory failure, while hyperinflammatory patients are more likely to die from circulatory failure. These findings highlight the need for a deeper understanding of patient pathways beyond initial treatment decisions.
Adjudicating causes of death and treatment outcomes manually is labour-intensive and subject to variability. Emerging artificial intelligence models, such as the C3PO project at Mass General, demonstrate how natural language processing (NLP) can improve accuracy in identifying heart failure admissions. By training AI models on electronic health records (EHRs), researchers achieved an 80% accuracy rate, significantly outperforming diagnosis codes alone. This approach has the potential to transform critical care trials by automating data extraction and improving endpoint classification.
Mega-ROX, an upcoming 40,000-patient trial, exemplifies the shift towards registry-based trials that rely on distributed data networks and EHRs. The integration of machine learning in these trials can enhance our understanding of patient trajectories, allowing for more precise treatment approaches tailored to individual patient responses.
Mortality remains a critical endpoint in clinical trials, but embracing both baseline and pathway heterogeneity is essential. The evolution of AI and large-scale data analysis offers promising avenues to refine our understanding of complex critical care pathways. Future research should focus on integrating computational models and registry-based trials to enhance outcome predictions and improve patient care.
Source: ISICEM 2025 Presentation
Image Credit: ISICEM 2025
