Why do some people make full recoveries from stroke and not others? What triggers a rise in blood glucose in the morning? How do viruses spread?
A fundamental goal of most sciences is discovering the rules behind why things happen. To successfully interact with the world by predicting future events, explaining the occurrence of phenomena, and intervening to produce particular outcomes, we cannot simply describe how a system behaves, we must know why it behaves as it does.
Say we learn that seizures and poor stroke outcomes are associated. If this correlation is due to an unmeasured cause of both, then outcome predictions based solely on seizures can fail unexpectedly. If seizures are predictive of a poor outcome just because they tell us about the severity of brain injury, then we cannot explain a patient’s outcome as being due to seizures. Further, an intervention treating the seizures not only may not improve outcomes but may also cause harm (due to the side effects of medication).
Beyond the need for causality, patients are highly complex systems that pose many challenges for traditional causal inference methods. The body’s various subsystems are densely interconnected and affected by properties of the whole; they are robust and contain backup mechanisms; and they change at many timescales. Depending on the level of temporal granularity, a seemingly nonstationary system may actually be cyclic (e.g. circadian rhythm), and new causal relationships may emerge over time in response to environmental triggers.
To face this challenge one might look to the field of complex systems. However, this research has often focused on networks (without causality) or complex causal relationships between two variables (ignoring large densely connected structures), which can find spurious relationships between effects of a common cause (such as yellowed fingers causing lung cancer if both are due to smoking). At the same time, computational methods for causal inference have not been able to handle complex systems, focusing instead on simple relationships between individual variables.
This work aims to unite causality and complex systems to address two key questions whose answers will advance computation, biomedical applications, and philosophy.
First, what are the elements of a causal relationship and how can we discover these automatically? Instead of limiting inference to user-defined sets of variables, we aim to discover relationships involving system properties and ultimately to find what causes a causal relationship to emerge. This means we will be able to find relationships like “when homeostasis of the brain is disrupted after a stroke, new compensatory links are created that bypass affected systems.”
Second, systems are organized on multiple spatial and temporal levels, but how do these interact? Rather than abstracting data to the same timescale or understanding these in disconnected ways, we will develop methods for inferring the temporal and spatial relationships between cause and effect. Through ongoing collaborations with clinical and biomedical researchers, the methods developed will be applied to better understand intensive care and outpatient data and improve human health.