Causal physiological models

Physical therapy is a profession, a practicing profession. It’s professionals: “diagnose and treat individuals of all ages, from newborns to the very oldest, who have medical problems or other health-related conditions that limit their abilities to move and perform functional activities in their daily lives” (APTA). A core component to the knowledge used in practice for this profession is physiological knowledge. If you understand physiology you can quickly move to an understanding of pathophysiology. You may not know all the language (categories) of all the pathological conditions, but you can reason through the consequences of various “physiology” becoming “non physiology” or pathophysiology. It is fairly safe to say that much of what happens at the causal model level of knowledge for practice includes physiology.

I have been promoting the use of causal models are representations of the knowledge for practice, and the use of graphical causal models for the clear articulation and sharing of such models. I have done further to suggest that DAGs (directed acyclic graphs) offer several advantages as graphical causal models. I still believe that, but there is a caveat that must be considered. DAGs are acyclic which means there are no feedback loops. Physiological models are full of feedback loops. Physical therapists often consider feedback loops in practice (i.e., immobility leads to deconditioning, deconditioning leads to immobility).

Do we therefore need to reconsider our use of DAGs as a core of graphical causal modeling for KBP? Or do we need to reconsider the use of feedback loops in causal physiological models?

I plan to reconsider the use of feedback loops in causal physiological models - at least models at the level of inquiry most interacted with by physical therapists (systems, organism, environmental).

Let’s look at the above feedback loop graphically (note that DAGitty automatically highlights the feedback path to alert the user that there is a cycle in the “acyclic” graph):


Fatigue and/or dyspnea on exertion (DOE) result in immobility (a relative decrease in mobility), which leads to deconditioning, which then leads to more fatigue and/or DOE.

I believe we can handle these loops with DAGs by considering how the events are occurring across time, and the state of the system being studied across time (which clearly changes). So the above could be considered as a DAG across three iterations (changes in state over time):


Here we see the same consequences, but the models explicitly recognizes the changes in state over time. Which, I believe, can be a helpful aspect of using DAGs in the modeling of the causal processes at play in a clinical scenario. After all, to get back to the initial state a patient must traverse each state along the way, unless there are jumps in state. If we can anchor the states with other variables that would then change over time we can consider how far, and by what path someone may return to a previous level of functions.

In summary, the fact the DAGs are acyclic does not provide a barrier to their use as representations of clinical practice knowledge, but rather an opportunity to consider the cycles we commonly refer to as they play out over time and the state of the system change over time.

Leave a Comment