Differential diagnosis can mean at least two things to PTs and considering them separately is not necessarily bad. However both basically come down the same thing. In one sense we think of differential diagnosis as screening for medical issues outside our scope of practice for referral. In another sense we consider differential diagnosis as it is, a complex reasoning process to identify the cause of some sign, symptom, or set of signs or symptoms. Due to causal chains we may be involved with multiple differential diagnostic processes with a single patient. For example,we may find that the cause of an activity limitation is pain; the cause of pain is a trigger point; the cause of the trigger point is bad posture; the cause of bad posture is tight muscle X, weak muscle Y; the cause of tight muscle X, weak muscle Y is a job sitting all day with no control over the daily routine, structure or schedule (i.e. low job control). You get the point.

We then attempt to intervene to remove (reverse, adapt, compensate) as many of the causes as possible (those more proximal to the signs or symptoms that are limiting function are typically addressed first), but eventually we attempt to broaden our interventions. Of course, as we move distally there be an increase in the number of variables in the causal structure (for example, imagine trying to come up with the causes of low job control). For those interested in seeing what I mean, I have taken the above simple chain and expanded it based on one possible confounder (stress and high sympathetic activation) in a DAG in the Physical Therapy DAGs GitHub repo at:


This particular sequence is in the TrPtPain folder (trigger point pain) with figures and explanatory text and starts with this model as a base:


In the repo I provide the consequences of fixing pain as the outcome and attempting to identify it’s cause which becomes more complicated because of the confounding influence of sympathetic activation; then to fixing the trigger point as the effect and attempting to identify it’s cause (also confounded). For more on confounding - it is helpful to see the posts on types of cause (exacerbating factors, etc).

Back to differential diagnosis. These two perspectives are ultimately the same process and therefore can both take advantage of a common understanding of causal models and reasoning. What makes them different is that within the process we may identify a cause within the causal network for a patient that we suspect requires further examination by someone else with a scope of practice and expertise to diagnose (identify the cause) and/or intervene on that cause. At that point we refer.

It is more efficient if we can identify such causes earlier since in the decision tree of examination / evaluation when you need to refer for a proximal cause you would rather find out right away, rather than doing a complete examination / evaluation and then figuring out you need to refer (saves everyone time). And this is why it is effective to consider differential diagnosis as two separate things. First we screen (differential diagnosis), then we examine/evaluate (differential diagnosis).

There are two reasons to consider these two separate processes as one process of differential diagnosis. First, since the reasoning is the same it is useful to understand the reasoning generally and how the slightly different sets of knowledge contribute to effectiveness in each. The sets of knowledge for each ultimately come down to sets of causal models that share effects and have different, possibly interactive, causes. For example, pain as an effect can have a causal model that we learn about in musculoskeletal PT related to muscle dysfunction, or it can have a cause we learn about in pathology related to liver dysfunction. Those are different causal models for sure, but they overlap minimally at their terminus (pain as the effect). For example, in this very simplistic causal model we consider multiple possible causes of thoracic pain which we know (adjusted) about as a cause of the activity limitation.


If a patient is referred to the PT then the referring provider has most likely come to the conclusion that the Thoracic Pain is originating from a cause that is within the PT’s scope of practice. So the likelihood of identifying a cause out of the PT scope is smaller (which makes up most PTs experiences thus far). As such the PT may look more quickly through a screen (if at all) of other possible causes.  However, If a patient walks off the street, then the likelihood of the possible causes that do not fall within the PTs scope of practice increases to at least the level of the prior probability of that condition. It is important to keep in mind that most therapist’s experiences are not with truly direct access patients. But with direct access PTs need to be mindful of the very important difference between those referred and those not referred. Those not referred have a higher probability of causes falling outside the PT’s scope.

Another feature of the causal structure (very common structure) in this DAG is something called an “inverted fork” where there are multiple causes for one effect. It turns out that it is well known that when this occurs and when there is adjustment on the effect being converged upon it can create the appearance of an association between the otherwise independent causes. For example, if we put 1000 patients with thoracic pain in a room, and the causes of thoracic pain are truly independent from one another, then having one cause rules out all others, giving the appearance within that sample that the causes are associated (see this post on selection bias). Studies investigating diagnostic test reliability have to obtain their samples somehow. They often do so by selecting people with a certain characteristic. Depending on the causal structure and the characteristic chosen, such studies can easily display selection bias, creating associations where they do not exist leading to errors in estimates of likelihood.

Well, as stated this post was on “differential diagnosis - some thoughts”. At this point I am not prepared to share any major insights on the use of causal models and DAGs (as above and in the repo referred to) in making differential diagnosis more effective and efficient, or on how to teach this to students differently than how it is currently done. But I do believe that simply pointing out the shared features, the common use of reasoning, the use of causal models and DAGs would be a helpful step and I do plan on exploring this topic further. As always - if interested in contributing, let me know. Once you learn the methods the application starts to come easily.