Differential Diagnosis & Marr’s 3 Levels, as a Paradigm for Integrating Literatures

By Luke Hafermann

This article describes a process for integrating incongruent terminologies, drawing on the concept of differential diagnosis from medicine, as well as Marr's 3 Levels of Analysis. I came up with the idea in order to better organize the literature review for my QSS thesis, & I'll probably put most of this into my thesis paper, following a little section called "Justification for Proposing New Terms", which hopefully assures the reader that I'm handling carefully their established terminology-efforts, and means that they're now on board with me synthesizing some new terminology (not included here; this wiki's just the process!).

Differential Diagnosis

In medicine, using the wrong terminology to describe a patient’s symptom-pattern is dangerous, so practitioners are cautious about assuming that the current diagnosis is correct. *Differential Diagnosis,* as a paradigm, includes engrained methods for differentiating between two diagnoses (or integrating components of two diagnoses) in order to better explain a patient’s symptom-pattern. Differential Diagnosis, with its focus on the patient’s well-being, attempts to avoid undue bias toward one diagnosis (set of terms) over another.

Differential Diagnosis inspires this paper’s approach toward integrating the attention and memory literature, since it provides a neutral process for creating consistent terminology without favoring one literature over another. It includes guidance on 1) what to do with duplicate terms & 2) what to do with a term that (while important within one literature) fails to capture an important distinction made by the other literature.

These 2 considerations are first illustrated within a relatively straightforward scenario, involving Differential Diagnosis of ADHD-like-symptoms, when no further integration is needed. Here, the patient’s physician can draw upon a hypothetical, neatly-integrated theory of ADHD which documents 2 potential causes of this symptom pattern. Diagnosis then involves a relatively straightforward process of identifying the subset of causes present for this particular patient (either both, either, or none), and reflects the value of a well-integrated theory where the causes are modular, non-overlapping, and mutually compatible. It is not even clear in this context why a “duplicate term” would arise, & each term easily incorporates additional distinctions as necessary (without compromising its use in other contexts). This straightforward scenario illustrates an ideal where there is no further work to be done on the terminology; the integration has already been achieved.

Differential Diagnosis also however includes a process for integrating incongruent diagnoses, & the second diagnostic scenario showcases this. As above, sometimes two diagnoses seem to duplicate each other at the symptom-level, with each potentially giving rise to the same pattern of symptoms reported by a certain patient. Contrary to above, the two overlapping symptom-patterns (anxiety and ADHD) have not yet been deconstructed by cognitive science into traits which are modular, non-overlapping, and mutually compatible. In the case where an initial diagnosis of anxiety does not respond to treatment as anticipated, the entire initial symptom category (anxiety) must now be scrapped & rethought, since none of the patient’s history & self-report surveys (framed in the terminology of anxiety) are immediately reinterpretable from the perspective of ADHD. The old documentation is organized according to traits from the anxiety model (where each trait groups the attributes as they commonly cooccur in the context of anxiety); ADHD, however, would group these very same attributes differently. Thus, the documented anxiety-traits must be decomposed into more-granular attributes that can then be recomposed into the ADHD-terminology’s preferred groupings. This recombination-process is intensely demanding of a medical practitioner, because expanding a named trait into more-granular attributes involves naming these attributes (and thus inventing an idiosyncratic & patient-specific terminology) as the old documentation is reanalyzed in terms of new patterns. An alternative approach would be to restart the diagnostic process from scratch in light of ADHD, & reuse none of the old documentation beyond an intuitive gist-based reframing. In summary: anxiety’s attribute-groupings do not map cleanly onto ADHD’s attribute-groupings, and thus the process of Differential Diagnosis impels practitioners to either perform an ad-hoc integration of the two terminologies, or scrap one terminology (& thus their effort on documentation) in order to try out the other diagnosis. This is a friction, and a cost in terms of the practitioner’s time and the patient’s delayed (or *never* successfully reached) diagnosis. It is a friction that would be reduced by a better-integrated terminology between the two symptom-patterns. This would mean standardizing a more-granular set of traits, perhaps in terms of underlying cognitive processes such as attention & memory, which can be satisfactorily grouped into a higher-level tier of ADHD traits, and grouped in a different way to produce a higher-level tier of anxiety traits. The underlying traits would thus be modular, non-overlapping, and mutually compatible, and would comprise an integrated terminology for the symptom-patterns & treatment-responses of both diagnoses.

An Integrated Terminology for How to Integrate Terminologies

Differential Diagnosis thus demonstrates the (medical) benefits of well-integrated terminology, and also illustrates an ad-hoc process *for integrating* terminology. In the following section I continue to describe this process more formally, by interacting it with related concepts drawn from Marr's 3 levels of analysis & Object Oriented Programming. Eventually I develop a fully granular set of terms which can satisfactorily explain all 3— Differential Diagnosis, Marr’s 3 Levels, & OOP. If this feels self-referential, in a “I’m so meta, even this acronym __ ____” sort of way, it should. It’s integrating terminologies for how to integrate terminologies; if the “process for integrating terminologies” were an operating rule (that operates on terminologies), here I portray an operating rule which can operate upon itself, and which thus comprises its own learning rule. The author finds this delightfully recursive in a Godel Incompleteness Theorem/Hofstadter’s Strange Loop kind of way.

A concept from Marr’s 3 Levels (and another it misses)

Here I elaborate a useful concept, abstracted from Marr’s 3 levels of analysis: the notion that 2 systems can appear the same at the specification level, while arising from 2 different sets of underlying mechanisms. In other words: two different algorithms can achieve the same specification, and if hidden inside a black box, they would be undistinguishable (since their behavior is isomorphic). Diagnosis thus involves reaching inside the black box to glimpse some component of the internal mechanisms. If two diagnoses manifest an identical symptom pattern, but respond differently to treatment or testing, then the act of treatment or testing effectively lifts a component from within the black box up to the surface, and takes something from the algorithmic level and inserts it as part of the specification; there’s now a distinction which differentiates the two algorithms, & which differentiates the diagnosis. Marr’s 3 levels (as far as I am aware) does not describe a process for moving something from the algorithm level up to the specification level, which is the entire focus of differential diagnosis. I will call this process “un-black-boxing” an attribute, or “un-glossing-over” a newly relevant distinction. I welcome additional candidate terms.

Triskaphobia (of Marr’s 3 levels)

Triskaphobia, quite prudently, advises us to beware of the number 3. Not all triskaphobias are rational, but this particular version, on the contrary, is quite well-founded: it serves as a counterheuristic, a protective mechanism, against humankind’s broader tendency to favor the number 3 when coming up with tidy explanations of reality. I suspect we overindex this number, favoring it as the proper number of categories within mental models, due to some combination of its mnemonic effectiveness and then applying the overgeneralization bias upon instances where the mental model’s 3 categories do happen to map well to reality (or when they pass an F-test with reality, so to speak, regarding how many variables to include). You can observe this overgeneralization bias internally, as a sense of finality and completeness (ie, the sensation as a conditioned response) which arises when encountering novel groups of 3. (A good experiment might be to see whether this self-report increases after priming participants with practical & time-tested trio-based mental models, such as that of pathos, ethos, and logos; to stop, drop, and roll; or to look both ways (and back again) before crossing the street.) Regarding the actual mnemonic effectiveness of “3” within mental models, it likely reflects a tradeoff where smaller numbers (i.e., remembering 2 causal factors instead of 3) have a reduced explanatory power for the world, yet larger numbers unduly burden WM capacity (with their additional details displacing the very cognitive processing which would seek to utilize them). This tradeoff & bias forms an attractor landscape between the numbers 3 4 and 5, as evidenced by cultures whose oral traditions note that the use of 2, 4, or even 5 categories is more common. Therefor the number 3 is best approached with caution when found within mental models, and the attentive reader should now be primed to explore dimensionality-reduction of any 3-factor model they come across, or at least to delight in the frequency illusion of 2-factor models appearing now seemingly everywhere… (and those with pattern-recognition abilities verging on schizophrenic might even take note of this paragraph’s 2-factor model of Triskaphobia.)

• Prime example: theory/practice only has two levels. (it’s an illustration of black-boxing bc without the diagnostic, the two symptom patterns appear the same. The surface appearance is identical.) And look, Marr just did black boxing, but with 3.
• So black-boxing is the cooler thing, bc combined with differential diagnosis (between multiple models, and using one to *predict the other)* it basically produces all the cool stuff that Marr’s 3 levels does.

Another thing which Marr’s 3 levels doesn’t have terminology for, and how this (makes more likely) a fallacy called “mistaking the map for the territory”

• Example of “mistaking the map for the territory”: misdiagnosing ADHD as anxiety. This isn’t necessarily a problem with the diagnostic process; sometimes you have to try out a diagnosis, & the implicated diagnostics & treatments, before you move on to testing another diagnosis. “mistaking the map for the territory” is more when you’re not aware there’s a difference, and to be appropriately cautious (and ready to revise the map).
• (Marr’s 3 Levels doesn’t have many practical tips for *noticing when 2 things aren’t the same at the specification level; fortunately, Differential Diagnosis does)*
• Thus, Marr’s 3 levels is insufficient to ward off the following problem, which arose in the schema/attention literature. Remember that if you say “schemas were storing _ in his brain,” this is a shorthand, & we should stay attuned and aware of what it would look like if our schema theory is being extended into territory it no longer applies. Marr’s3 should take a lesson from Differential Diagnosis.