The Semantic Memory of Physicians - and More...

 

I have the somewhat grandiose plan to model psychiatric diagnosis based on the cognition of a physician rather than focusing on the externals.  By the externals I mean classification systems and critiques of classification systems.  At a later date – I might try to comment on how this approach compares with AI.  For now, I will try to keep it focused on human diagnosticians.  I have an interest in this is because I have made and witnessed incredible diagnoses and treatments by physicians and psychiatrists who I have been affiliated with. I don’t think there has been much of a focus on the process.  A secondary consideration is that cognitive neuroscience is a neglected subject in psychiatry and I hope to make the point that should change. I would go as far as suggesting that cognitive neuroscience should be taught to all psychiatrists more urgently than focusing on another DSM.    

Since the early 1970s, memory functions are divided along various lines clinically and functionally. The first division is long term memory and working memory (also called short term memory).  On the long-term side there is a further division to declarative and procedural memory.  Declarative memory is divided into episodic and semantic memory.  Episodic memory is the ability to recall discrete events.  Semantic memory can have a number of graded definitions.  A minimalist definition is factual knowledge independent of the source (7). A definition more informed by recent research in cognitive psychology: “General (encyclopedic) knowledge as well as schematic representations of events distilled from lifelong experiences, retrieved independently from their original spatial or temporal context” (9).  The authors in that case give examples of knowing who wrote the book “1984” and what generally happens at a birthday party.  That naturally raises the question how does all of this freestanding knowledge occur in the first place?  And also – does that imply a connection to episodic memory? In other words, does semantic memory occur when the context surrounding episodic memory is forgotten?

In the case of physicians there is a very long list of formative experiences across the course of one’s career.  The ability to recall them often assists in making diagnoses and provides an advantage over a physician who has not experienced that event.  Semantic memory is about concepts, words, and their relationship independent of a specific event or experience.  It typically consists of a collection of general facts and word meanings.  For example, it would include facts that apples can be red, green, or yellow and what a mechanic does.

Anyone familiar with cognitive screening examinations has probably asked questions focused on semantic memory.  Naming, word similarities, verbal fluency by word generation, general knowledge questions, are all examples. 

The semantic memory of a physician will contain many unique concepts and they will vary based on experience and exposure to clinical scenarios.  The general categories can be described as the following:

1:  Meaningful prior experiences – even though episodic memory stores specific events at specific intervals, semantic memory contains the specific meaning.  In the case of psychiatry an example would be seeing the effects of CMV encephalitis in a major university transplant unit and a decade later seeing similar behavior and consulting on a case in a general community hospital for similar findings.  That similarity triggers non-analytic hypothesis generation.

2:  Prototypes - the patterns noted in the above example can be averaged over a group of patients and those averages can be consolidated into prototypes.  In the above case a psychiatrist may have seen many cases of encephalitis and many cases of meningitis resulting in encephalitis and meningitis prototypes.  Similar prototypes may exist for all major neurological, medical, and psychiatric condition that they have encountered.  Note that the prototype differs from diagnostic criteria (the typical focus) because it is recall of all of the relevant and in many cases unique clinical features that were experienced.

3:  Specific patient memories (exemplars) – all physicians recall specific patients.  These memories are important for non-analytical reasoning like pattern matching.

4:  Knowledge Encapsulation – medicine like most professions is based on a system of graduated learning.  Basic science transitions rapidly into clinical medicine and then into clinical practice and lifelong learning.  At each stage prior knowledge is reorganized in a more efficient way.  In this case – general biomedical knowledge from basic science is organized under higher level concepts. 

An example in one of the references is a person with an infection who is experiencing progressive physiological problems.  At the medical student/basic science level the analysis might proceed from the basic science level and pathophysiology first.  At the clinician level the relevant pathophysiology is organized as sepsis and that provides a more immediate pathway for intervention.  The encapsulation encompasses and efficiently organizes the lower-level information.  At the same time experts must retain a significant amount of that earlier information.  

5:  Illness Scripts – are mental representations of diseases containing three different dimensions.  The first is enabling conditions like risk factors, demographics, predisposition, and context.  The second is fault or underlying pathophysiology.  The third is consequences including signs, symptoms, lab findings, and course or natural history.  Experts have a significant collection of these features.   

One of the questions in this area is what kind of illness script do physicians have?  Should they all be from their particular specialty or should these scripts encompass the totality of their training?  Some authors suggest that the pathophysiological mechanisms from basic science needs to be retained for true expertise – so my conclusion is that the illness scripts from the entirety of a physicians training probably remain relevant.

This is important in psychiatry because the general pathophysiology important in today’s environment was probably not taught is any detail in medical school and most conditions that are not secondary to medical conditions or the effects of drugs do not easily lend themselves to physiological explanations.  I would suggest that medical stability, generalized seizures and seizure variants, increased intracranial pressure, meningitis, encephalitis, cerebral localization, cerebellar dysfunction, peripheral neuropathies, coma, confusion/stupor/delirium, intoxication, and cranial nerve deficits are some of the illness scripts that every psychiatrist must have.

6:  Semantic Qualifiers - every physician has a lexicon of semantic qualifiers acquired in both medical school and post graduate training. They include anatomic descriptions (areas, more specific locations), pathological descriptions, disease course descriptions, and many others. Framing clinical scenarios with these qualifiers is often all that is needed to acquire associations to the disease of interest.

7:   Base rates and Context – experts by way of their clinical practice have an intuitive grasp of the base rates of various clinical conditions and how they typically present in their practice.  These rates of presentations and findings are integrated with the other features of semantic memory (disease scripts, patterns, etc) for more analysis and hypothesis generation.

These features of semantic memory are of course models of brain function for the most part determined by experimental models in cognitive psychology. Examples include testing for specific functions and seeing how those modelled functions vary among trainees and experts at various stages of development. 

Apart from the descriptive approaches used in many studies on physicians at various levels of training are there any more general models that could apply?  Cognitive neuroscience and cognitive psychology offer a more complete model of memory and knowledge structures as well as the underlying biology.  The lead figure for this post is a case in point and has the potential to consolidate many of the descriptions under a more comprehensive model based on experimental validation.

At levels B and C in the diagram we see a perceptual episode being processed from the left to the right in the diagram.  The activated or instantiated schema is a template for extracting relevant features and repressing irrelevant features.  In the diagram circles represent general concepts and squares are action scripts. Gist in the case of the model is a representation of a single episode where much of the detailed information is removed.  The overall sequence at level B depicts how a schema serves to form semantic type memory (gists) and at the same time can be altered or accommodated by new information.

Level A in the diagram illustrates what is known about the localization of these processes largely from human fMRI and preclinical studies.  Memory schemas are stored in various sites including the retrosplenial cortex (RSPL), middle and superior temporal gyrus (MTG/STG), anterior temporal lobe (ATL), and temporoparietal junction (TPJ).  These sites are bound per the diagram to the ventromedial prefrontal cortex (vmPFC).    Solid lines are context sensitive associative pathways biased by the vmPFC. Broken lines in the diagram represent context irrelevant associations that are not activated or inhibited.

How might all of this model work for psychiatry?  In general physicians are seeing a lot of patients in their training and practice.  In the course of that work - schemas are developed for diagnoses, signs, symptoms, and situations.  Here is a comparison of two scenarios that all psychiatrists are trained to recognize acute encephalitis and bipolar disorder, manic with psychotic features. 

	Encephalitis	Bipolar disorder, manic with psychosis
Schema	Acute illness, acute altered mental status, fever, seizures, focal neurological deficits, CSF/MRI abnormalities	Acute illness, euphoria/irritability/anger, hyperactivity, functional impairment, psychosis, temporal pattern, exclusion features
Subschema	Predisposing factors, pathophysiology patterns, temporal pattern	Euphoric expansive Irritable dysphoric Spontaneous v. precipitated
Gist	Acute confusion + fever + temporal lobe MRI changes = treat as HSV until proven otherwise" "Young woman + new psychosis + movement disorder = think anti-NMDAR, look for teratoma" "Summer encephalitis + flaccid paralysis = arboviral, likely West Nile" “Immunocompromised man with acute agitation = think CMV encephalitis	Episodic psychosis +/- mood changes (diagnosis gist) Mood stabilizer + antipsychotic (treatment gist) Severe postpartum psychosis = think bipolar disorder, manic with psychotic features Catatonia – think bipolar disorder, manic/depressed/mixed with psychotic features.

 

I came up with the following graphic (click to enlarge) based on the descriptive categories and the cognitive neuroscience model of Gilboa and Marlatte (12).  From left to right – the  “heterogenous construct supported by clinical utility” characterization is probably the most charitable one from philosophers.  Others like “this disorder does not exist” or “this disorder is not real” are two additional examples.  The central semantic memory category includes investigations and models of diagnostic reasoning conducted largely on medical students and physicians.  The cognitive neuroscience model contains schema and I have attempted to show how the concepts and actions map from the semantic memory to the schema model.  In both the semantic memory and cognitive neuroscience model, although the focus is memory the conceptualizations are really knowledge structures emphasizing a dynamic role for the schema in incorporating features of reality – in this case patient encounters. The cognitive neuroscience and semantic memory models also map on to brain anatomy – with a more comprehensive map for the cognitive neuroscience model as illustrated in the figure at the top.

What have I learned about this so far:

1:  The pattern matching of yesterday is more complicated today – I taught a course in diagnosis and diagnostic reasoning for 15 years into the early part of this century.  Pattern matching and pattern completion was a big part of that course.  The patterns were fairly simple and involved visual diagnoses (diabetic retinopathy, rashes) comparing physicians at various levels of training.  The most dynamic aspect was the implication that experts were better at matching incomplete patterns than novices.  Today’s conceptualizations of knowledge structures and schemas contain concepts, actions, and dynamically alter what is retained in memory and what is not. 

2:  There are clear implications for psychiatric diagnosis -   the DSM classification and all of the criteria do not capture the reality of medical and psychiatric diagnoses.  There is a qualifier in the manual that it is not a substitute for experience but that is never defined.  That reason becomes a lot clearer looking these cognitive models.  Classification systems attempt to operationalize the diagnostic reasoning of a physician by averaging a verbal description of those events.  I don’t think that is possible and I will cite a couple of examples.

Example 1:  A psychotherapist refers a 27-year-old woman to a psychiatrist because of concerns that she has histrionic personality disorder.  She has not been able to make progress in therapy.  The psychiatrist seeing the patient knows within minutes that she is manic.

Example 2:  An intern is presenting the history of a 68-year-old man to his psychiatric attending.  The patient is extremely depressed to the point that he believes that he is cursed based on a trivial event that occurred in his childhood. Within the first 5 minutes the attending realizes that the patient is delusional and communicates that to the intern. The intern acknowledges that this is true and wonders how he failed to make that diagnosis.

Both cases highlight that knowledge of a classification system is not enough.  The psychotherapist and the intern both know the DSM and use it regularly. They have both had didactics in classification of mental disorders.  The only difference is that the psychiatrist in both cases has experienced cases of the disorder and had knowledge structures and schema to make the diagnosis.  Written descriptions of schema and knowledge structures are an incomplete approach to diagnostic reasoning. 

3: Classifications artificially separate actions from concepts – any reading of the DSM gives the impression that “this is the universe of psychiatric disorders – in order to function as a psychiatrist, pick one and then come up with a treatment plan.”  This is problematic at two levels.  First, if the cognitive neuroscience model of memories and knowledge structures is correct – a classification system is operating at a sublevel that averages features.  It is blind to the overall gist that despite this averaging no two people are alike.  Second, it removes action features that are necessary to function as a physician.  That would include top level schemas like “This patient is medically unstable and requires medical or surgical care first” or “This is a life-threatening problem that requires a safe and closely monitored environment." Some will argue that is not the goal of classification.  I would argue that many consider classification to be a diagnosis and in order for it to function that way – it needs to include action items in addition to a general rule out of causative intoxication states and medical problems. The DSM as it exists is classification without diagnosis.

4:  Cognitive neuroscience models highlight the fact that the separation between diagnosis and treatment is artificial.  All physicians are taught to do exhaustive evaluations of medical problems.  That is the initial step in a career.  It is also critical to learn when that exhaustive process needs to be immediately interrupted to focus on a more acute problem. I can still recall seeing a 7-year-old boy who have been hit by a car while playing in the street. He was alert but had significant abdominal pain.  The car bumper struck him just below his left rib cage.  It took me less than 5 minutes to determine that he had an acute abdomen and call the trauma surgeons. That non-linear process happens frequently in acute care psychiatry and in outpatient psychiatry with patients in crisis who need verbal interventions to assist in the diagnostic and treatment process.  

5:  Psychotherapy – there are recent perspectives on how cognitive psychology applies to the psychotherapeutic process at both the psychological and biological levels using these models.  Basically, maladaptive schemas are confronted and modified during the therapy.  There is some empirical evidence that this may happen particularly in the area of positive and negative self-schemas.  Much of this literature draws on existing cognitive behavioral therapy.  That leads to a question of what is the difference between a therapy focused on a cognition or an isolated memory compared with a schema focused therapy?

At the highest level of analysis memory focused therapies generally involve isolated autobiographical memories and schema focused therapies are about knowledge structures abstracted across multiple events that involve emotion, cognition, and behavior.  In theory the schema focused therapies may be useful in cases where the memory focused therapy is not effective, but a competing consideration is that schemas can be entrenched and difficult to change.  The memory focused therapy could be considered a bottom-up type of approach and the schema focused a top-down approach. 

6:  Criticisms – Criticizing the DSM as a diagnostic system is a cottage industry in the US and the UK.  As we approach a new version of the DSM expect most media sites to start months and even years of criticism. Practically everybody does it rarely discussing their motivations, understanding, and the limitations of their proposed system if they have one.  If diagnostic reasoning is a complex process consistent with the cognitive neuroscience models and requires direct experience, criticism of the manual rings hollow.  It is equivalent to reading about things that might exist and proclaiming you are an expert.  Psychiatrists with criticisms are also limited if they have insufficient experience in the areas they are criticizing.  Psychiatrists with the broadest experience will produce the best criticism. If you are criticizing a list of diagnostic criteria in a classification system in isolation – that is exactly what you are doing.  It is trivial compared with an actual diagnosis by a trained and experienced psychiatrist.        

This brief focus on the cognitive neuroscience of diagnosis should highlight that psychiatric education and practice is seriously lagging in this knowledge base.  If we are taking the “diagnosis” in DSM seriously it has to be modified to include this important brain science.  All of the current competing models face the same criticism.  A diagnosis by a physician is much more than typed criteria attempting to capture a dynamic process.  Secondly, psychiatry needs modern approaches to the mind. Approaches that correlate with neurobiology and have a clear empirical basis. Much of the DSM claims a sketchy atheoretical basis that should no longer be acceptable when powerful explanatory theories may exist.  Philosophy is no substitute.  Finally, we must find a way to implement these across all of our training programs and practitioners.  We should be devoting as many resources to integrating cognitive neuroscience into psychiatry as we do modifying the DSM.

And that should be the first step.  What does a DSM looked like with cognitive neuroscience baked in?  The answer goes a lot farther than “dimensions”.      

   

George Dawson, MD, DFAPA

 

 

References:

 

1:  Norman G, Young M, Brooks L. Non-analytical models of clinical reasoning: the role of experience. Med Educ. 2007 Dec;41(12):1140-5. doi: 10.1111/j.1365-2923.2007.02914.x. Epub 2007 Nov 13. PMID: 18004990.

2:  Brush JE Jr, Sherbino J, Norman GR. Diagnostic reasoning in cardiovascular medicine. BMJ. 2022 Jan 5;376:e064389. doi: 10.1136/bmj-2021-064389. PMID: 34987062.

3:  Custers EJ. Thirty years of illness scripts: Theoretical origins and practical applications. Med Teach. 2015 May;37(5):457-62. doi: 10.3109/0142159X.2014.956052. Epub 2014 Sep 2. PMID: 25180878.

4:  Koufidis C, Manninen K, Nieminen J, Wohlin M, Silén C. Unravelling the polyphony in clinical reasoning research in medical education. J Eval Clin Pract. 2021 Apr;27(2):438-450. doi: 10.1111/jep.13432. Epub 2020 Jun 22. PMID: 32573080.

 5:  Binder JR, Desai RH, Graves WW, Conant LL. Where is the semantic system? A critical review and meta-analysis of 120 functional neuroimaging studies. Cereb Cortex. 2009 Dec;19(12):2767-96. doi: 10.1093/cercor/bhp055. Epub 2009 Mar 27. PMID: 19329570; PMCID: PMC2774390.

6:  Duff MC, Covington NV, Hilverman C, Cohen NJ. Semantic Memory and the Hippocampus: Revisiting, Reaffirming, and Extending the Reach of Their Critical Relationship. Front Hum Neurosci. 2020 Jan 24;13:471. doi: 10.3389/fnhum.2019.00471. PMID: 32038203; PMCID: PMC6993580.

7:  Insaustu R, Amaral DG. Hippocampal Formation. In: Mai JK, Paxinos G (eds) The Human Nervous System, 3^rd ed.  Elsevier, London, 2012: p. 933.

8:  Mazoué A, Gaultier A, Rocher L, Deruet AL, Vercelletto M, Boutoleau-Bretonnière C. Does a rabbit have feathers or fur? Development of a 42-item semantic memory test (SMT-42). J Clin Exp Neuropsychol. 2022 Sep;44(7):514-531. doi: 10.1080/13803395.2022.2133088. PMID: 36269845.

9:  Renoult L, Irish M, Moscovitch M, Rugg MD. From Knowing to Remembering: The Semantic-Episodic Distinction. Trends Cogn Sci. 2019 Dec;23(12):1041-1057. doi: 10.1016/j.tics.2019.09.008. Epub 2019 Oct 28. PMID: 31672430.

10:  Brown TI, Rissman J, Chow TE, Uncapher MR, Wagner AD. Differential Medial Temporal Lobe and Parietal Cortical Contributions to Real-world Autobiographical Episodic and Autobiographical Semantic Memory. Sci Rep. 2018 Apr 18;8(1):6190. doi: 10.1038/s41598-018-24549-y. PMID: 29670138; PMCID: PMC5906442.

11:  Teghil A, Bonavita A, Procida F, Giove F, Boccia M. Temporal Organization of Episodic and Experience-near Semantic Autobiographical Memories: Neural Correlates and Context-dependent Connectivity. J Cogn Neurosci. 2022 Nov 1;34(12):2256-2274. doi: 10.1162/jocn_a_01906. PMID: 36007071.

12:  Gilboa A, Marlatte H. Neurobiology of Schemas and Schema-Mediated Memory. Trends Cogn Sci. 2017 Aug;21(8):618-631. doi: 10.1016/j.tics.2017.04.013. Epub 2017 May 24. PMID: 28551107.

13:  Reyna VF, Edelson S, Hayes B, Garavito D. Supporting Health and Medical Decision Making: Findings and Insights from Fuzzy-Trace Theory. Med Decis Making. 2022 Aug;42(6):741-754. doi: 10.1177/0272989X221105473. Epub 2022 Jun 23. PMID: 35735225; PMCID: PMC9283268.

14:  Wilhelms EA, Fraenkel L, Reyna VF. Effects of Probabilities, Adverse Outcomes, and Status Quo on Perceived Riskiness of Medications: Testing Explanatory Hypotheses Concerning Gist, Worry, and Numeracy. Appl Cogn Psychol. 2018 Nov-Dec;32(6):714-726. doi: 10.1002/acp.3448. Epub 2018 Sep 1. PMID: 30686857; PMCID: PMC6345391.

15:  Hawke LD, Provencher MD, Parikh SV. Schema therapy for bipolar disorder: a conceptual model and future directions. J Affect Disord. 2013 May 15;148(1):118-22. doi: 10.1016/j.jad.2012.10.034. Epub 2012 Dec 4. PMID: 23218898.

16:  Lane RD, Ryan L, Nadel L, Greenberg L. Memory reconsolidation, emotional arousal, and the process of change in psychotherapy: New insights from brain science. Behav Brain Sci. 2015;38:e1. doi: 10.1017/S0140525X14000041. Epub 2014 May 15. PMID: 24827452.

Graphics Credit:

1:  The lead graphic as noted is from Cell Press and reference #12.  It is reproduced here with permission from Elsevier and this is their acknowledgement:

Reprinted from Trends in Cognitive Sciences, August 21(8), Gilboa A, Marlatte H. Neurobiology of Schemas and Schema-Mediated Memory, p. 618., Copyright 2017, with permission from Elsevier.  License 6278000229455, May 29, 2026 

2:  Second graphic was made by me using Microsoft Visio.

Supplementary 1:  Nobel Laureate and Psychiatrist Eric Kandel noted the importance of cognitive neuroscience years ago and this was a quote from his book:  The Age of Insight.

 

Medical Reasoning vs. A Diagnostic Manual

 

I taught a course on medical decision making and how not to mistake a physical illness for a psychiatric disorder from about 1990 to 2002. The main theorists at the time were all internists – Stephen Pauker, Jerome Kassirer, Richard Kopelman, David Eddy, and Harold Sox.  I read their papers and attended their courses.  State-of-the-art in those days involved extensive differential diagnosis, Bayesian analysis, and an awareness of an extensive list of potential cognitive biases. I had been impressed with the need for pattern matching and pattern completion and incorporated all those elements into my course.  I eventually pared it down to about 9 sections in the lecture notes illustrated with case vignettes.

My original emphasis was to recognize that there are several considerations when assessing the medical aspects of psychiatric care.  The first is the medical stability of the patient.  Can they be cared for on a psychiatric unit or do their medical needs require medicine or in some cases surgery?  Do they need referral to a generalist of specialist?  This is more complicated than it sounds because the patient is there seeing a psychiatrist for what is supposed to be a psychiatric problem.  But that presentation is complicated by several factors including most patients have no primary care physician and no routine health care maintenance. Many will come into the emergency department concerned about a medical problem but get sent to psychiatry. In that situation, people still get all of the acute medical illnesses including heart attacks, strokes, asthma attacks, pulmonary emboli, seizures, pneumonia, meningitis, encephalitis, and acute cholecystitis to name a few.  Many exhibit non-specific behaviors like agitation, crying out, aggression, or unresponsiveness that can be due to either a psychiatric disorder or a medical problem.    

The second is a psychiatric presentation of a physical illness in a communicating patient. The classic presentations involve brain pathology that is infection, inflammatory, vascular, trauma, or neurodegenerative.  Systemic endocrinopathies and inflammatory disorders are a close second. 

Finally, there is the patient with a clear psychiatric disorder who has intercurrent illness that is or is not known.  Examples that I have seen many times include current or new onset diabetes mellitus, profound anemia usually secondary to an upper or lower GI bleed, dermatology conditions that have often been neglected, symptomatic nutritional deficiencies (B12, folate, D), sexually transmitted diseases, complications of substance use like cirrhosis, and various acute and chronic infectious diseases.

Given that large population with diverse medical and psychiatric problems as well as diverse presentations that can include denying any physical problems – I typically reviewed how the diagnoses occurred.  Pattern matching was the fastest.  The physician has seen a physical finding, lab, behavior, etc – many times before, knows what it is, diagnoses it and treats it.  A good example is a rash.  Dermatologists are rash experts and can correctly classify rashes and marginal cases much faster than primary care physicians (4).  The same is true for diabetic retinopathy and ophthalmologists (5).  Until you have seen a person with severe mania or catatonia, neuroleptic malignant syndrome, or serotonin syndrome it is less likely that you can diagnose the conditions by reading criteria in a book.  Patterns are important for all medical specialists.

On the other end of the spectrum is the contemplative side of diagnosis.  There are several possible diagnoses, and it takes additional data, thought, and reasoning to come to a final diagnosis. Every medical student does this in their initial internal medicine rotation.  There is encouragement to produce a list of many diagnoses that might account for the presentation – but even as the case is being recorded or presented that list rapidly narrows to the apparent diagnosis.

In psychiatry, it may take much more data and collateral information to make a specific diagnosis at the initial presentation.  First episode psychosis (FEP) is a case in point. It is very important to determine what the symptoms onset was like and whether there were any associated mood symptoms or substance use problems. The patient may not be able to describe the phenomenology and depending on the circumstances treatment may be initiated while the diagnostic process is ongoing.  Teaching about the diagnostic process, we would spend time discussing what that might look like combined with a recursive approach to the patient and an awareness of cognitive and emotional biases.  I provided several examples of non-psychiatric physicians making errors due to emotional biases.

Since my course, the literature on medical decision making has changed to some degree.  There is some literature that addresses expertise in general at both the level of cognitive psychology (1) and neurobiology (2).  The general approaches have been to analyze expertise and diagnostic reasoning from the perspective of typical domains (cognitive, perceptual, motor) or to look at a general model and how that has developed over the years.

A dual processing model (3) is generally considered the best current representation of clinical reasoning and decision making.  In this model, there is a fast automatic, heuristic, and unconscious system called Type 1 and a slower conscious, analytical, and effortful system called Type 2.  Additional properties are indicated in the following table.

Parameter	Type 1	Type 2
Speed	Fast, automatic, unconscious/preconscious, little effort	Slow, deliberate, analytical, varying degrees of effort
Control	Minimum control, similar to automatic associations in everyday life except more focused	Control over thought process and direction
Systems and Processing	Pattern recognition and completion, implicit learning, access to long term memory	Working memory and manipulation of data in working memory, planning and reasoning based on that data
Memory Systems	Long term memory	Short term and working memory
Localization	-Orbitofrontal cortex (OFC) -Basal ganglia (caudate, putamen) -Insula -Anterior cingulate cortex -Amygdala -Hippocampus	-Dorsolateral prefrontal cortex (DLPFC) -Left inferior frontal gyrus -Middle frontal gyrus -Inferior parietal lobule -Precuneus -Hippocampus

 A clinical example of Type 1 reasoning is when a trained clinician recognizes a classic presentation of a medical illness, diagnosis, or finding.  An example I frequently use is when one of my Infectious Disease attendings who was an expert in Streptococcal infections recognized characteristic rash from across the room on a patient we were consulted for a different problem.  He made the diagnosis within seconds and told us how it could be confirmed.  In studies of the process the orbitofrontal cortex and limbic connections are activated.  Training is a critical element, especially seeing a maximum number of patterns and their variations.  Although the characterization is that this is a fast and automatic process, there is some room for deliberation.  For example, recognizing or attempting to classify equivocal cases without classic presentations. 

Type 2 reasoning is considered more of the typical process of differential diagnosis.  The findings are compared, analyzed, and accepted or rejected based on additional data and clinical judgment. This process is thought to localize in dorsolateral prefrontal cortex (DLPFC) the home of the working memory where data can be maintained and analyzed.  The left inferior frontal gyrus contributes to rule-based reasoning and hypothesis testing.  A clinical example from my experience is the case of the agitated stuporous patient.  These cases require a great deal of caution because they are most likely to represent a serious or life-threatening illness.  It requires a clinician who knows how to examine patients with stupor or coma and rapidly makes sense of the history and findings. It is a problem that can rarely be solved by Type 1 reasoning alone due to a fairly non-specific presentation.  Some of the critical points for hypothesis testing will be signs of increased intracranial pressure, purposeful response to painful stimuli, eye movements, reflex and musculoskeletal exam abnormalities, signs of infection, and meningeal signs.

The interaction between Type 1 and Type 2 systems is not necessarily sequential but it can be with the Type 1 system matching patterns that lead to hypothesis generation.  There is some evidence that in most clinical situations most of the diagnoses occur with Type 1 reasoning.  Experts can operate at the level of Type 1 reasoning due to extensive experience.  There is not necessarily a hard separation based on the properties in the table. Some hypothesis testing can occur at both levels.  Both systems are commonly grounded in both the limbic system and the hippocampus.

The human brain is capable of parallel distributed processing of data or information.  This means that there are many processing areas in the brain that are interconnected and they can all be working at once.  The modern conceptualization is brain networks that are active processing areas connected by white matter tracts widely distributed through the brain.  

That brings me to my model of diagnostic reasoning (see lead graphic and click to enlarge).  It is based on the course I taught, neuroanatomy and neurology, and what I have observed clinically. When I was talking about pattern matching 20 years ago based on my observations and reading studies in dermatology, ophthalmology, radiology, and pathology – the term seemed to fade rapidly from the diagnostic reasoning literature.  It was revived somewhat by the more recent focus on AI and comparison of that modality to humans.

There was a lull in Bayesian analysis after the invention of computerized programs like Quick Medical Reference (QMR) and Iliad.  They were designed to facilitate medical diagnoses by providing an exhaustive list of findings and their probabilities. These were 20^th century personal computer programs and not AI.  A study of these and 2 additional programs suggests that the programs got 52-71% of 105 diagnostic cases correct with 19-37% being the mean portion of correct diagnoses (6). Despite those figures the programs provided an additional 2 diagnoses per case that experts considered as relevant.  The authors recommended that the programs be used only by physicians who could include the relevant and exclude the irrelevant information provided by the programs.  The programs were discontinued without further modification or updates.  

That is the 8-mile-high view.  I plan to do a deeper dive into the neuroanatomy and neurophysiology.  But the clear reality of the situation is the ability to make a psychiatric diagnosis resides in the brain of a psychiatrist and not a classification manual or a checklist.   Manuals and checklists are crude approximations of some of the cognitive features that psychiatric experts possess.  Like all experts – skill will vary based on practice, exposure, and interest because of the effects on these brain systems.  But we are well past the point of equating what a psychiatrist does to a crude manual.  A manual never saved or treated anyone.  Further – the diagnostic reasoning process emphasizes elements that are important for education and training. It seems that in the past decades there has been a preoccupation with evidence-based research rather than the evidence itself. It does not do the physician or patient any good to be in a situation where that physician is unable to communicate with a person who is in a critical state and has no idea how to assess that problem.  Rearranging diagnostic criteria in a manual for the ninth or tenth time does not get you there.   

 

George Dawson, MD, DFAPA

Supplementary 1:   Before anyone says the diagram is too complex - it is a general diagram for any human diagnostician.  The main modifications for physicians and psychiatrists are the interactive aspects that include empathic comments, formulations, and numerous verbal interventions that other diagnosticians may not need to use.  The specifics about how these memory systems interact are not known at this point - I will be researching that over the next several months.  I borrowed the superposition concept from quantum mechanics - even though there are no wave functions for memory.         

 References:

.

1:  Bilalić M.  The Neuroscience of Expertise.  Cambridge University Press. Cambridge, United Kingdom. 2017.

2:  Maguire EA, Gadian DG, Johnsrude IS, Good CD, Ashburner J,  Frackowiak RSJ, Frith CD. 2000. Navigation-related structural change in the hippocampi of taxi drivers. Proc Natl Acad Sci USA 97:4398–4403.

3:  Norman GR, Monteiro SD, Sherbino J, Ilgen JS, Schmidt HG, Mamede S. The Causes of Errors in Clinical Reasoning: Cognitive Biases, Knowledge Deficits, and Dual Process Thinking. Acad Med. 2017 Jan;92(1):23-30. doi: 10.1097/ACM.0000000000001421. PMID: 27782919.

4:  Federman DG, Concato J, Kirsner RS. Comparison of dermatologic diagnoses by primary care practitioners and dermatologists. A review of the literature. Arch Fam Med. 1999 Mar-Apr;8(2):170-2. doi: 10.1001/archfami.8.2.170. PMID: 10101989

5:  Sussman EJ, Tsiaras WG, Soper KA. Diagnosis of Diabetic Eye Disease. JAMA. 1982;247(23):3231–3234. doi:10.1001/jama.1982.03320480047025

6:  Berner ES, Webster GD, Shugerman AA, Jackson JR, Algina J, Baker AL, Ball EV, Cobbs CG, Dennis VW, Frenkel EP, et al. Performance of four computer-based diagnostic systems. N Engl J Med. 1994 Jun 23;330(25):1792-6. doi: 10.1056/NEJM199406233302506. PMID: 8190157.

Adding Rather than Subtracting Bias - An Underlying Basis for Polypharmacy?

There was an interesting piece in Nature this week (1,2) about cognitive biases in complex problem solving.  The research psychologists asked subjects to solve problems of varying complexity and structure from the perspective of whether additional structures or steps were necessary or whether an optimal solution could be obtained by subtracting structures or steps. I will briefly describe each of the problems in the table below (pending permission to use one of their graphics).

Task	Description
Abstract grid task	Transform a grid pattern to make it symmetrical
Suggested changes to a large public university	Changes to improve the sense of community, enable student learning, and prepare students for a lifetime of service
Lego block structure	Improve the 8 or 10 block structure
Lego block structures	3 possible structures built from 12 blocks of a pool of 24 blocks on a 6” x 8” base.
Lego block structure	Revision of original structures made from a possible 20 blocks to make a 10 block structure
Lego block structure	Modify a Lego structure so that it can hold a brick over the head of an action figure in the structure
Read and summarize an article	Make a 6-8 sentence summary and then edit it to a shorter version
Read and summarize an article	Edit someone else’s summary and edit it to “omit needless words”
Day trip to Washington DC	Inspect a trip itinerary and suggest changes to improve it
Make a grilled cheese sandwich	Make a grilled cheese from 27 ingredients -
Modify a soup recipe	From 5, 10, and 15 ingredient soup recipes – modify from a list of ingredients and modify to improves the soup.

Inspection shows that the cognitive tasks cover many domains ranging from 2D and 3D visuospatial tasks, language tasks, and more theoretical tasks that involve speculative rather than confirmed outcomes. The authors suggest an all-encompassing definition: “the cognitive science of problem solving describes iterative processes to imagining and evaluating actions and outcomes to determine if they would produce an improved state.”(p. 258).  They define subtractive transformations as fewer components than the original and additive transformations as more components than the original.  The authors noted a bias in anecdotal literature to making conscious subtractive transformations and that suggested to them that strategy may be less common or undervalued. 

Across all experiments, the tendency toward subtractive strategies with the general instruction were lower but probabilistic.  For example, across all experiments, subtractions ranged from 21-41%.  A second set of conditions with subtle subtraction cues increased the rate of subtractive transformations to 43-61% across the same experiments.  At one point the researchers added a cognitive load task that was basically a distractor to use more attentional resources. In these conditions cognitive shortcuts are less accessible. Under those conditions subjects failed to identify a subtractive solution more frequently.  The authors also studied subjects form Germany and Japan suggesting that there is cultural generalizability of the additive over subtractive strategies.

The authors consider that the differences could be accounted for by generating a number of additive and subtractive ideas and selecting the additive or they simply default to the additive.  They elected to look at the default to the additive mode. They describe heuristic memory searches allowing for the timely access of relevant information.  They suggest a number of reasons what additive strategies may be favored including – the processing may be easier, semantic biases such as more being better, cognitive biases may favor the status quo or less change, and it may be more probable that additive rather than subtractive changes offer a better outcome.

This is an interesting paper from a number of perspectives.  First, it presents a cognitive psychology approach with no purported biological mechanisms. There are no functional imaging studies or brain systems described.  The theories and design of experiments depends on a psychological model of cognitive function. Second, the model is probabilistic.  Although the title suggests systematic overlooking of subtractive strategies, it turns out that many don’t and this bias can be modified by experimental conditions such as subtraction cues. Third, the effect of increased cognitive load can be demonstrated to increase the likelihood of additive rather than subtractive biases. Fourth, the biases extend across a number of domains including physical, social, and intellectual. Fifth, the authors suggest that there may be a number of “cognitive, cultural, and socioecological reasons for favoring the additive bias over the subtractive one.  Sixth, although the additive transformation was more likely to occur that does not mean it offers the best solution to the problem.  It may simply be the most commonly used solution. 

Real world experience illustrates how the additive transformations can be reinforced.  Advertising is a common one. The goal of advertising is basically to sell someone something that they don’t need or change their preferences for something that they do need to a different product.  If it works, it is an additive strategy on top of additive behavior.  If the product being sold affects other learning centers in the brain like reward-based learning that can lead to further additive effects. The photo at the top of this post illustrates another example.  This kitchen drawer for spoons and spatulas is a solution to the cooking problem of how many are needed to accomplish what the cook in this case needs to accomplish. The drawer is packed to the point where it barely closes and at that point, the cook is forced to reassess and decide about cleaning the drawer out and starting over.  Homeowners often forced to make similar downsizing or subtractive decisions after 20-30 years of additive ones and being forced with either space constraints or a smaller family.  

What about medical and psychiatric treatment?  I don’t think there is any doubt that additive transformations are operating. Most treatments that involve medication have a step approach with the addition of medications for symptoms that do not respond or partially respond to the initial treatment. This occurs after an explicit subtractive bias or at least a bias to maintain the status quo 20 years ago.  At that time, hospitals and clinics were reviewed based on criteria to limit the amount of polypharmacy defined as more than one drug from the same class. Today, polypharmacy is common.  Reference 3 below gives an example of polypharmacy defined as 5 or more medications taken concomitantly and hyper-polypharmacy was defined as 10 or more medications taken concomitantly in a 3-month sample of 404 geriatric patients with cardiovascular disease admitted to a hospital during 3-month period.  They found the prevalence of polypharmacy was 95%.  The prevalence of hyper-polypharmacy was 60%.  Most patients (77.5%) also had a potential drug-drug interaction.  Their suggestion be vigilant is a strategy discussed as being potentially successful in containing the additive strategies (2).  

From psychiatry, I am including a common problem that I encountered as a tertiary consultant.  That problem is what to do about a person with a depression that has not responded to high dose venlafaxine. There are geographic areas in the US, where very high dose venlafaxine is used with and without pharmacogenomic testing.  From the options listed in the diagram it is apparent that there are 4 additive (black arrows) strategies and 2 subtractive (red arrows). There is a robust literature on the additive strategies and not so much with the subtractive. As a result, it is common these days to encounter patients who have tried numerous combinations right up to and including “California Rocket Fuel” (4) of the combination of an SNRI like venlafaxine with mirtazapine.  The ways to analyze this situation, especially if there has not been any improvement are significant and depend a lot on patient preferences and side effects in addition to the lack of response. I have found that very high dose venlafaxine, can be sedating to a significant number of people and that they feel better when it is tapered.  I have also seen many people far along the augmentation strategies when tapering or discontinuing the venlafaxine was never considered. In some of these cases, the patient reports that venlafaxine is historically the only antidepressant that has worked for them in the past.

That brings up the issue of additive versus subtractive biases on the part of the patient. We have all been bombarded by pharmaceutical commercials suggesting the best way to mood stabilization is adding another medication – typically aripiprazole or brexpiprazole. In fact, those commercials speak directly to additive biases. It is often very difficult to convince a person to discontinue or reduce a medication that they have talked for years – even when careful review suggests it has been ineffective or creates significant side effects. 

Could a discussion of additive versus subtractive transformations be useful in those situations? There is currently no empirical guidance, but these might be additional experiments to consider for both prescribing physicians and the patients they are seeing. Certainly the expectations that they patient has for any given treatment needs to be discussed and whether that expectation is reasonable given their personal experience and the objective evidence. On the side of prescribing physicians, it is fairly easy to flag medication combinations that are problematic either from the perspectives of too many medications being used at once, physical and side effects not being analyzed closely enough, or medications being changed too frequently. Would discussing additive and subtractive strategies be useful in that setting?  Would a discussion of basic rules to address additive biases such as discontinuing a medication when it is replaced be useful?

Remaining vigilant that there are subtractive strategies out there is a useful lesson from this paper. Physicians are aware of the concept of parsimony and how that can be applied to medical care. Given the fact that the additive strategies are probabilistic and modifiable with conscious strategies – that should still prove to useful in containing polypharmacy.  

 

George Dawson, MD, DFAPA

Supplementary:

Another common additive strategy that I have encountered in the past 10 years is performance enhancement.  The patient presents not so much for treatment of a psychiatric problem but because they believe that adding a medication or two or three will improve their overall ability to function. Common examples would include:

1.  Presenting for treatment of ADHD (with a stimulant medication) not because of an attentional problem but because the stimulant creates increased energy and the feeling of enhanced productivity.

2.  Presenting for treatment of insomnia in the context of drinking excessive amounts of caffeine in the daytime and the caffeine is viewed as necessary to enhance energy at work or in the gym.  In some cases, stimulants are taken in the daytime and the idea is that the medication for insomnia would counter the effect of stimulants or caffeine taken late into the day.

3.  Taking anabolic androgenic steroids (AAS) and expecting to treat the side effects of mood disturbances, insomnia, anger, and irritability in order to keep taking the AAS.  Many AAS users also take other medications for this purpose as well as various vitamins, supplements, and stimulants to enhance work outs.

4.  Taking excessive numbers of supplements with no proven value and seeking to use medications for nondescript symptoms associated with the supplement use. In many cases, patients with psychiatric disorders are sold on elaborate mixtures of minerals and supplements with the promise that they address their symptoms.  In many cases it is difficult to determine if the associated vitamins and supplements interact with the indicated medical treatment or not.

All of these are additive strategies with no proven value that I have seen in the outpatient settings.  It is obviously important to know if the patient being treated is using these strategies.  There are often competing considerations – for example does the patient have a substance use disorder and are substance use disorders another predisposing condition to additive biases (I suspect they strongly are).

 

References:

1:  Meyvis T, Yoon H. Adding is favoured over subtracting in problem solving. Nature. 2021 Apr;592(7853):189-190. doi: 10.1038/d41586-021-00592-0. PMID: 33828311.

2:  Adams GS, Converse BA, Hales AH, Klotz LE. People systematically overlook subtractive changes. Nature. 2021 Apr;592(7853):258-261. doi: 10.1038/s41586-021-03380-y. Epub 2021 Apr 7. PMID: 33828317.

3:  Sheikh-Taha M, Asmar M. Polypharmacy and severe potential drug-drug interactions among older adults with cardiovascular disease in the United States. BMC Geriatr. 2021 Apr 7;21(1):233. doi: 10.1186/s12877-021-02183-0. PMID: 33827442; PMCID: PMC8028718.

4:  Stahl, SM . Essential psychopharmacology: neuroscientific basis and practical applications. Cambridge University Press, Cambridge 2000. p. 363.

 

Graphics Credit:

So far they are all mine.  Yes that is one of my kitchen drawers but I am fairly good at avoiding polypharmacy.  Click on any graphic to enlarge.