The Highest Accurate Abstraction
Compressed Medicine · 3. The Highest Accurate Abstraction
Tuesday morning, ED triage room three. A 22-year-old with type 1 diabetes is brought in by his roommate, slurring his words and breathing fast. The intern picks up the case. Two minutes later the intern has a working impression. Here are two ways to write it.
Version one: "22yo male, type 1 DM. Confusion, dyspnea, vomiting times eight hours. Vitals: T 99.1, HR 124, RR 32, BP 102/68. Exam: dry mucous membranes, Kussmaul respirations, no focal neuro deficits. Labs: glucose 480, anion gap 24, bicarb 8, pH 7.18, ketones positive, K 5.2, BUN 28, Cr 1.4."
Version two: "22yo type 1 DM with DKA. Anion gap 24, pH 7.18. Started on fluids and insulin drip per protocol."
Same patient, same minute, two messages. The second activates the entire DKA model in the receiver's head: expected trajectory, standard treatment cascade, the potassium watch that becomes important once insulin starts, the gap closure timeline, the cerebral edema risk in children, the eventual transition criteria. The receiver does not have to assemble the picture from raw data.
The first message contains more facts. The second does more clinical work.
The principle
Compress to the highest level the receiver can accurately decompress. Within the diagnostic dimension of a clinical message, disease names compress more than syndrome names, syndrome names compress more than complaint patterns, complaint patterns compress more than raw observations. The hierarchy measures how much model the receiver gets for free.
The principle operates on every axis of clinical compression: state (what is happening), trajectory (how it is changing), action (what is being done), and confidence (how reliable each of these is). The state axis carries the most cognitive load per token, so this essay walks state in detail. The same logic transfers to the other three, and we return to them after the worked example.
Why the hierarchy exists
A diagnosis is a model loader. More generally, every level of the hierarchy is a handle: a token pointing at a pattern in clinical reality. The receiver's stored model of that pattern is what loads when the handle arrives. Disease names point at rich patterns: presentation, trajectory, complications, treatment cascade, exit criteria, predictable side effects of the treatment itself. Syndrome names point at management pathways without etiologic commitment. Complaint patterns point at differential shapes. Raw observations point at single values at single moments.
Compression ratio of a handle = (information in the pattern it points at) / (information in the handle itself)
The numerator scales with the richness of the pattern: how many predictions about trajectory, treatment, complications, and disposition the receiver loads when the handle arrives. The denominator scales with the handle's surface cost. "DKA" is a three-letter handle pointing at a pattern that unpacks into dozens of predictions and a multi-hour management trajectory; "anion gap 24" is a longer handle pointing at a single value. This is a clinical instance of an information-theoretic principle natural language already exploits: word lengths are optimized so that less-predictable, higher-information content gets longer encodings while predictable content rides in shorter forms (Piantadosi et al. 2011, following Zipf and Shannon). Across cognitive domains, certain category levels are also privileged because they maximize within-category similarity while preserving between-category distinctiveness (Rosch 1978). Clinical reasoning inherits both: the disease level is usually where the patterns are richest and the handles are shortest.
This is how trained clinicians actually reason: by activating illness scripts (the clinical instantiation of scripts as compressed knowledge structures in cognitive science) and matching the case against them (Schank & Abelson 1977; Schmidt & Rikers 2007). The handle the sender uses determines which script the receiver loads.
Accurate means predictive
"Accurate" does not mean complete or final. Every clinical abstraction is either a model of reality or a description of measured values, and accuracy is unified across both: the predictions the message carries verify when checked.
Higher-level handles carry rich models that predict trajectory, treatment response, and complications. The accuracy test for "DKA" is whether the downstream predictions hold: gap closes on insulin, potassium drifts as expected, the patient transitions on schedule. Lower-level handles carry narrow point-predictions about specific values. The accuracy test for "anion gap 24" is whether the value verifies if remeasured. Density of prediction differs across levels; the criterion does not. Predictive fidelity is the same standard at every level.
The codec constraint
A handle only loads a pattern in a receiver who has the matching codec: the stored handle-to-pattern mapping. "DKA" loads the full management cascade in an emergency physician; in a patient's family member it might load nothing, or load a wrong pattern built from a single past hospitalization. The hierarchy is receiver-relative, and more precisely codec-gap-relative. Communication theorists call this work "grounding": the joint effort sender and receiver do to establish enough shared representation that the message lands as intended (Clark & Brennan 1991). When codecs align tightly (intensivist to intensivist), high compression works: "DKA" lands and the rich pattern loads on arrival. When they diverge (specialist to generalist on a boundary case, clinician to layperson, clinician to AI trained on a different distribution), compression must descend to where alignment is still sufficient.
Function-relativity (Part 2: The Function of the Message) and level-relativity both operate through codec-gap. Function chooses which axes appear in the message and with what weight; level on each axis is then set by where sender and receiver codecs currently align. Structured handoff protocols like I-PASS (Starmer et al. 2014) operationalize this implicitly, prescribing which axes the message must cover. Teaching is the special case where the codec is being built rather than used, so the message lives at observation level deliberately.
The four levels of state
Disease / diagnosis when the model is earned. The highest-density handle. "DKA, gap closing, K borderline low." "NSTEMI, hemodynamically stable, awaiting cath." "Uncomplicated appendicitis on CT, surgery aware." The disease name carries the model; the additional words modulate it.
Syndrome / clinical state when the disease is not yet justified. High compression without commitment beyond what the evidence supports. "Undifferentiated shock." "Acute hypoxemic respiratory failure." "Sepsis physiology without source." These load management pathways without forcing an etiology that the data has not earned.
Complaint-pattern when even the syndrome is not defined. Chief complaint plus discriminating features. "Exertional pressure-like chest pain with left arm radiation." "Sudden maximal-onset headache with neck stiffness." These activate pattern-recognition for the right differential without claiming a specific entity.
Raw observations when the model has not been built yet. Numbers, findings, history details. They belong in the message when they are abnormal, changing, threshold-relevant, decision-relevant, surprising, or required to justify the level above. They do not belong by default.
Compress upward until prediction would break
That sentence is the operational rule. Pick the highest level you can defend; descend only when the level above fails.
The intern in triage three has lab values that nail the diagnosis. "DKA" is justified, and it is the right starting compression. The team that picks up the case loads the cascade and moves on.
A hypotensive patient with cold extremities and an unknown source has earned "undifferentiated shock" but not yet "septic shock" or "cardiogenic shock." Naming the syndrome keeps the receiver's mind open across the differential that drives the next workup. Compressing further to a specific disease would be guessing.
A patient with ambiguous ECG, pending troponin, unremarkable exam, and an atypical story may not yet have earned "ACS," "PE," or even "low-risk chest pain" as a stable management syndrome. "Exertional pressure-like chest pain with risk factors for ACS, workup pending" is the honest compression: it activates the right pattern without committing to a syndrome the case has not justified.
A patient in their first ten minutes of an undifferentiated workup lives at the observation level. The work of those ten minutes is to compress upward.
The failure modes
Premature compression. Using a level higher than the evidence supports. "Sepsis" before a source is even possible to consider. "ACS" before any biomarker has returned. "Anxiety" while postpartum status has not been addressed. The premature handle loads a pattern the data does not yet justify, and once loaded, the pattern anchors the rest of the workup. Disconfirming evidence has to fight an extra layer. This is how anchoring errors propagate through a team: the label becomes the working model. Anchoring and premature closure are among the best-documented contributors to diagnostic error (Croskerry 2003); they are what premature compression looks like from the inside.
Lazy compression. The opposite. Refusing to compress when the evidence does support it. The chart note that lists vitals and labs when the diagnosis is established and the team is already managing to it. The handoff that recites the history when the disease is named and the next four hours' work is what matters. The dashboard that returns lists of observations when the chart already encodes a diagnosis. Lazy compression offloads model-building onto the receiver every time, even when the receiver already knows the model.
The working rule sits between these two failures. Compress to the highest level the evidence supports. Not higher. Not lower.
The other axes
The state axis is the worked example because it carries the most cognitive load per token. The same compress-upward rule operates on the other three:
- Trajectory. Disposition outcome → trend class → quantitative trend → time series → individual measurements. Risk is a sub-aspect: the variance and tail of the prediction.
- Action. Overall plan → key decisions → specific orders → execution details.
- Confidence. Definitive → likely → possible → speculative → unknown.
The axes are practically independent: the same situation produces different per-axis profiles across messages. A night-team handoff: "DKA, gap closing, K borderline low, recheck mag AM" (compressed on all four). A cardiology consult: "DKA day 2, can we resume home metoprolol or wait twenty-four hours?" (state and trajectory compressed, action narrowed to one decision). A family update: "Your son's diabetes made his blood acidic; he's improving and will need IV insulin until tomorrow" (state at the family's codec level, action explained rather than named, the disease handle dropped because the family lacks the codec).
These four are minimum-sufficient for action, not an exhaustive ontology. An agent's full model includes more dimensions (cause, goals, affordances, the team's belief states, the patient's preferences, resource limits) that ride through the four by default and get promoted to standalone axes only when function demands it: palliative care promotes goals, addiction medicine promotes cause, psychiatry promotes other-agents' belief states. The four are practical, not complete.
One refinement on which axis leads. Diagnosis is usually the densest handle — disease names carry more model per token than any other clinical token does. But the highest-value handle is not always diagnostic. In unstable, uncertain, or action-constrained settings the densest useful handle lives elsewhere: on the trajectory axis (rapidly worsening, gap closing), on the action axis (airway watch, needs source control, cannot anticoagulate), or on the confidence axis (ACS unlikely but not excluded, safe for discharge after repeat troponin). The sharper operational rule is therefore: lead with the highest accurate handle on whichever axis best compresses the receiver's next action-relevant model. State is usually that axis. Sometimes it isn't, and the message that leads with diagnosis when the action is what matters has compressed at the wrong level even when every word is correct.
The clinical-AI implication
A clinical AI system inherits this principle on every axis. The output should start at the highest abstraction the data and reasoning support, and descend on demand or when the level above is not defendable. The system also has to expose three things the human user otherwise has to guess at: which axis and level it is operating at, the evidence that earns that level, and the uncertainty that prevents the next level up.
A summary that says "DKA" should be able to show the labs that justify it. A summary that says "undifferentiated shock" should be able to show what would have to come in to commit to a source. A summary that says "improving, discharge tomorrow" should show the trajectory data that earns it. Each axis can be compressed too high or too low independently, so each axis needs its own evidence trail.
The lazy-compression failure for AI is a differential generator that returns "ACS, PE, GERD, anxiety" when the case has earned "low-risk chest pain with nonischemic ECG and serial troponins negative." A scribe that captures the conversation but never proposes the disease-level handle has offloaded the highest-value compression step back to the clinician.
The premature-compression failure is worse. An AI that asserts "sepsis" on a single abnormal lactate, or "ACS" on nonspecific chest pain, performs premature compression with confidence. The clinician then has to fight both the bad inference and the model that has already loaded around it.
At the bedside
The intern in triage three has done the hard part: built a model from raw observations in two minutes, compressed to the disease-level handle, and emitted a four-line note that loads the receiver's full DKA framework. The protocol runs. The fluids hang. The drip starts. The compression did the coordination work.
The next three parts operationalize what happens once the level is chosen. Part 4 handles the order of elements within the chosen level. Part 5 handles what to leave out. Part 6 handles when to add the evidence. All of them depend on this one: the right level first, then everything else.
Sources
Theoretical foundations
- Clark HH, Brennan SE. Grounding in communication. In: Resnick LB, Levine JM, Teasley SD, editors. Perspectives on Socially Shared Cognition. Washington, DC: American Psychological Association; 1991. p. 127-149.
- Piantadosi ST, Tily H, Gibson E. Word lengths are optimized for efficient communication. Proceedings of the National Academy of Sciences 2011; 108(9): 3526-3529.
- Rosch E. Principles of categorization. In: Rosch E, Lloyd BB, editors. Cognition and Categorization. Hillsdale, NJ: Erlbaum; 1978. p. 27-48.
- Schank RC, Abelson RP. Scripts, Plans, Goals, and Understanding: An Inquiry into Human Knowledge Structures. Hillsdale, NJ: Lawrence Erlbaum Associates; 1977.
Clinical applications
- Croskerry P. The importance of cognitive errors in diagnosis and strategies to minimize them. Academic Medicine 2003; 78(8): 775-780.
- Schmidt HG, Rikers RMJP. How expertise develops in medicine: knowledge encapsulation and illness script formation. Medical Education 2007; 41(12): 1133-1139.
- Starmer AJ, Spector ND, Srivastava R, et al. Changes in medical errors after implementation of a handoff program. New England Journal of Medicine 2014; 371(19): 1803-1812.
Compressed Medicine · Preface · 1. The Compression Substrate · 2. The Function of the Message · 3. The Highest Accurate Abstraction · 4. The Decompression Order · 5. The Minimum Sufficient Message · 6. The Grounding Constraint · 7. The Belief-State Object · 8. The Same Wall · 9. The Defense Architecture · 10. The Temporal Loop · 10.1 The Connection · 10.2 Quiet Verification · 10.3 Quiet Acquisition · 11. The Irreversible-Action Check