Two posts, one missing piece

Avinash Kumar's recent posts describe the same insight from two directions. "Cohort-Level Clinical Intelligence" argues that once you apply the Avinash Principle across a whole patient cohort rather than one chart, the combinatorial noise of population health records collapses into a small set of scale-free macro-patterns: presentation clusters, a "key decision node" where the population branches into survival vs. failure trajectories, an intervention "hub," and longitudinal tracking signatures that correlate intermediate physiology with final outcome.

"Accelerating the Evidence Pyramid" takes that structure and asks what happens when a clinician needs it right now, at the bedside. Its answer is a template-driven query engine: a clinician's question is matched against pre-computed cohort analytics and dropped into a standardized brief — query match %, active hub node, presentation metrics, required intervention, predicted outcome — instead of triggering a free-text LLM essay.

Both posts describe the same missing piece without naming it: for a "query" to match a "cohort," and a "hub node" to mean the same thing every time, the underlying clinical concepts have to be coded, not just written. That's the job SNOMED CT does — and it's what the rest of this tour builds, live, one stop at a time.

The enabling layer

SNOMED CT (Systematized Nomenclature of Medicine — Clinical Terms) is what turns the Avinash Principle from a compelling metaphor into something a machine can execute accurately:

• One normalized surface for the whole cohort. Every finding, procedure, and situation gets one canonical concept ID no matter how it was phrased — the precondition for computing a real "cohort frequency match" instead of a plausible-sounding guess.
• Compositional grammar, not a flat code list. Concepts combine with relationships like finding site, severity, and clinical course, so "hemostatic failure at T+15m" and "at T+45m" are structurally related — the substrate for the temporal-velocity mapping the cohort post describes.
• A hierarchy built for roll-up. Thousands of individually coded presentations collapse to a handful of hub nodes because the parent/child structure already defines the roll-up.
• Formal logic that catches nonsense before it reaches a clinician — directly addressing the "black box" hallucination risk the second post calls out.
• Cross-mapping to LOINC and drug terminologies, so a hub node links mechanically to the lab order that measures it and the intervention that treats it.

In short: the Avinash Principle needs a way to say the same clinical thing the same way, every time, with enough internal structure to support time windows and hierarchy roll-ups. That's a description of SNOMED CT's design goals — and it's the layer every stop below is built on.

Ask a bedside question in plain language

This is where the pipeline actually starts: a clinician types what they're seeing, in their own words. No coding knowledge required at this stage — that's the point. Type your own question below, or pick one of the examples, then press Run query. The rest of this tour fills in live as you scroll, using whatever you typed here.

Natural language becomes SNOMED CT concepts

A lightweight, deterministic clinical-NER layer — not a freeform generative one — maps each clinical phrase in your query to a canonical SNOMED CT concept. This is the accuracy checkpoint of the whole pipeline: if a phrase can't be resolved to a concept with confidence, the system should say so rather than guess, the same discipline the accelerating-evidence-pyramid post asks of the downstream brief.

Concepts compose into an ECL cohort query

The extracted concepts are composed into a SNOMED CT Expression Constraint Language (ECL) query — SNOMED CT's own formal query language, purpose-built for exactly this. The << operator pulls in descendant concepts automatically, so a query for "renal dysfunction" includes AKI, CKD-on-conservative-management, and oliguria subtypes without anyone enumerating every leaf term by hand.

The ECL query retrieves a matching cohort

The ECL result set is joined against a pre-computed cohort analytics layer — presentation timelines, decision nodes, intervention timelines, tracking correlations — all already indexed by SNOMED CT concept ID. Graph centrality metrics (the "super-connected hub" language from the cohort post) are computed once, offline, over the whole cohort. The query at runtime is a lookup, not a live computation — which is what makes sub-second template population possible.

Meet a few real patients inside this cohort

Before the aggregate — this is what the cohort is actually built from. Each tab below is one real patient's coded trajectory, laid out exactly like a single-case chart review: lanes for comorbidity → symptom → diagnosis → intervention → escalation → outcome, timed by day-offset from admission. The cohort graph two stops from now is nothing more than many charts like these, overlaid and counted.

The match populates a Clinical Intelligence Brief

The matched hub node and its linked intervention and tracking concepts drop into the fixed brief structure from the accelerating-evidence-pyramid post — no free-text generation in the clinical-fact fields, only in optional surrounding narrative. This is the artifact a clinician actually reads at the bedside.

The same match, written out as a templatized insight report

A graph is for pivoting; a report is for reading and citing. This stop takes the identical matched cohort and pours it through a second, fixed template — the three-dimension structure from "Cohort-Level Clinical Intelligence": Presentation Insights, Intervention Insights, and Tracking & Correlation Insights, each broken into the post's own named sub-fields. No sentence here is generated free-form — every value is a direct read from the same cohort object that drew the graph below; only the surrounding template prose is fixed English.

Two worked examples

These two are pre-run so you can read a complete report without typing anything — the same template, filled from the two demo cohorts on this page.

The cohort renders as a trajectory graph

This is where a single-case chart-review tool becomes a cohort tool. Each node below represents every patient in the matched cohort who shares that SNOMED CT concept at that day-offset window — sized by patient count, colored toward red as its adverse-outcome rate rises, and ringed in gold if it's the active hub node. Click a node to pivot the graph around it: connected nodes stay lit, everything else dims, exactly like tracing one patient's chart, except every edge weight is now a real cohort transition frequency.

Click the hub node's pivot chip below to see the pre-computed hub route — the exact chain the brief and insight report above describe — highlighted in gold across the graph.