1 · Pick the disease

Pick a body system, then the disease to observe. From shows what follows it; to shows what precedes it. Use Build progression below to chain steps.

2 · Model confidence (non-noise)

RF ≥0.70

GNN ≥0.70

Combine

3 · Temporal correlation (tSPM)

|ρ| ≥0.00

adjusted p ≤0.05

Filter on the magnitude of ρ; the signed value is shown on the map and in the inspector. p uses the multiple-testing adjusted value.

4 · Time between conditions

0–14d 15–30d 30–90d 90+d

5 · Genetic layer (r_g)

Amber = shared genetics, blue = inverse, gray = measured-near-zero. Thin gray = no genetic mapping.

time →

model confidence ↑

Choose a condition to begin.Then tighten model confidence and ρ to trim toward a defensible shortlist.

Dig in

Click any edge on the map to seed a sequence, then extend it in either direction: backward into what precedes it, forward into what follows. Each step keeps your current confidence, ρ, and p settings. Or just double-click any node on the map, or hit Open progression network, to expand into the full step-by-step network.

Vignettes

Worked progressions, and how to read them through the three layers

Each vignette is one path a user can rebuild in the Explore tab. The logic is always the same. The classifiers (RF and GNN) decide whether a transition is real signal and trim co-occurrence noise. The temporal correlation ρ then measures how tight the ordering is. The genetic correlation r_g tests whether biology independently agrees, which separates inherited cascades from acquired, exposure-driven ones. A step with no r_g is not weak; for infection and exposure trajectories, silence on genetics is the expected and informative result.

Aging · cross-system bridge

1. Cerebrovascular injury to epilepsy

A cardioembolic source seeds a stroke, and the stroke opens a late path to epilepsy.

Heart failure

↓90+ d · ρ+0.13 · RF 0.99 / GNN 1.00 · no r_g

Acute phlebitis; thrombophlebitis and thromboembolism

↓90+ d · ρ+0.12 · RF 1.00 / GNN 1.00 · r_g +0.10

Cerebral infarction

↓90+ d · ρ+0.27 · RF 1.00 / GNN 1.00 · r_g +0.19*

Other nervous system disorders

↓90+ d · ρ+0.33 · RF 1.00 / GNN 1.00 · r_g +0.71*

Epilepsy; convulsions

How the layers read

The classifiers keep every step, and confidence climbs as the chain becomes more specific, reaching RF 1.00 and GNN 1.00 at the stroke-to-neurology transitions. The temporal correlation is modest early and strengthens into the neurological tail. Genetics corroborate exactly where it matters: the stroke-to-nervous-system step carries rg+0.19 and the nervous-system-to-epilepsy step rg+0.71, so the late, clinically important link is the one biology supports most.

Explore it yourself

Seed on Cerebral infarction in the Nervous system group, set direction to from, and double-click it to build forward. Switch direction to to in order to trace the cardioembolic precedents. Keep the time filter at 90+ days, where this cascade lives.

How to read it

This is the directed-bridge pattern: a transition that is weak as raw co-occurrence but survives because the models and the genetics agree. Read it as a hypothesis that post-stroke epilepsy risk concentrates months out, not as a claim that every stroke leads there.

Aging · shared biology

2. The cardiometabolic cascade

Obesity opens a metabolic sequence that ends at the coronary event, and the genetics tighten at every step.

Obesity

↓90+ d · ρ+0.35 · RF 0.98 / GNN 1.00 · no r_g

Essential hypertension

↓90+ d · ρ+0.59 · RF 1.00 / GNN 1.00 · r_g +0.73*

Disorders of lipid metabolism

↓90+ d · ρ+0.52 · RF 1.00 / GNN 1.00 · r_g +0.77*

Coronary atherosclerosis and other heart disease

↓15–30 d · ρ+0.39 · RF 0.96 / GNN 0.99 · r_g +0.98*

Acute myocardial infarction

How the layers read

Each step clears the gate at RF and GNN near 1.00. The temporal correlation is strong through the middle of the chain. The genetic correlations escalate as the chain advances, rg+0.73 from hypertension to lipid disorder, rg+0.77 into coronary disease, and rg+0.98 from coronary disease to the infarction. Type 2 diabetes feeds the same coronary node directly at rg+0.43.

Explore it yourself

Seed on Essential hypertension or Diabetes mellitus, Type 2 in the Circulatory or Endocrine group. Set the genetic layer to genetically supported to keep only the corroborated steps, then build forward.

How to read it

The near-unity genetic correlation between coronary disease and infarction means those two behave as one biological object recorded as a fast 15 to 30 day transition. Read the chain as a window on intervention timing, since the modifiable steps sit early, before the coronary node.

Aging · musculoskeletal

3. Fragility fracture and the musculoskeletal pain module

A fracture acts as a sentinel, and the pain phenotypes form a tight, genetically shared loop.

fracture

↓30–90 d · ρ+0.33 · RF 0.98 / GNN 1.00 · no r_g

Osteoporosis

↓90+ d · ρ+0.20 · RF 0.99 / GNN 1.00 · r_g -0.08

Low back pain

↓90+ d · ρ+0.41 · RF 0.99 / GNN 1.00 · r_g +0.61*

Musculoskeletal pain, not low back pain

↓90+ d · ρ+0.50 · RF 0.99 / GNN 1.00 · r_g +0.80*

Osteoarthritis

How the layers read

The classifiers trim hard around osteoarthritis, dropping more than half of its temporally significant transitions as age-confounded noise. What survives is tight, with the strongest steps reaching ρ+0.50. Genetics light up the pain loop, rg+0.61 from low back to general musculoskeletal pain and rg+0.80 from there to osteoarthritis, while the fracture-to-osteoporosis sentinel step is strong on timing but not genetically marked.

Explore it yourself

Seed on Osteoarthritis in the Musculoskeletal group and build forward to see the pain module. Switch direction to to and seed on Osteoporosis to see the fragility-fracture precedent.

How to read it

The high genetic correlations and the near-loop between osteoarthritis and musculoskeletal pain say this is a shared-biology module, not a one-way cascade. The fracture step is the actionable signal, since a fragility fracture may flag both bone and downstream pain trajectories.

Post-inflammatory cascade

4. Ascending infection, sepsis, and chronic heart failure

A bacterial source escalates to sepsis, and sepsis leaves a chronic cardiovascular mark.

Intestinal infection

↓90+ d · ρ+0.22 · RF 1.00 / GNN 1.00 · no r_g

Bacterial infections

↓90+ d · ρ+0.39 · RF 1.00 / GNN 1.00 · no r_g

Pneumonia (except that caused by tuberculosis)

↓90+ d · ρ+0.36 · RF 0.99 / GNN 1.00 · no r_g

Septicemia

↓90+ d · ρ+0.39 · RF 1.00 / GNN 1.00 · no r_g

Hypotension

↓90+ d · ρ+0.39 · RF 1.00 / GNN 1.00 · no r_g

Heart failure

How the layers read

Every step clears the gate at RF and GNN at or near 1.00, so the models read this as clean signal rather than incidental co-coding. The temporal correlations are steady across the infectious and the post-sepsis steps. No step carries a genetic correlation, which is expected, because this is an acquired cascade and the classifiers and timing carry the weight while genetics stay silent.

Explore it yourself

Seed on Septicemia in the Infectious group. Build forward to the post-sepsis cardiovascular tail, then switch direction to to in order to recover the infectious origin.

How to read it

The absence of genetic support is the point. A real, model-confident, well-ordered cascade with no genetic correlation is the signature of an exposure-driven trajectory, which is what post-sepsis cardiovascular morbidity should look like.

Post-inflammatory cascade · renal

5. Urinary infection, glomerular injury, and chronic kidney disease

Recurrent infection and proteinuria precede glomerular injury, which settles into chronic kidney disease.

Urinary tract infections

↓30–90 d · ρ+0.14 · RF 0.97 / GNN 0.99 · no r_g

Proteinuria

↓30–90 d · ρ+0.20 · RF 0.95 / GNN 1.00 · no r_g

Nephritis; nephrosis; renal sclerosis

↓90+ d · ρ+0.32 · RF 0.98 / GNN 1.00 · no r_g

Chronic kidney disease

↓90+ d · ρ+0.60 · RF 1.00 / GNN 1.00 · no r_g

Diabetes mellitus; renal complication

How the layers read

The classifiers keep all steps, with confidence rising into the chronic phase. The temporal correlation strengthens as the kidney injury consolidates, from ρ+0.14 early to ρ+0.60 at the endpoint. The early inflammatory steps are not genetically marked, consistent with acquired injury, while the chronic endpoint reflects shared metabolic-renal biology.

Explore it yourself

Seed on Nephritis; nephrosis; renal sclerosis in the Genitourinary group and build forward to chronic kidney disease. Switch to to in order to surface the proteinuria and infection precedents.

How to read it

The rising correlation tracks a real clinical consolidation, with acute and inflammatory renal signals giving way to a stable chronic state. The endpoint hypothesis is that diabetic kidney disease and other chronic kidney disease share a final common path worth modeling jointly.

Environmental exposure

6. Inhaled environmental lung injury to cor pulmonale

Lung injury from an external agent drives obstruction, failure, and finally right-heart disease.

Lung disease due to external agents

↓90+ d · ρ+0.15 · RF 0.81 / GNN 0.99 · no r_g

Chronic obstructive pulmonary disease and bronchiectasis

↓90+ d · ρ+0.32 · RF 1.00 / GNN 1.00 · no r_g

Respiratory failure; insufficiency; arrest

↓90+ d · ρ+0.16 · RF 0.97 / GNN 1.00 · no r_g

Pulmonary heart disease

↓90+ d · ρ+0.21 · RF 0.99 / GNN 1.00 · no r_g

Heart failure

How the layers read

The first step is the only soft one, RF 0.81, because external-agent lung disease is heterogeneous, and the gate still keeps it. From obstruction onward the models sit at RF and GNN near 1.00. The temporal correlation peaks at the obstruction-to-failure step. No step carries genetic support, the expected signature of an exposure cascade rather than an inherited one.

Explore it yourself

Seed on Chronic obstructive pulmonary disease and bronchiectasis in the Respiratory group and build forward to pulmonary heart disease. Set the genetic layer to all progressions, since this trajectory is environmental and will not appear under the genetic filter.

How to read it

This is the cleanest environmental example in the set, a directed, model-confident, genetics-silent path from an inhaled exposure to right-heart failure. It is a natural target for linking an external exposure series to a downstream cardiopulmonary endpoint.

Environmental exposure · toxic

7. Nephrotoxic exposure, acute injury, and chronic kidney disease

A drug or toxic exposure precipitates acute kidney injury that hardens into chronic disease.

Drug induced or toxic related condition

↓90+ d · ρ+0.22 · RF 0.98 / GNN 0.98 · no r_g

Acute and unspecified renal failure

↓90+ d · ρ+0.62 · RF 0.98 / GNN 1.00 · no r_g

Chronic kidney disease

↓90+ d · ρ+0.60 · RF 1.00 / GNN 1.00 · no r_g

Diabetes mellitus; renal complication

How the layers read

All steps clear the gate at RF and GNN near 1.00. The temporal correlation jumps at the acute-to-chronic transition to ρ+0.62, the strongest single link in the exposure set, marking where reversible injury becomes permanent. The exposure steps carry no genetic correlation, as expected, while the chronic-renal endpoint reflects shared renal-metabolic biology.

Explore it yourself

Seed on Acute and unspecified renal failure in the Genitourinary group. Build forward to the chronic endpoint, then switch to to in order to recover the toxic exposure trigger.

How to read it

The acute-to-chronic step is the hypothesis: the window where a nephrotoxic exposure stops being recoverable. It is the kind of timing claim the map is built to generate and an exposure study could test.

Generated from the ClaiD corpus (New England, US). Hypothesis-generating associations with timing, not causal claims. Gate: adjusted p ≤ 0.05, RF ≥ 0.70, GNN ≥ 0.70.

ClaiD

Mapping the Phylophenome of Human Disease · pronounced “Clade”

The Vision

Historically, medical research has treated human disease as a series of isolated clinical snapshots. This static approach fails to capture the dynamic, evolving reality of patient health. To truly understand, predict, and alter the course of illness, we must move beyond static phenotypes and map the continuous progression of disease over time. ClaiD, pronounced as Clade, is the foundational architecture built to achieve this. It is a computable, directed temporal knowledge base designed to track, model, and predict the complete evolutionary pathways of human health.

Defining the Phylophenome

ClaiD introduces a new scientific paradigm: the Phylophenome. Just as the genome maps human DNA and the proteome maps proteins, the Phylophenome is the definitive, comprehensive map of all temporal clinical phenotypes. It captures how specific health states interact, mutate, and branch out over a patient's lifespan, transforming complex, longitudinal health records into a structured, computable universe of disease progression.

The Architecture

The Concept of the Clade

The engine’s name, ClaiD, is rooted in both its origin — the Clinical Augmented Intelligence (CLAI) group — and its core structural philosophy derived from phylogenetics.

In evolutionary biology, a clade is a complete branch on the tree of life, containing a single ancestor and all of its descendants. ClaiD applies this exact mathematical and biological structure to clinical informatics:

The Ancestor: a patient's baseline clinical state or genetic foundation.
The Descendants: the cascading, branching sequence of subsequent diagnoses, complications, and temporal phenotypes that emerge from that baseline.

Within this framework, a clinical “clade” is a specific, highly detailed temporal trajectory of disease. By mapping these clades, ClaiD enables researchers and autonomous AI systems to trace the exact lineage of a complication back to its root, or look forward to predict the cascading effects of a single medical event.

The Impact

ClaiD is not just a repository; it is a navigational engine. By providing a structured, directed map of human disease trajectories, it creates the essential infrastructure required for the next generation of medical AI. It empowers clinical researchers to identify critical temporal bottlenecks, optimize intervention timing, and fundamentally reshape how we predict the future of a patient's health.

The Corpus

The current ClaiD corpus was derived from the longitudinal electronic health records of about 300,000 patients in the Mass General Brigham system in New England, United States. This is a first instantiation of the phylophenome, and it will expand as further populations and data sources are added.

Documentation

How to read and use the ClaiD explorer

What the map shows

Every dot is a downstream condition reached from the disease you pick. Its horizontal position is the time between the two conditions, binned as 0 to 14, 15 to 30, 30 to 90, and 90 plus days. Its vertical position is model confidence that the progression is real signal rather than administrative co-occurrence. A direction control reads what follows a condition (from), what precedes it (to), or both.

The controls

1 · Pick the disease

Choose a body system, then the condition to observe, and a direction. Progressions fan out across the time axis.

2 · Model confidence

Raise the random forest and graph network thresholds, and choose whether an edge must satisfy both models or either one. Low-confidence noise falls away first.

3 · Temporal correlation

Raise the minimum on the magnitude of the tSPM correlation, and lower the maximum adjusted p. The filter uses the size of |ρ|, while the signed value stays visible on the map. The adjusted p starts at 0.05, so the default view is already restricted to multiple-testing-significant associations.

4 · Time and 5 · Genetics

Restrict to the lag windows you care about, and switch the genetic layer to all progressions, those with a measured genetic correlation, only the genetically supported, or only the measured-null. Amber means shared genetics, blue means inverse, gray means measured near zero.

The calibration trace

The trace on the right asks whether higher model confidence tracks higher genetic support. Across the corpus it rises monotonically, so the model recovers biology it was never shown. The faint line is the full-corpus reference; the bright line is your current view, with a guard when too few genetic anchors are present to calibrate.

Building a progression

Double-click any node on the map, or use Open progression network in the dig-in panel, and the builder expands full-screen. The disease you start from anchors a bold path that flows left to right. Its likely next conditions radiate off it as green branches, each labeled with the condition name and a small tag giving the time gap and the temporal correlation ρ. Click a branch to take that step; it joins the bold path and the next branches grow from there. Build A → B → C → D → E. The paler branches are the routes you did not take.

Each step can be taken at its own time bucket: the same next condition can appear as more than one branch when it is reachable at several time gaps, so A → B at one gap does not force B → C onto the same gap. A condition never appears twice in one chain, and a condition never progresses to itself. Click any committed step to back up, and export the chain as CSV.

Co-occurrence versus progression

The shortest bin, 0 to 14 days, is treated as co-occurrence rather than progression: two conditions recorded within two weeks are usually presenting together, not one leading to the other. Those links are labeled as co-occurrence and drawn dashed. Links at 15 days and beyond are progressions.