Pediatric epilepsies in dismech: census, AAP coverage, and the mechanism-module gap

2026-09-10. Landscape analysis plus one KB addition (a module collection). The disorder-level census itself is generated, not hand-written: it lives in research/pediatric_epilepsy_census.md and is produced by scripts/pediatric_epilepsy_census.py. Regenerate it with uv run python scripts/pediatric_epilepsy_census.py --out research/pediatric_epilepsy_census.md, which needs the MONDO SQLite build (just fetch-ontology-dbs mondo).

Summary

dismech has 109 disorder entries inside the MONDO epilepsy closure (MONDO:0005027), and 84 of them are pediatric. Another 36 entries are treated as epilepsies by the KB but sit outside that closure in MONDO, and 46 more carry seizures as an obligate or very frequent phenotype without being epilepsy entries. The American Academy of Pediatrics list of pediatric epilepsy types is almost fully covered: 20 of 23 types have a dedicated entry, and the remaining three are partial rather than absent.

The gap is not coverage. It is mechanism. Of the 191 entries in the census, 91 conform to epilepsy_excitation_inhibition_imbalance, and for 78 of them it is the only module they conform to. That module is a five-node generic chain from ion-channel dysfunction to recurrent seizures. Whatever is specific about SCN1A in Dravet, about thalamocortical circuits in absence epilepsy, about pyridoxine responsiveness, or about immune-driven epileptogenesis in Rasmussen encephalitis, none of it is in the module layer. It is prose in individual entries.

This report records what is there, what the AAP list asks for, and which mechanism lanes deserve modules. It also adds one thing to the KB: the Mechanisms of the Epilepsies module collection, which groups the six existing modules that are genuine routes to seizures.

How the census decides what counts

Three tiers, applied by script so the answer survives the next curation wave.

Tier	Test	Count
1	The entry's `disease_term`, a `has_subtypes[].subtype_term`, or an exact/narrow MONDO mapping is an `is_a` descendant of `MONDO:0005027`	109
2	MONDO does not place it under epilepsy, but the KB does: named epilepsy/seizure/DEE, or conforming to `epilepsy_excitation_inhibition_imbalance`, or listed in an epilepsy grouping, or named in the script's ILAE syndrome map	36
3	Outside tiers 1 and 2, a phenotype bound to an HP seizure term with `frequency` OBLIGATE or VERY_FREQUENT	46

Tier 2 is the interesting one. It is where the KB and MONDO disagree about what an epilepsy is. CDKL5_Deficiency_Disorder is a member of the Early_Infantile_Developmental_and_Epileptic_Encephalopathies grouping and is not in MONDO's epilepsy subtree. Neither UNC13A entry is, though both are listed in the Epilepsy_Excitation_Inhibition_Imbalance_Disorders grouping. Most of the rest are neurodevelopmental disorders whose MONDO label happens to end in "with or without seizures". None of this is an error to fix. It is a boundary worth knowing about before anyone writes a query that trusts the closure alone.

The tier-2 test began as a name match and that was not enough. Four entries no reasonable reader would leave out of a pediatric epilepsy census fell through it: CDKL5 deficiency disorder, Sturge-Weber syndrome, hemimegalencephaly, and ring chromosome 20 syndrome are epilepsies whose names do not say so. Adding module conformance, grouping membership, and presence in the ILAE map as alternative signals recovers all of them. The census now reconciles exactly against a grep: 91 files in kb/disorders/ conform to epilepsy_excitation_inhibition_imbalance and all 91 appear, a line the generated report prints so the two can be compared without arithmetic.

Age is then assigned from the ILAE 2022 Task Force position papers rather than from the entries' own annotations, because the annotations are mostly absent: 28 of the 109 tier-1 entries carry an onset_category at all. The script keys each named syndrome to its ILAE age group (neonatal and infantile, PMID:35503712; childhood, PMID:35503717; idiopathic generalized, PMID:35503716; variable age, PMID:35503725) and treats numbered gene-defined DEEs as infancy-or-early-childhood by definition. Sixteen tier-1 entries are ILAE variable-age syndromes and are reported as VARIABLE, not as pediatric. Six are UNKNOWN: they are in the closure, they are not named ILAE syndromes, and they carry no onset data.

This is a place the KB could be improved cheaply. The onset information exists in the source literature for nearly all of these entries and the schema already has the slot. Populating onset_category on the tier-1 entries would make the age question answerable from the KB rather than from a hard-coded map inside a script.

AAP coverage

20 of the 23 types on the AAP pediatric epilepsy types page have a dedicated entry. The full table is in the generated census. The three partial ones:

Reflex epilepsies. Only the photosensitive forms are curated (Photosensitive_Epilepsy, Photosensitive_Occipital_Lobe_Epilepsy). The reflex-epilepsy umbrella MONDO:0017768 has no entry, and reading epilepsy, startle epilepsy, and hot-water epilepsy are absent.
Self-limited familial and non-familial neonatal-infantile seizures. SeLNE and SeLIE are curated. The SCN2A self-limited familial neonatal-infantile form appears only as a phenotype-spectrum sentence inside the SCN2A DEE entry.
Sleep-related hypermotor epilepsy. The familial form is curated. Sporadic SHE is not a separate entry.

Two things the AAP page names that are worth a note. Its "Epileptic Encephalopathy with Continuous Spike and Wave during Sleep (CSWS)" maps to two dismech entries, because ILAE 2022 folded Landau-Kleffner syndrome into DEE-SWAS and both are kept here. Its "Doose Syndrome" also maps to two, and both carry MONDO:0014633 as disease_term, which is the duplicate-disease_term pattern tracked in issue #10113.

The mechanism gap

Module conformance across all 191 census entries:

Module	Conformers
`epilepsy_excitation_inhibition_imbalance`	91
`synaptic_vesicle_cycle`	8
`congenital_disorder_of_glycosylation`	7
`microtubule_dependent_neuronal_migration_failure`	4
everything else	1 to 3 each

73 of the 191 declare no conforms_to at all.

The shape of that table is the finding. One module absorbs the whole population, and it is the most generic one available. epilepsy_excitation_inhibition_imbalance has five nodes and no treatments, no mechanistic hypotheses, and no discussions. It says that something goes wrong with channels or synapses, then excitation exceeds inhibition, then neurons fire together, then seizures. That is true of every epilepsy, which is why 91 entries reach it, and it is why reaching it distinguishes nothing.

Meanwhile the recurring mechanism themes are visible in the entries' own pathophysiology node names and have no module to conform to:

Theme	Tier-1 entries whose pathophysiology nodes name it	Module?
Developmental and epileptic encephalopathy (seizures worsening development)	26	No
Thalamocortical spike-wave oscillation	12	No
Potassium-channel excitability	10	No
Interneuron dysfunction	9	Partly (developmental module only)
mTOR pathway	7	Yes (`pi3k_akt_mtor_cortical_overgrowth`)
Sodium-channel excitability	6	No
Neuroinflammation in epileptogenesis	6	No
Blood-brain barrier disruption	6	No
NMDA receptor dysfunction	5	Only the hypofunction direction
GABA-A receptor dysfunction	3	No
Vitamin B6 dependence	2	No

Modules worth writing, in order

developmental_and_epileptic_encephalopathy. The highest-value one. It is the concept that makes a DEE a DEE: the epileptic activity itself contributes to developmental impairment beyond what the etiology alone would cause. 26 entries already model it as a node, and the ILAE definition papers plus PMID:41014440 give it an evidence base. It is also the piece that would let the two existing DEE groupings audit their members against a criterion instead of a name.
thalamocortical_spike_wave_oscillation. The absence-epilepsy circuit: T-type calcium currents in thalamic relay neurons, GABA-B-mediated reticular inhibition, and the generalized 3 Hz spike-wave discharge. It has 12 candidate conformers, a distinct treatment logic (ethosuximide targets the T-type current; carbamazepine worsens absence), and it is a genuine circuit-level mechanism rather than a restatement of hyperexcitability.
voltage_gated_sodium_channel_neuronal_excitability. The single most therapeutically loaded lane in pediatric epilepsy, because the gain-of-function and loss-of-function directions demand opposite drugs. SCN1A loss in Dravet contraindicates sodium-channel blockers; SCN2A and SCN8A gain of function calls for them. That inversion is currently prose in each entry and would be one module with two branches.
neuroinflammatory_epileptogenesis. IL-1beta, HMGB1-TLR4, and blood-brain barrier breakdown as a route from insult to chronic epilepsy. Covers FIRES, Rasmussen encephalitis, post-traumatic epilepsy, and the BBB content currently sitting in the umbrella Epilepsy entry. Needs careful scoping: it is the lane most prone to being overstated as a universal explanation, and the existing epilepsy project brief already says so.
vitamin_b6_dependent_epilepsy. Small, but the cleanest treatable mechanism in the whole pool: an enzymatic block starves pyridoxal 5'-phosphate-dependent decarboxylases, and the seizures answer to the vitamin rather than to antiseizure medication. Two conformers today (Pyridoxine-Dependent_Epilepsy, PNPO_Deficiency), more later.

A sixth candidate, a GABA-A receptor module, is real but thin at three entries; it may be better as a branch of the sodium-channel module's successor rather than its own file.

What was added

kb/module_collections/Mechanisms_of_the_Epilepsies.yaml, a MECHANISTIC_FAMILY collection with six members: the excitation-inhibition convergence module plus five upstream lanes (interneuron specification and tangential migration failure, the synaptic vesicle cycle, excitatory synapse scaffold disruption, PI3K-AKT-mTOR cortical overgrowth, and FAME pentanucleotide repeat RNA toxicity).

Membership is modules that are an evidenced route to seizures or to excitation-inhibition imbalance in at least one curated entry. It is deliberately not a rule about a module's terminal node, because two correct members would fail that test: synaptic_vesicle_cycle ends at neurotransmitter release failure and never says the word seizure, and excitatory_synapse_scaffold_disruption passes through excitation-inhibition imbalance at node three and then continues to a generic neurodevelopmental terminus. The first has ten conformers in the KB, of which eight are epilepsy or DEE entries; the other two, SYT1 Baker-Gordon syndrome and the VAMP2-related disorder, are neurodevelopmental disorders that fall outside this census, and SYT1 carries no seizure phenotype at all. The module earns its place on the eight.

Four modules that recur in epilepsy entries were deliberately left out, and the collection's notes records why. In each, seizures are incidental to the chain rather than its point. glutamate_excitotoxicity is a route to neuronal death. metabolic_intoxication_decompensation is a route to acute encephalopathy. epigenetic_machinery_neurodevelopmental_dysregulation and congenital_disorder_of_glycosylation are general neurodevelopmental modules in which seizures are one output among many. A collection that admitted those would be an inventory of modules epilepsy entries happen to touch, which is the failure mode the create-module skill warns about.

The collection is not pediatric-only, on purpose. Mechanism lanes are shared across ages and age at onset is a property of the disease entries. That is why the FAME module is a member despite familial adult myoclonus epilepsy being adult-onset.

Follow-ups

Populate onset_category on the tier-1 entries that lack it, so the pediatric verdict comes from the KB rather than from the script's ILAE map.
Write the five proposed modules, highest value first, and retrofit conformance.
Decide whether tier-2 entries such as CDKL5 deficiency disorder should carry a mappings.mondo_mappings entry that places them under epilepsy, or whether the MONDO boundary is correct and the KB grouping is the right home.
Resolve the two entries sharing MONDO:0014633 (issue #10113).
Curate the reflex-epilepsy umbrella and sporadic SHE if the AAP list is to be fully covered.