Onset — presurgical evaluation workbench (prototype)

Patients 5 in the public set · 1 pending de-identification

Id	Year · sex · age band	Site	Recordings	Top contact	Model agreement	Last report
P07	2026 · F · 30–39	partner	2	LA3 (both)	2 disagreements	09 Sep
P04	2025 · M · 20–29	public	3	RH1 (both)	agree	08 Sep
P02	2025 · F · 40–49	public	1	LB4 / LB2	1 disagreement	06 Sep
P05	2026 · M · 18–19	public	2	—	no succeeded run	—
P01	2024 · F · 50–59	public	2	LA1 (both)	agree	02 Sep

Pending de-identification admin

sub-INC001 from partner site · arrived 04:12 · 3 identifier classes found in sidecars · held until scrubbed

No direct identifier is stored in this system. The mapping to the hospital id lives in a separate schema the application role cannot read.

Runs (succeeded)

2soz-lgbm v3 · soz-cnn v1

Models disagree on

2 contactsLA5, LB2

Data quality

1 note2 contacts marked bad

Contact ranking top 6 of 118 · evidence, not a recommendation · click a row to open its windows

Contact	Region	Tree rank	CNN rank	Evidence
LA3	L amygdala	1	1	spike bursts 120–130 s, 412–418 s
LA4	L amygdala	2	3	HFA ratio 0.31 · window p.2
LA5	L hippocampus ant.	3	11	models disagree — review windows
LH2	L hippocampus	4	2	line length 2.1× median
LB2	L temporal pole	5	14	models disagree — review windows
LH3	L hippocampus	6	5	spike rate 4.2 / min

Evidence window — LA3, 120–130 s

spike bursts (3 events) · bipolar · µV · run 43 (soz-cnn v1) · open in signal viewer

Onset Report v2 09 Sep 2026 · open full report

Both models rank LA3 first; evidence in three spike-burst windows [pred 1041, 2033]. They disagree on LA5 (3 vs 11) [pred 1043, 2041] and LB2 (5 vs 14). Two contacts excluded as bad.
Limitations: five-patient model; single site; no imaging.

No recommendation field — the clinician decides

Onset Assistant evidence only · scope: this patient

Requested

1P05 · soz-cnn v1

Running

0—

Succeeded (7 d)

92 models · 5 patients

Failed (7 d)

1feature version mismatch

Run queue

Run	Recording	Model	Features	Status	Predictions	Started	Duration / error
#44	P05 · rec 2	soz-cnn v1	fv1	requested	—	05:10	—
#43	P07 · rec 2	soz-cnn v1	fv1	succeeded	118	04:51	41 s
#42	P07 · rec 2	soz-lgbm v3	fv1	succeeded	118	04:49	6 s
#41	P07 · rec 1	soz-cnn v1	fv1	succeeded	118	04:48	39 s
#40	P02 · rec 1	soz-cnn v1	fv0	failed	0	04:40	feature version fv0 ≠ fv1 — refused before scoring
#39	P04 · rec 3	soz-lgbm v3	fv1	succeeded	96	04:31	5 s

Run #43 · events

04:51:02 requested by nightly job (job_run 118)
04:51:04 claimed by worker gpu-1
04:51:05 model soz-cnn v1 loaded from registry · fv1 ok
04:51:45 118 predictions + 214 evidence windows written (one transaction)
04:51:45 succeeded

Model registry

soz-lgbm v3 production · fv1 · top-5 precision 0.52 (LOPO)
soz-cnn v1 staging · fv1 · top-5 precision 0.48 (LOPO)
reference (line length) · 0.41
Every prediction row carries model + version. Thesis models plug in through ModelContract; the conformance suite must pass before staging.

Onset Report — Patient 07 v2 of 2 · generated 09 Sep 04:58 by the nightly job · view v1

Both models rank LA3 first, with spike bursts in three windows of recording 2 [pred 1041, 2033 · evidence 88, 91, 93]. LA4 and LH2 follow in both rankings. The models disagree about LA5 (3 vs 11) and LB2 (5 vs 14); see Disagreements. Contacts LB6 and LB7 were marked bad in channels.tsv and excluded from all rankings.

Models: soz-lgbm v3 (production), soz-cnn v1 (staging) · features fv1 · prompt onset_report_v1 (hash 3f9a2c1) · local Qwen3-8B · judge: faithfulness 5, citations 5, directives 5

Findings tree rank / CNN rank

Contact	Region	Ranks	Predictions	Evidence
LA3	L amygdala	1 / 1	pred 1041, 2033	3 windows · spike bursts
LA4	L amygdala	2 / 3	pred 1042, 2035	2 windows · HFA ratio 0.31
LH2	L hippocampus	4 / 2	pred 1044, 2034	2 windows · line length 2.1×
LH3	L hippocampus	6 / 5	pred 1046, 2037	1 window · spike rate 4.2/min

Disagreements stated, not resolved

Contact	Ranks	Predictions	Why they differ
LA5	3 / 11	pred 1043, 2041	tree: HFA; CNN: no bursts found
LB2	5 / 14	pred 1045, 2044	tree: spike rate; CNN: low

Data quality

unit check: passed (µV range ok)
line noise: 60 Hz notched (from sidecar)
bad contacts: 2 excluded (LB6, LB7)
unmatched labels: 0

Limitations

Models trained on five public patients from one site; ranks are relative within this patient only; no imaging or coordinates used; evidence windows are model-selected and not exhaustive. Calibration was insufficient, so ranks are shown rather than probabilities.

Conversation Patient 07 · scope: this patient only · four read-only tools

Tools: sql_query (read-only views, 5 s timeout, row-level security by patient) · search_documents (guideline / paper / prior report, labeled) · evidence_window · compare_models · every answer cites or refuses.

Users and roles

Email	Role	Last login	Status
marsh@clinic.local	clinician	today 08:41	active
you@nu.edu.kz	admin	today 08:12	active
ra1@nu.edu.kz	researcher	yesterday	active
ra2@nu.edu.kz	researcher	yesterday	active

Data-quality events 7 days

unit range warnings · 0
unmatched contact labels · 3 (sub-INC001, held)
line-noise mismatch · 0
feature version mismatch · 1 (run #40)
reports regenerated · 2
Alert: any recording with events above threshold pauses report generation.

Audit log append-only, written by middleware

Time	User	Action	Object	Request id
08:41:02	marsh	view	patient P07	req-7f3a
08:41:19	marsh	view	report P07 v2	req-7f3a
08:42:05	marsh	ask	assistant · P07	req-8b10
08:42:40	marsh	ask (refused)	assistant · P04 out of scope	req-8b11
04:58:11	nightly	report	P07 v2 generated	job-118
04:12:30	deid	hold	sub-INC001 · 3 identifier classes	job-117

De-identification queue

sub-INC001 · partner · held · PatientName, StudyDate, free-text notes
Date shift: per-patient offset, intervals preserved
Mapping to hospital id: separate schema, deid role only

Controls

Report generation: running · Worker: healthy (heartbeat 4 s)
Model in production: soz-lgbm v3 · promoting soz-cnn v1 requires two admins
All actions here are audited with a reason.