Identify evidence of depression from a patient's post-index clinical notes: explicit diagnosis, depressive symptoms, antidepressant use, psychiatry referral, and PHQ-9 severity. Tier and final Depression/No Depression decision are computed. Evidence-cited. Triggers on: depression, PHQ-9, depressive disorder, MDD, antidepressant.
Identify evidence of depression from a patient's post-index clinical notes: explicit diagnosis, depressive symptoms, antidepressant use, psychiatry referral, and PHQ-9 severity. Tier and final Depression/No Depression decision are computed. Evidence-cited. Triggers on: depression, PHQ-9, depressive disorder, MDD, antidepressant.
Procedure
This is a notes-only phenotype task for a GLP-1 / obesity-T2D cohort. You
answer five leaf fields directly:
high_confidence_diagnosis — yes / no / no_info.
depressive_symptoms — yes / no / no_info.
antidepressants — yes / no / indication_not_verified / no_info.
study1_tier, phq9_threshold_met, and final_decision are computed
fields — do NOT answer them directly.
This is a two-pass procedure. Do not read a note and immediately commit a
field answer from it — that is exactly the failure mode this task is prone
to (this chart type has the same diagnosis and PHQ-9 score repeated across
many visits; answering from the first note you read will silently drop the
other five, ten notes that also support it).
Pass 1 — per-note scan (build the evidence index)
Call list_structured_data FIRST. Its response includes index_date
for this patient — record it. Then call list_notes, which returns each
note's date. Only notes with a date strictly AFTER index_date are
in scope — a note dated exactly ON the index date is OUT of scope.
Drop every out-of-scope note from your working set now, before Pass 1
starts; do not open or cite them.
list_criteria + read_criteria([...]) for all five leaf fields, before
reading any notes, so you know what to look for.
Go through the in-scope notes per your session's search mode
(smart-search: search_notes for high-signal terms, then
read_note/read_notes on matches; comprehensive: read every in-scope
note in full, one at a time, in date order).
For EVERY note you read, before moving to the next note: scan it
against all five fields and call select_evidence(evidence, field_id, category, rationale) once for each passage in that note relevant to
each field it bears on. One note commonly informs more than one
field (e.g. a visit note with both a diagnosis line and a PHQ-9 score
pins two select_evidence calls, one per field_id) — pin all of them
before advancing. Do this even for a note you're fairly sure won't
change the final answer; the index has to be complete, not just
sufficient. Skipping this step for a note is equivalent to not having
read it.
You must reach the last in-scope note before starting Pass 2. Stopping
early because you already have "enough" evidence for a field is the
discipline failure this procedure exists to prevent.
Pass 2 — synthesize (commit the five leaf fields)
Call get_review_state to retrieve everything you pinned in Pass 1.
For each of the five leaf fields, gather every pinned item with that
field_id and commit ONE answer via set_field_assessment(field_id, answer, confidence, evidence, rationale). The answer MUST be one of
the field's enum values.
Multi-citation rule (mandatory):evidence is an ARRAY. Copy in
EVERY distinct note's pinned item that supports the committed answer —
if six separate visits each document the same diagnosis, or two notes
each mention a different depressive symptom, all of them go in as
separate evidence items. A single-item evidence array is only correct
when Pass 1 pinned exactly one supporting item for that field — check
your Pass-1 index before assuming that. The rationale should
synthesize across all cited spans and name any tension between notes
(e.g. a diagnosis mentioned in five visits but denied in a sixth).
Span discipline — cite the SMALLEST span that supports each point:
Quote the single sentence/phrase that justifies it (well under ~300
chars) per evidence item. Use find_quote_offsets to get exact offsets
so the faithfulness gate passes.
Do NOT cite a whole note as one block.
Every cited span must be affirmative — never cite a negated
sentence ("denies depression", "no suicidal ideation") to support a
positive answer; a negation supports no.
For no / no_info / not_documented: always cite at least one short
span — the section(s) you checked where the info would appear if
present.
Decision rules (apply across all leaf fields)
Priority sections: HPI, Assessment/Plan (incl. Problem List),
Social History, Medications/Home Medications. Exclude: generic
Discharge Instructions boilerplate ("Call your doctor if you have suicidal
feelings") and patient-instruction templates — these are not patient
evidence.
Negation: if a depression term is preceded within ~80 characters by
"no", "not", "denies", "denied", "denying", "without", "negative",
"absence of", "absent", "never" — it is negated. Do not count it.
Non-psychiatric "depression": EKG/cardiology ("ST depression",
"ST elevation or depression"), cardiac function ("depression of systolic
function"), orthopedic/anatomical ("depression of the lateral tibial
plateau", "depressed fracture") are never evidence for
high_confidence_diagnosis.
Antidepressant alternate indications — check the note for a documented
non-depression reason before counting a drug as positive:
Neuropathic pain, migraine prophylaxis, ENT dizziness
Escitalopram (Lexapro)
GAD, PPPD, panic disorder
Sertraline (Zoloft)
Anxiety, OCD, PTSD
If a non-depression indication IS documented → no. If the drug is
present but no indication is documented at all → indication_not_verified
(still meaningful evidence, just unverified — do not silently drop it).
GLP-1 confound: if the only "symptoms" present are weight loss, poor
appetite, or fatigue, AND the note documents concurrent GLP-1 therapy,
treat these as a likely medication effect rather than depressive symptoms
— do not answer depressive_symptoms=yes on these alone. Note the
ambiguity in rationale. A genuine depressive-symptom mention (mood,
anhedonia, hopelessness, guilt, SI, concentration) is unaffected by this
caution.
PHQ-9 extraction:
You must scan every in-scope note for a PHQ-9 mention before
committing this field — the highest score is frequently NOT in the first
note you find one in. Cite every distinct post-index PHQ-9 occurrence you
found as a separate evidence item (per the multi-citation rule above),
not just the one that set the band.
Use the highest post-index PHQ-9 total score found anywhere in the
chart to pick the band: 0–4 minimal, 5–9 mild, 10–14 moderate,
15–19 moderately_severe, 20–27 severe.
"PHQ 18/27" → total 18 (27-point base confirms PHQ-9).
Multiple PHQ-9 mentions in the same note → use the last occurrence
in that note.
A carried-forward score cited on a later date is NOT a new assessment —
attribute it to its original date, and don't let it override a higher,
more recent, genuinely new score.
Qualitative wording alone ("mildly depressed") without a number → do
NOT assign a band from it; keep looking for a real PHQ-9 score.
PHQ-2 (max 6 points) is NOT PHQ-9 — never substitute it in.
If PHQ-9 and GAD-7 both appear in the same note, verify carefully which
number belongs to PHQ-9.
No PHQ-9 documented anywhere post-index → not_documented.
Record the actual numeric score(s) you found in rationale even though
the committed answer is the band.
Confidence:high = explicit, unambiguous documentation (diagnosis
term in Assessment/Plan, a clearly-dated PHQ-9 number); medium =
inferred/pattern-based (e.g. antidepressant with no indication documented,
symptom cluster without explicit diagnosis); low = ambiguous or
borderline — prefer no_info/not_documented over a low-confidence guess.
Confirm all FIVE leaf fields (high_confidence_diagnosis,
depressive_symptoms, antidepressants, psychiatry_referral,
phq9_severity_band) have a set_field_assessment — every leaf must
have a value (use no_info/not_documented if genuinely absent, but
only after Pass 1 covered every in-scope note). Do NOT commitstudy1_tier, phq9_threshold_met, or final_decision — they are
derived automatically. Do NOT call set_review_status. Once the
five leaves are committed, emit a one-line summary and stop.
