chart-review-t1d

name

chart-review-t1d

description

T1D diabetes-type confirmation phenotype. From a patient's clinical notes and EHR structured data (OMOP conditions/drugs/measurements), extract the evidence that decides whether the patient is true/likely type 1 diabetes vs not T1D vs uncertain, and identify completed DDS-17 / T1D-DDS / PHQ-9 with scores. Evidence-cited; the DiCAYA proportion, Klompas criteria, and classification are computed from the extracted leaves. Triggers on: T1D, type 1 diabetes, diabetes type confirmation, Klompas, C-peptide, GAD65 autoantibody, DDS-17, T1D-DDS, PHQ-9.

Procedure

This is a structured-data + notes phenotype task. You extract evidence leaves; the platform computes the DiCAYA proportion, the five Klompas criteria, the Klompas roll-up, and the final classification from your leaves — you do NOT answer those computed fields.

Two kinds of leaves — two different rules for "absent"

1. Section-1 confirmation leaves feed the computed rules. ALWAYS commit them, even when there is no evidence — use the explicit negative/absent enum value (no, never, none, not_measured, not_tested, no_info). If you leave one blank, every Klompas criterion and the classification that depends on it collapse to "Pending". The code counts are integers — commit 0 when there are no codes.
2. Section-2 survey SCORES feed no computation. Leave the score unanswered (null) when the chart documents no number — never write 0 (a PHQ-9 of 0 is a real, valid score). Commit the survey status enum always; commit the score only when a number is documented.

Leaf fields YOU commit

Section 1 — diabetes codes (integers, commit 0 if none): t1d_code_count, t2d_code_count.

Section 1 — medications / supplies (enum, always commit): insulin_use (current/historical/never/no_info), glucagon_rx (yes/no), noninsulin_med (none/metformin_only/other/no_info), ketone_strip_rx (yes/no).

Section 1 — labs (enum, always commit): autoantibody_result (positive/negative/borderline/not_tested), c_peptide_result (low/normal/high/not_measured).

Section 1 — provider diagnosis (enum, always commit): endo_provider_dx (t1d/t2d/none), nonspecialist_dx (t1d/t2d/none).

Section 1 — longitudinal: earliest_t1d_date, earliest_t2d_date, first_insulin_date, first_glp1_date (ISO YYYY-MM-DD or leave blank if undated); explicit_t1d_statement (yes/no_info), glp1_indication_documented (yes/no_info), continued_t1d_treatment (yes/no_info), longterm_noninsulin_before_insulin (yes/no/no_info, always commit).

Section 2 — surveys (status always; score/date only when documented): dds17_status, dds17_score, dds17_date; t1ddds_status, t1ddds_score, t1ddds_date; phq9_status, phq9_score, phq9_date, phq9_item9.

Computed fields — do NOT answer these

t1d_code_proportion, gt50_t1d, klompas_1_glucagon, klompas_2_noninsulin, klompas_autoantibody, klompas_c_peptide, klompas_ketone_strip, overall_klompas_rule_met, t1d_insulin_preceded_t2d, and t1d_classification (the final proposal) are derived from your leaves and shown on the Computed panel. To change them, fix a leaf.

Workflow

1. Structured first. list_structured_data then read_structured_data for conditions (ICD diabetes codes), drugs (insulin, glucagon, metformin/GLP-1/ SGLT2, ketone-strip supplies), measurements (C-peptide, GAD65/IA-2/ZnT8/IAA autoantibodies). Count distinct T1D vs T2D coded events; read lab values + units.
2. Notes next. list_notes; search_notes for high-signal terms ("type 1", "T1DM", "insulin", "glucagon", "C-peptide", "GAD", "autoantibody", "endocrinology", "DDS", "T1D-DDS", "PHQ-9"); read_note/read_notes on candidates. Use the true event date — a note that says "diagnosed as a child" is historical; do not stamp it with the note date.
3. list_criteria + read_criteria([...]) for each field's allowed values and guidance. Follow each criterion exactly.
4. Commit every leaf via set_field_assessment(field_id, answer, confidence, evidence, rationale). answer must match the field's schema (a listed enum value, the raw integer count, a documented number for scores, or an ISO date).

Evidence rules

• Prefer structured citations. When the fact comes from an OMOP row, cite source:"omop" with table + row_id (and concept_id/value/unit/ evidence_date when helpful) — no note quote needed.
• Note citations cite the SMALLEST affirmative span: source:"note" with note_id, span_offsets, and the verbatim_quote. Use find_quote_offsets for exact offsets so the faithfulness gate passes. Never cite a negated or family-history sentence to support a positive.

Decision rules

• Patient + affirmative only. Exclude family history, planned/ordered items ("GAD panel ordered", "start insulin next visit"), and negations.
• Codes: count each distinct coded diabetes event; E10.*/250.x1/250.x3 → T1D, E11.*/250.x0/250.x2 → T2D. Do NOT count E08/E09/E13/O24, gestational, prediabetes, or hyperglycemia toward either.
• C-peptide: low only after verifying the value is < 0.8 ng/mL in the documented units; otherwise normal/high; not_measured if absent.
• Autoantibody: positive only for a definitive positive result; a negative does not exclude T1D; ambiguous → borderline.
• Ketone strips: a patient-specific urine-ketone/acetone strip prescription → yes; a ketone lab result is NOT a strip prescription.
• longterm_noninsulin_before_insulin: yes only when the med history clearly shows sustained non-insulin diabetes treatment before any insulin start.
• Surveys: confirm the specific instrument (DDS-17 = 17-item mean 1–6; T1D-DDS = 28-item mean 1–6; PHQ-9 = total 0–27). A generic screening code (CPT 96127, G0444, 96160/96161) does NOT identify a specific instrument — require the name / score / items / a verified local form. status ∈ present_with_score / present_without_score / mentioned_planned / blank_template / copied_forward / not_present / uncertain. Record phq9_item9 (self-harm) separately when present.
• Confidence: high = explicit structured row or documented value; medium = narrative inference within the rules; low = ambiguous → prefer the absent value over a guess.

Ground truth is Human-Only

t1d_classification is derived as the machine proposal by the guideline's sequential rule (endo dx → no insulin → objective lab → long-term non-insulin → nonspecialist dx). The final ground-truth call is Human-Only: the reviewer confirms or overrides it during VALIDATE. Your job is to get every leaf right so the proposal is well-grounded.

Commit every leaf, do NOT set the computed fields, do NOT call set_review_status, then emit a one-line summary and stop.