chart-review-platform

ACTS Alzheimer's/dementia phenotyping — extract from a patient's clinical notes: impaired cognition (impaired_cognition), the documented APOE genotype (apoe_genotype → ε2/ε3/ε4 allele flags computed), postmenopausal status + last menstrual period, documented cognitive / depression / neuropsychiatric scale scores (MoCA, MMSE, CDR, Hachinski, Mattis DRS, TICS, GDS, Cornell, NPI, Global Deterioration stage), education years, and smoking status. Evidence-cited. Triggers on: impaired cognition, MCI, dementia, APOE, ε2/ε3/ε4, postmenopause, LMP, MoCA, MMSE, CDR, Hachinski, NPI, GDS, smoking, ACTS.

BSO-AD (Biological / Social / Other determinants of Alzheimer's Disease) NER scope skill. Activates when extracting entities under the BSO-AD ontology — 9 entity-type subtrees covering Demographic, Behavior / Lifestyle, Economic Stability, Education / Literacy, Food, Health Care, Neighborhood, Social and Community Context, and Dementia. Use this skill in combination with the universal chart-review-ner skill when the task_id is "bso-ad-ner" (or any task with task_kind=ner that pins the BSO-AD ontology).

PSMA PET/CT clinical-context chart review. Activate when assembling the prostate-cancer context a radiologist needs to interpret a PSMA PET/CT — Grade Group, PSA + trend, prior metastatic sites, treatment history (prostatectomy, radiation, ADT, ARPI, chemo, radioligand), the scan indication, and prior-imaging status — answering the PC0–PC3 chart-review questions per patient from pathology, oncology, urology, radiation-oncology, and prior imaging notes plus structured EHR data.

Asthma adherence task. Activates when extracting structured guideline-concordance signals from an asthma patient's chart — tier 0 (eligibility), tier 1 (control assessment: ACT score, exacerbation history, controller use), tier 2 (management: step therapy, spirometry follow-up, written action plan). Use this skill when task_kind=adherence and the task_id is asthma-adherence. The pipeline-extract-adherence package reads the references/questions/*.yaml + references/rules/*.yaml in this bundle to build the tier prompts and concordance verdicts.

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.

Score RUCAM (Roussel Uclaf Causality Assessment Method) for drug-induced liver injury (DILI) from EHR structured data + clinical notes. Triggers on: RUCAM, DILI causality, drug-induced liver injury, hepatotoxicity, R ratio, time to onset, rechallenge, liver injury scoring.

Extracts named entities from a biomedical / clinical text (PubMed abstract, EHR note, etc.) and normalizes each one to a concept_name from the BSO-AD ontology. Uses MCP tools on the `ner_mcp` server to enumerate the supported entity types, fetch the concept subtree for each type, and map (entity_type, candidate_label) tuples to canonical concept names defined in `concepts.json`. Writes one structured JSON record to `results/ner/` via `write_ner.py`. Use this skill whenever the user asks to extract entities from a clinical / biomedical text and normalize them against the BSO-AD ontology — phrased as "annotate this abstract", "extract entities from this note", "normalize these mentions", or "run NER on PMID 12345".

Extracts named entities from a biomedical / clinical text (PubMed abstract, EHR note, etc.) and normalizes each one to a concept_name from the BSO-AD ontology. Uses MCP tools on the `ner_mcp` server to enumerate the supported entity types, fetch the concept subtree for each type, and map (entity_type, candidate_label) tuples to canonical concept names defined in `concepts.json`. Writes one structured JSON record to `results/ner/` via `write_ner.py`. Use this skill whenever the user asks to extract entities from a clinical / biomedical text and normalize them against the BSO-AD ontology — phrased as "annotate this abstract", "extract entities from this note", "normalize these mentions", or "run NER on PMID 12345".

Universal NER (named-entity recognition) reviewer for chart-review-platform-v2. Activates when an NER task asks the agent to extract entity spans from a patient's notes and normalize each span to a concept_name from the task's ontology. Composes with a per-task scope skill (chart-review-<task-id>) that supplies the ontology and any task-specific guidance. Calls the chart_review_ner MCP server's 7 tools — list_entity_types, get_concept_tree, normalize_to_ontology, locate_in_source, set_span_label, set_span_status, get_span_review_state. Use this skill whenever the user asks to "run NER on this patient", "extract entities from these notes", "label this chart against the <name> ontology", or activates a task with task_kind=ner.

Extract cancer histology type, distant-metastasis status, and local-recurrence status from a patient's clinical notes; disease extent is computed from the latter two. Evidence-cited. Triggers on: cancer type, histology, disease extent, metastatic, local recurrence.

Lung cancer (NSCLC) molecular-testing guideline-concordance review. Activate when assessing whether an NSCLC patient's molecular workup adhered to NCCN — genomic testing performed, comprehensive genomic profiling vs single-gene, PD-L1 IHC, testing-before-therapy sequencing, essential-biomarker completeness, and targeted-therapy alignment for actionable mutations — answering the MT0–MT12 chart-review questions per patient from pathology, molecular/genomics, radiology, and oncology notes plus structured EHR data.

Colorectal cancer (CRC) NCCN guideline-concordance review. Activate when assessing whether a CRC patient's care adhered to NCCN — molecular (MSI/MMR) testing, staging workup, adequate lymph-node harvest, and stage-appropriate adjuvant chemo / neoadjuvant chemoradiation — answering concordance questions per patient from pathology, radiology, oncology, and operative notes plus structured EHR data.

Drafts a new chart-review guideline package from a research objective and reference materials. Use when the user says "draft a guideline", "create a chart review protocol", "design a phenotype rubric", "I want to validate [condition]", "set up a chart review for [study]", or provides published guidelines, SOPs, or papers and asks to turn them into a structured rubric. Produces a complete guideline directory under .claude/skills/chart-review-<task-id>/ with meta.yaml (status: draft), criteria/*.yaml, and optional seed keyword_sets, code_sets, and edge_cases. Composes with chart-review-build for interactive step-by-step authoring, and with chart-review-calibrate to validate the draft before locking. For interactive drafting via the Studio Authoring tab, prefer chart-review-build instead.

Builds a chart-review guideline interactively, interviewing the reviewer in plain conversational text then drafting the guideline package once enough is known. Use when the user says "build a guideline with me", "help me design a chart review protocol", "I want to draft this iteratively", "walk me through creating a rubric", or starts the Studio Authoring Builder flow. Subsumes chart-review-author for interactive use: produces the same .claude/skills/chart-review-<task-id>/ package but step-by-step through conversation. Composes with chart-review-calibrate (validate the draft) and chart-review-improve (iterate after calibration).

Calibrates a draft chart-review guideline before locking by computing per-criterion inter-rater reliability (Cohen's kappa) from blind dual-reviewer samples. Use when the user says "calibrate this draft", "is this guideline ready to lock", "compute kappa for this protocol", "check inter-rater agreement", "validate the rubric", "run a calibration sample", or asks to gate locking on agreement metrics. Reports per-criterion kappa with Landis and Koch buckets, surfaces disagreements with confusion matrices, and writes a calibration report. Acts as the release gate between draft and locked; composes with chart-review-improve to act on criteria that fail the kappa threshold.

Generates efficiency artifacts (keyword sets, code sets, note-type filters) from a locked guideline + its validated cohort. Use when the user says "codify the artifacts", "generate keyword sets", "extract code anchors from cohort", "speed up the agent for next runs", or after locking a guideline and wanting to make subsequent agent runs cheaper. Writes to references/{keyword_sets,code_sets,note_type_filters}/ on the locked guideline; updates each criterion's `uses:` block to reference the new artifacts. Composes with chart-review-improve (which produces guideline edit proposals; codify produces static reference artifacts) and runs post-lock — never modifies guideline shape.

Analyzes a cohort of completed chart reviews under a locked guideline to detect drift, surface override patterns, and generate Role-C-style feedback. Use when the user says "what's drifting", "any patterns in recent overrides", "Role C analysis", "cohort feedback", "is the guideline still working", "agent quality over time", "run cohort QA", or asks for a periodic or quality-assurance review of how a locked guideline is performing in production. Outputs a feedback report with detected drift, clustered override patterns, and pointers to issues that warrant a chart-review-improve proposal. Does not propose edits itself — surfaces findings only. Composes with chart-review-improve to act on the findings.

Read-only review copilot for the human reviewer during chart-review validation. Use when the reviewer asks any question while validating a patient's already-drafted assessments — "summarize this patient", "why is this medium risk?", "show me the evidence for active disease", "why did the agent pick X?", "find evidence against this", "what does the guideline say about Y?", "help me write the override reason", "before I lock, anything I missed?". Explains the drafting agent's reasoning, retrieves and groups evidence by strength, looks up guideline rules verbatim, and helps document override reasons — but never commits answers. The structured review form is the only commit path. Composes with chart-review (which creates the draft this copilot explains).

Improves an existing chart-review guideline by sampling patient charts, comparing agent reviews against human ground truth, clustering disagreements, and proposing concrete edits. Use when the user says "improve this guideline", "the agent keeps getting [criterion] wrong", "we keep overriding [field]", "tune the protocol", "fix the [criterion] guidance", "calibrate based on disagreements", or after running calibration or cohort-feedback and wanting to act on the findings. Writes proposals to proposals/<guideline-id>/<proposal-id>.yaml; never modifies the locked guideline directly. Composes with chart-review-calibrate (which surfaces κ failures that become this skill's input) and chart-review-cohort (which surfaces drift signals).

LLM-as-judge for the calibration phase. Activates when the platform asks to "judge an agent disagreement", "analyze low-confidence answer", "pre-screen for the human reviewer", or "produce judge analysis". Inspects a single (patient, criterion) cell where two reviewer agents disagree OR where one agent reported low confidence, then emits a structured JSON suggestion the reviewer sees in the VALIDATE form. Read-only — never commits answers.

Drafts an academic-paper Methods section describing a chart-review study from a locked guideline plus cohort QA statistics. Use when the user says "write the methods section", "draft paper methods", "academic methods text", "methods paragraph for the manuscript", "describe the chart review for the paper", "STROBE methods", "RECORD methods", or is preparing to publish a study based on a locked phenotype protocol. Produces past-tense, third-person markdown text (~300-500 words) with the four-paragraph structure (protocol overview, criterion structure, reviewer process + reliability, limitations) — extended to a five-paragraph structure with a deployment-stage validation paragraph when a deployment-kappa.json report is provided. Composes with chart-review-calibrate (which produces the calibration kappa) and the cohort + sample-validation workflow (which produces the deployment kappa).

LLM-as-judge for the NER calibration phase. Activates when the platform asks to "judge a span disagreement", "analyze low-confidence span", or "pre-screen a novel_candidate span" for the human reviewer. Inspects one span-pair where two NER agents disagreed (boundary, concept, or miss) OR a single agent's novel_candidate span, then emits a structured JSON suggestion the reviewer sees in the VALIDATE form. Read-only — never commits answers. Use this skill whenever the user says "judge this span", "analyze this NER disagreement", "pre-screen these spans", or the platform invokes the chart_review_ner judge phase.

Reviews a patient's electronic health record against a clinical phenotyping rubric (also called a chart-review guideline or abstraction protocol). Use whenever the user says "review this chart", "review this patient", "abstract this case", "is this lung cancer confirmed", "fill out the review form", "does this patient meet the criteria", or asks any question that maps to a guideline criterion. Reads notes, pathology reports, imaging, ICD codes, and OMOP-style structured data; consults the active phenotype skill's criteria, code sets, edge cases, and exemplars; commits answers via the chart_review_state MCP tools (set_field_assessment, select_evidence, set_summary, find_quote_offsets). Composes with phenotype scope-skills (chart-review-<noun>-phenotype) that provide the rubric definition.

Run the Playwright smoke suite against the Studio UI to catch regressions in the session workflow, phase gates, sidebar consistency, and cohort/agent rendering. Use after any UI change to client/src/ui/Workspace/ or before declaring a frontend task complete. Triggers: "run UI smoke tests", "verify the UI", "check the session workflow", "did I break the sidebar", "before I push", "smoke test the workspace".

Look up ICD-10, LOINC, CPT, and SNOMED-CT codes; expand a parent code to its descendant set; map between code systems. Use during chart review when validating a coded value or expanding a study's inclusion code list.

Aggregate audit JSONLs across reviewed records and surface fields where override rates have shifted recently. Use to detect when a chart-review-guideline criterion is becoming systematically wrong and may need a guideline update.

Verify that a quoted clinical span actually appears verbatim in the source chart at the claimed location. Use to enforce evidence-grounded answers and reject hallucinated quotes during chart review.

Compute Cohen's kappa per criterion across two or more reviewers from the audit log of locked records. Use to measure agreement on chart-review-guideline application.

Search chart note text for terms or phrases; return hits with note_id and span offsets. Use when extracting evidence from a patient's notes during chart review, especially when looking for specific clinical concepts or codes.

Apply a new SKILL bundle SHA to past locked records by archiving each affected record and reopening it for re-review. Use when a guideline change requires re-evaluating previously locked decisions under the new rules.

Draft a v.next of the chart-review-guideline-skill when drift detection flags a criterion as needing revision. Use to mechanize Role C proposals from drift signals into a methodologist-reviewable patch.

Compute a per-field semantic diff between two SKILL bundle versions; report added, removed, changed, and unchanged criteria with payload-level diffs for guidance, gates, schemas, and derivations. Use to surface what changed between two locked task SHAs.