Clinical Research Session

M4 Clinical Research Workflow

Structured clinical research from hypothesis through analysis. All work is tracked in a vitrine study — see CLAUDE.md for the vitrine quick reference or /vitrine-api for the full API.

When This Skill Activates

• User invokes /research command
• User describes research intent: "I want to study...", "Can we analyze...", "What's the mortality rate for..."
• User mentions cohort analysis, hypothesis testing, or comparative studies

Terminal and Vitrine

The researcher has the terminal and vitrine open side by side. Vitrine is where structured interaction happens — forms, data review, approvals. The terminal is where you discuss, explain reasoning, and refine.

When blocking for input (wait=True), always narrate the handoff in the terminal before the show() call. Tell the researcher what you've posted and what you need: "I've posted the study parameters form in vitrine — please fill in your outcome and exclusion criteria."

Study Setup

Every research session is organized as a study — one study per research question, spanning one or more conversations.

from vitrine import (
    show, section, register_output_dir, study_context,
    list_studies, export, Form, Question,
)

STUDY = "early-vasopressors-sepsis-v1"
output_dir = register_output_dir(study=STUDY)

# FIRST card: study description as the opening vitrine card
show("""# Early Vasopressor Use in Sepsis
...research question, design, key definitions...
""", title="Study Description", study=STUDY)

Continuing a study: Call list_studies() and study_context(study) to re-orient. Use section() to mark a new conversation within an ongoing study — not a new study.

Branching: Create a new version (v2) when the researcher wants a different approach.

Output Structure

Cards tell the story. Scripts ARE the science. Create this structure at study start:

output_dir/
├── PROTOCOL.md
├── RESULTS.md
├── scripts/
│   ├── 01_cohort_definition.py
│   ├── 02_baseline_characteristics.py
│   ├── 03_outcome_analysis.py
│   └── ...
├── data/
│   ├── cohort.parquet
│   ├── baseline_table.parquet
│   └── ...
└── plots/
    ├── age_distribution.json
    ├── kaplan_meier.json
    └── ...

Script-First Workflow

Every analysis step that produces a result shown in vitrine MUST be executed from a stored script. The script IS the analysis — not a retrospective summary of interactive work. This eliminates duplicate effort: writing the script is doing the analysis.

The pattern — for every analysis step:

1. Write the script to scripts/NN_name.py
2. Run it — outputs (parquets, PNGs) land in data/ and plots/
3. Show results in vitrine by loading the script's outputs

Interactive exploration (checking schemas, small test queries to understand data shape) is fine — not everything needs a script. But the moment you produce a result you'll show to the researcher, it comes from a stored script.

Script requirements:

• Self-contained: imports, set_dataset(), SQL strings, analysis code, output writes — everything to run python scripts/01_cohort_definition.py from the output directory
• Relative paths: use out = Path(__file__).resolve().parent.parent to locate data/ and plots/
• Saves outputs: .parquet to data/, .json to plots/ via fig.write_json() (never .html or .png)
• Plotly reload: plotly.io.from_json(open("plots/fig.json").read()) to reconstruct a Figure for show()
• Independent: each script runs on its own; later scripts load earlier outputs from data/

Example — one analysis step, start to finish:

# 1. Write the script
(output_dir / "scripts" / "01_cohort_definition.py").write_text('''\
"""01 — Define sepsis cohort from MIMIC-IV."""
from pathlib import Path
from m4 import execute_query, set_dataset

set_dataset("mimic-iv")
out = Path(__file__).resolve().parent.parent

sql = """
SELECT s.stay_id, s.subject_id, i.admission_age,
       a.hospital_expire_flag
FROM mimiciv_derived.sepsis3 s
INNER JOIN mimiciv_derived.icustay_detail i ON s.stay_id = i.stay_id
INNER JOIN mimiciv_hosp.admissions a ON s.hadm_id = a.hadm_id
WHERE i.first_icu_stay = true AND i.admission_age >= 18
"""
cohort = execute_query(sql)
cohort.to_parquet(out / "data" / "cohort.parquet")
print(f"Cohort: {len(cohort)} patients")
''')

# 2. Run it (via Bash tool)

# 3. Show results in vitrine
cohort = pd.read_parquet(output_dir / "data" / "cohort.parquet")
show(cohort, title="Sepsis Cohort", study=STUDY)

Phase 1: Research Interview

Collect study parameters through vitrine forms. The interview is adaptive — compose questions based on what you already know from the user's initial description.

Guidelines:

• Skip questions the user already answered in their prompt
• Add questions not in the library if the research question demands it (time windows, subgroup definitions, method preferences)
• Split into multiple forms when it makes sense — quick basics first, targeted follow-up after processing answers
• Narrate each blocking call in the terminal before posting

Question Library

Use from vitrine import Form, Question and compose from these. Use multiple=True for any question where multiple answers make sense (e.g., exclusion criteria, confounders). Without it, options render as single-select radio buttons.

Research question:

Question("question", question="Research Question",
         options=[
             ("Association study", "Is variable X associated with outcome Y?"),
             ("Prediction model", "Can we predict outcome Y from variables X?"),
             ("Cohort characterization", "What are the characteristics of population P?"),
         ], allow_other=True)

Study design:

Question("design", question="Study Design",
         options=[
             ("Descriptive", "Characterize a cohort — demographics, severity, outcomes"),
             ("Comparative", "Compare groups — treatment vs control, exposed vs unexposed"),
             ("Predictive", "Build or validate a prediction model"),
             ("Exploratory", "Hypothesis-generating — clustering, pattern discovery"),
         ])

Primary outcome:

Question("outcome", question="Primary Outcome",
         options=[
             ("In-hospital mortality", "hospital_expire_flag in admissions"),
             ("28-day mortality", "dod relative to admission or ICU entry"),
             ("90-day mortality", "dod relative to admission or ICU entry"),
             ("ICU length of stay", "los in icustays — beware survivor bias"),
             ("Hospital length of stay", "dischtime minus admittime — beware survivor bias"),
             ("Ventilator-free days", "28 minus days on mechanical ventilation"),
             ("Vasopressor-free days", "28 minus days on vasopressors"),
             ("AKI incidence", "KDIGO stage 2+ after exposure window"),
         ])

Exposure / intervention:

Question("exposure", question="Exposure / Intervention",
         options=[
             ("Treatment timing", "Early vs late initiation of a therapy"),
             ("Treatment dose / intensity", "High vs low dose, or trajectory over time"),
             ("Treatment received vs not", "Binary: any use within a defined window"),
             ("Severity score / biomarker", "Continuous or categorical exposure variable"),
             ("None (descriptive study)", "No exposure — cohort characterization only"),
         ])

Population:

Question("population", question="Base Population",
         options=[
             ("Sepsis (Sepsis-3)", "SOFA >= 2 + suspected infection"),
             ("Septic shock", "Sepsis-3 + vasopressor + lactate > 2 mmol/L"),
             ("ARDS / respiratory failure", "Berlin criteria or P/F ratio-based"),
             ("Cardiac arrest", "In- or out-of-hospital cardiac arrest"),
             ("Heart failure / cardiogenic shock", "Acute decompensated HF"),
             ("Acute kidney injury", "KDIGO criteria"),
             ("General ICU", "All ICU admissions, no disease-specific filter"),
         ])

Exclusion criteria:

Question("exclusions", question="Exclusion Criteria", multiple=True,
         options=[
             ("First ICU stay only", "Exclude readmissions — one observation per patient"),
             ("Age < 18", "Exclude pediatric patients"),
             ("ICU stay < 24h", "Minimum observation window — watch for immortal time bias"),
             ("Early death", "Exclude death within N hours — specify N"),
             ("DNR / comfort care on admission", "Exclude treatment limitations"),
             ("Chronic dialysis / ESRD", "Exclude pre-existing end-stage renal disease"),
             ("Missing key variables", "Exclude if critical data points are absent"),
         ])

Confounders:

Question("confounders", question="Key Confounders", multiple=True,
         options=[
             ("Age, sex", "Basic demographics"),
             ("Illness severity (SOFA)", "Organ dysfunction at baseline"),
             ("Illness severity (APACHE III / SAPS-II)", "Composite severity scores"),
             ("Charlson / Elixhauser comorbidities", "Pre-existing chronic conditions"),
             ("Admission type", "Medical vs surgical vs trauma"),
             ("Baseline labs", "Lactate, creatinine, bilirubin, platelets, etc."),
             ("Mechanical ventilation status", "On/off MV at baseline"),
             ("Vasopressor use at baseline", "Already on vasopressors at time zero"),
         ])

Dataset:

Question("dataset", question="Primary Dataset",
         options=[
             ("mimic-iv", "Full MIMIC-IV"),
             ("mimic-iv-demo", "100 patients, good for testing"),
             ("eicu", "Multi-center ICU database"),
         ], allow_other=False)

After the Interview

Review answers in the terminal. Pay attention to "Other" entries — help make them precise.

Key refinements to consider:

• Research question — Make specific and answerable: "Are sicker patients dying more?" → "Is day-1 SOFA independently associated with 30-day mortality in sepsis?"
• Outcome — Confirm operationalization (which table/column, survivor bias for LOS, follow-up window for X-free days)
• Exposure — Nail down time window, comparator, and immortal time bias risk
• Population & Exclusions — Check that exclusions don't introduce bias for this specific design
• Confounders — Check for mediators on the causal path (should NOT be adjusted for); consider propensity scores for treatment comparisons

Phase 2: Research Protocol

Draft a structured protocol. Save to output_dir / "PROTOCOL.md" and show with wait=True for approval.

## Research Protocol: [Title]

### Research Question
[Specific, answerable question]

### Study Design
[Descriptive/Comparative/Predictive/Exploratory]

### Population
**Inclusion:** [criteria]
**Exclusion:** [criteria with rationale]

### Variables
**Primary Outcome:** [definition and measurement]
**Exposure:** [definition and timing]
**Covariates:** [list with definitions]

### Analysis Plan
1. [Step with rationale]
2. ...

### Potential Biases & Limitations
- [Known limitation]

### M4 Skills to Use
- [Skill]: [Why]

Phase 3: Scientific Integrity Guardrails

Apply throughout the analysis.

Bias Prevention

Immortal Time Bias

• Define exposure at a FIXED time point (admission, 24h, 48h)
• Never use "ever received during stay" for treatments
• Use landmark analysis when appropriate

Selection Bias

• Report all exclusions with counts (CONSORT diagram)
• Analyze whether excluded patients differ
• Avoid conditioning on post-treatment variables

Information Leakage

• ICD codes are assigned at DISCHARGE — don't use for admission predictions
• Length of stay is only known at discharge
• Labs/vitals must be timestamped appropriately

Confounding by Indication

• Treatments are given to sicker patients
• Always adjust for severity (SOFA, APACHE, SAPS)
• Consider propensity scores for treatment comparisons

Statistical Rigor

• Pre-specify primary outcome; apply Bonferroni/FDR for secondary analyses
• Report cohort sizes at each step; be cautious with small subgroups
• Report missingness; consider imputation vs complete case; perform sensitivity analyses

Visualizations

Use plots liberally — a chart often reveals what a table hides.

Distributions → plots, not key-value cards. Categorical variables (race, gender, admission type) → horizontal bar chart. Continuous variables (age, LOS, SOFA) → histogram. Never show(dict) for a distribution with 5+ categories — use show(fig) instead. Reserve key-value cards for small summary stats (n, median, IQR).

Every plot card MUST have a description. Always pass a description= to show() when displaying a Plotly figure — a 1-4 sentences explanation of what the plot shows and why it matters. Example: show(fig, title="Age Distribution", description="Right-skewed distribution with median age 65; most patients are 50-80.", study=STUDY).

Other plots: Kaplan-Meier curves for survival, forest plots for effect sizes, covariate balance after matching, CONSORT flow diagrams. If you're staring at numbers and deciding what they mean, make a plot instead.

Reproducibility

Follow the script-first workflow from Study Setup — every result card traces back to a script in scripts/.

• Write → run → show. Never show results from throwaway interactive code. If you explored interactively to understand the data, crystallize the step into a script before showing results.
• Iterate on scripts, not inline. If a step needs fixing, edit the script file and re-run — don't create throwaway intermediates alongside it.
• Later scripts read earlier outputs. Step 03 loads data/cohort.parquet produced by step 01 — not by re-running the query. This makes dependencies explicit and each step independently verifiable.
• Use section() for phase transitions.
• Export the complete study at the end.

Phase 4: M4 Skills Reference

Severity Scores

Cohort Definitions

Clinical Concepts

Data & Methodology

Example Flow

User: "I want to study if early vasopressor use affects mortality in sepsis"

What you already know: comparative design, sepsis population, vasopressor exposure, mortality-related. Skip those questions.

Form: Ask only what's missing — outcome definition, "early" window, exclusion criteria, dataset. Narrate in terminal: "I've posted a form in vitrine for the details I couldn't infer — outcome, time window, exclusions, and dataset."

After response, refine in terminal:

• Anchor the time window to suspected infection onset (suspicion-of-infection skill), not ICU admission
• Recommend excluding death within the exposure window to avoid immortal time bias
• For treatment comparison, suggest propensity score matching over simple regression
• Show a follow-up form only if needed (e.g., matching method preference)

Protocol → approval → execute:

• Save protocol to output_dir / "PROTOCOL.md", show with wait=True
• Use section() to mark phases: Cohort Definition → Matching → Analysis → Conclusions
• Each phase = write scripts/NN_name.py → run → show from data/ and plots/ outputs
• Block for approval at decision points
• Export at the end: export("output/report.html", study=STUDY)

Red Flags

Stop and reconsider if you see:

• "Patients who survived to receive..." → Immortal time bias
• "Using ICD codes to identify patients at admission" → Information leakage
• "Complete cases only (N drops from X to Y)" → Selection bias
• "Treatment group had higher mortality" → Confounding by indication
• "47 significant associations" → Multiple comparisons
• "Small sample size but p < 0.05" → Likely false positive

After Completion

1. Save RESULTS.md to output_dir with findings, effect sizes, and confidence intervals
2. Show a summary card with key findings
3. Acknowledge limitations explicitly — show as a card for the record
4. Suggest validation on independent data (e.g., eICU if used MIMIC)
5. Export: export("output/study-report.html", study=STUDY)