M4 Clinical Research Workflow

This skill guides you through a structured clinical research session. It ensures scientific rigor from hypothesis formation through analysis execution.

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

Phase 1: Research Interview

Before writing any queries, interview the user to establish:

1. Research Question

Ask: "What specific clinical question are you trying to answer?"

Good questions are:

• Specific and answerable with available data
• Clinically meaningful
• Novel or confirmatory of existing findings

Help refine vague questions:

• "Are sicker patients dying more?" → "Is day-1 SOFA score independently associated with 30-day mortality in sepsis patients?"

2. Study Design

Ask: "What type of study is this?"

• Descriptive: Characterize a population (demographics, distributions)
• Comparative: Compare groups (exposed vs unexposed, treatment A vs B)
• Predictive: Build or validate a prediction model
• Exploratory: Hypothesis-generating analysis

3. Outcome Variable

Ask: "What is your primary outcome?"

Common outcomes and how to define them:

• In-hospital mortality:

  hospital_expire_flag

  in admissions table
• 30-day mortality: Compare

  dod

  to discharge time + 30 days
• ICU length of stay:

  los

  in icustays, be wary of survivor bias
• Ventilation duration: Requires careful definition (see m4 skills)
• Readmission: Subsequent

  hadm_id

  for same

  subject_id

4. Exposure/Intervention

Ask: "What exposure or intervention are you studying?"

For treatment comparisons:

• How is treatment defined? (any use vs duration vs dose)
• When is exposure status determined? (admission, 24h, 48h)
• What's the comparator? (no treatment, alternative treatment)

5. Population (Inclusion/Exclusion)

Ask: "Who should be included in this study?"

Standard considerations:

• First ICU stay only? (avoid correlated observations)
• Age restrictions? (pediatric exclusion common)
• Minimum ICU stay? (be careful of immortal time bias)
• Specific diagnoses? (how defined - ICD codes have limitations)

6. Confounders

Ask: "What factors might confound your results?"

Common confounders in ICU research:

• Age, sex, comorbidities
• Illness severity (SOFA, APACHE, SAPS)
• Admission type (medical vs surgical vs trauma)
• Hospital/unit effects

7. Dataset Selection

Ask: "Which dataset should we use?"

• mimic-iv: Full MIMIC-IV (requires access)
• mimic-iv-demo: 100 patients, good for testing queries
• mimic-iv-note: MIMIC-IV with clinical notes
• eicu: Multi-center ICU data (different schema)

Phase 2: Draft Research Protocol

After the interview, produce a structured research plan:

## Research Protocol: [Title]

### Research Question
[Specific, answerable question]

### Hypothesis
[If applicable - null and alternative]

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

### Dataset
[Selected dataset with justification]

### Population
**Inclusion Criteria:**
- [Criterion 1]
- [Criterion 2]

**Exclusion Criteria:**
- [Criterion 1]
- [Criterion 2]

### Variables
**Primary Outcome:** [Definition and how measured]
**Exposure:** [Definition and timing]
**Covariates:** [List with definitions]

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

### Potential Biases & Limitations
- [Known limitation 1]
- [Known limitation 2]

### M4 Skills to Use
- [Relevant skill 1]: [Why]
- [Relevant skill 2]: [Why]

Phase 3: Scientific Integrity Guardrails

Apply these checks 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

Multiple Comparisons

• Pre-specify primary outcome
• Apply Bonferroni/FDR correction for secondary analyses
• Report all analyses performed, not just significant ones

Sample Size

• Report cohort sizes at each step
• Be cautious with small subgroups
• Consider power for planned comparisons

Missing Data

• Report missingness rates for all variables
• Consider imputation vs complete case analysis
• Perform sensitivity analyses

Reproducibility

Query Documentation

• Save all SQL queries with timestamps
• Document data versions used
• Note any manual data cleaning steps

Analysis Trail

• Number analyses sequentially
• Distinguish exploratory from confirmatory
• Record decision points and rationale

Phase 4: Using M4 Skills

Match skills to research needs:

Severity Scores

Use when adjusting for baseline illness severity:

sofa-score

apsiii-score

sapsii-score

oasis-score

sirs-criteria

Cohort Definitions

Use when defining study populations:

sepsis-3-cohort

first-icu-stay

suspicion-of-infection

Clinical Concepts

Use when defining exposures or outcomes:

kdigo-aki-staging

vasopressor-equivalents

baseline-creatinine

gcs-calculation

Data Quality

Use when building queries:

clinical-research-pitfalls

mimic-table-relationships

mimic-eicu-mapping

Code Execution

Use for complex analyses:

m4-api

Example Interview Flow

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

AI Interview:

1. Research Question: "Let me help refine this. By 'early', do you mean within 1 hour, 6 hours, or 24 hours of sepsis onset? And by 'mortality', do you mean in-hospital, 28-day, or 90-day?"

2. Study Design: "This is a comparative study - you're comparing early vs. late/no vasopressor use. We'll need to carefully define the comparison groups to avoid immortal time bias."

3. Outcome: "Let's use 28-day mortality so we capture both in-hospital and post-discharge deaths."

4. Exposure: "I recommend defining 'early' as vasopressor initiation within 6 hours of suspected infection onset (using the suspicion-of-infection skill). The comparison group would be those who received vasopressors >6h or never."

5. Population: "I suggest: adult patients (>=18), first ICU stay, meeting Sepsis-3 criteria. Exclude patients who died within 6 hours (immortal time) and those on vasopressors at admission."

6. Confounders: "We should adjust for: age, SOFA score at sepsis onset, admission source, comorbidities. I recommend using propensity score matching given this is an observational treatment comparison."

7. Dataset: "Let's use mimic-iv. The demo dataset is too small for treatment effect studies."

Common Research Patterns

Pattern: Mortality Risk Factors

1. Define cohort (first-icu-stay)
2. Extract baseline characteristics
3. Calculate severity (sofa-score or apsiii-score)
4. Define mortality outcome
5. Multivariable regression

Pattern: Treatment Effect

1. Define cohort and time zero
2. Define exposure window (fixed time)
3. Extract confounders at baseline
4. Propensity score matching
5. Compare outcomes

Pattern: Cohort Description

1. Define cohort
2. Demographics, comorbidities
3. Severity scores
4. Treatments received
5. Outcomes (mortality, LOS, complications)

Red Flags to Watch For

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
• "We found 47 significant associations" → Multiple comparisons
• "Small sample size but p < 0.05" → Underpowered, likely false positive

After Analysis Completion

1. Summarize findings with effect sizes and confidence intervals
2. Acknowledge limitations explicitly
3. Suggest validation on independent data (e.g., eICU if used MIMIC)
4. Provide reproducibility info: queries used, cohort flow, data version