app.py

import pandas as pd
import numpy as np
from sklearn.linear_model import LogisticRegression, LinearRegression
import gradio as gr

REQUIRED_COLS = [
    "treatment",            # 0/1 (0 = control, 1 = new drug)
    "outcome",              # 0/1 or continuous outcome
    "age",
    "sex",                  # 0/1 or M/F convertible
    "baseline_risk_score",
    "comorbidity_index",
]


def propensity_covariate_adjustment(file):
    if file is None:
        return "❌ Please upload a CSV file."

    try:
        df = pd.read_csv(file.name)
    except Exception as e:
        return f"❌ Error reading file: {e}"

    # Check required columns
    missing = [c for c in REQUIRED_COLS if c not in df.columns]
    if missing:
        return (
            "❌ Missing required columns: "
            + ", ".join(missing)
            + f"\n\nYour columns: {list(df.columns)}"
        )

    # Make a copy to avoid warning issues
    df = df.copy()

    # Basic cleaning
    # Ensure numeric types where needed
    df["treatment"] = pd.to_numeric(df["treatment"], errors="coerce")
    df["outcome"] = pd.to_numeric(df["outcome"], errors="coerce")
    df["age"] = pd.to_numeric(df["age"], errors="coerce")
    df["baseline_risk_score"] = pd.to_numeric(df["baseline_risk_score"], errors="coerce")
    df["comorbidity_index"] = pd.to_numeric(df["comorbidity_index"], errors="coerce")

    # Handle sex if it's "M"/"F"
    if df["sex"].dtype == object:
        df["sex"] = df["sex"].str.upper().map({"M": 0, "F": 1})
    df["sex"] = pd.to_numeric(df["sex"], errors="coerce")

    # Drop rows with any missing key values
    df = df.dropna(subset=REQUIRED_COLS)
    if df.shape[0] == 0:
        return "❌ After cleaning, no valid rows remain. Please check your data."

    # Crude (unadjusted) treatment effect: difference in mean outcome
    treated = df[df["treatment"] == 1]
    control = df[df["treatment"] == 0]

    if treated.shape[0] == 0 or control.shape[0] == 0:
        return "❌ Need both treated (treatment=1) and control (treatment=0) subjects."

    crude_effect = treated["outcome"].mean() - control["outcome"].mean()

    # ----------------------------
    # Step 1: Propensity score model
    # ----------------------------
    X_ps = df[["age", "sex", "baseline_risk_score", "comorbidity_index"]]
    y_treat = df["treatment"]

    try:
        ps_model = LogisticRegression(max_iter=1000)
        ps_model.fit(X_ps, y_treat)
    except Exception as e:
        return f"❌ Error fitting propensity score model: {e}"

    # Predicted propensity scores
    df["propensity_score"] = ps_model.predict_proba(X_ps)[:, 1]

    # ----------------------------
    # Step 2: IPTW (Inverse Probability of Treatment Weighting)
    # ----------------------------
    # IPTW weights: treated = 1/PS, control = 1/(1-PS)
    df["iptw_weight"] = np.where(
        df["treatment"] == 1,
        1.0 / df["propensity_score"],
        1.0 / (1.0 - df["propensity_score"])
    )
    
    # Stabilized weights (optional but often used)
    # p_treated = df["treatment"].mean()
    # df["iptw_stabilized"] = np.where(
    #     df["treatment"] == 1,
    #     p_treated / df["propensity_score"],
    #     (1 - p_treated) / (1.0 - df["propensity_score"])
    # )
    
    # Recalculate treated/control with updated df
    treated = df[df["treatment"] == 1]
    control = df[df["treatment"] == 0]

    # Weighted means for outcomes
    weighted_mean_outcome_treated = np.average(treated["outcome"], weights=treated["iptw_weight"])
    weighted_mean_outcome_control = np.average(control["outcome"], weights=control["iptw_weight"])
    iptw_effect = weighted_mean_outcome_treated - weighted_mean_outcome_control

    # ----------------------------
    # Step 3: Standardized Mean Differences (SMD)
    # ----------------------------
    def calculate_smd(mean1, mean2, std1, std2):
        """Calculate standardized mean difference"""
        pooled_std = np.sqrt((std1**2 + std2**2) / 2)
        if pooled_std == 0:
            return 0.0
        return (mean1 - mean2) / pooled_std
    
    def calculate_weighted_std(values, weights):
        """Calculate weighted standard deviation"""
        weighted_mean = np.average(values, weights=weights)
        weighted_var = np.average((values - weighted_mean)**2, weights=weights)
        return np.sqrt(weighted_var)

    # Covariates to check balance for
    covariates = ["age", "sex", "baseline_risk_score", "comorbidity_index", "propensity_score"]
    
    smd_results = []
    for cov in covariates:
        # Before adjustment (unadjusted)
        mean_treated_before = treated[cov].mean()
        mean_control_before = control[cov].mean()
        std_treated_before = treated[cov].std()
        std_control_before = control[cov].std()
        smd_before = calculate_smd(mean_treated_before, mean_control_before, 
                                   std_treated_before, std_control_before)
        
        # After adjustment (IPTW weighted)
        mean_treated_after = np.average(treated[cov], weights=treated["iptw_weight"])
        mean_control_after = np.average(control[cov], weights=control["iptw_weight"])
        std_treated_after = calculate_weighted_std(treated[cov], treated["iptw_weight"])
        std_control_after = calculate_weighted_std(control[cov], control["iptw_weight"])
        smd_after = calculate_smd(mean_treated_after, mean_control_after,
                                 std_treated_after, std_control_after)
        
        smd_results.append({
            "Covariate": cov,
            "Mean_Treated_Before": mean_treated_before,
            "Mean_Control_Before": mean_control_before,
            "SMD_Before": smd_before,
            "Mean_Treated_After": mean_treated_after,
            "Mean_Control_After": mean_control_after,
            "SMD_After": smd_after
        })
    
    # Create balance table
    balance_table = "| Covariate | Mean (Treated) Before | Mean (Control) Before | SMD Before | Mean (Treated) After | Mean (Control) After | SMD After |\n"
    balance_table += "|-----------|----------------------|----------------------|------------|---------------------|---------------------|-----------|\n"
    for r in smd_results:
        balance_table += (
            f"| {r['Covariate']} | {r['Mean_Treated_Before']:.3f} | {r['Mean_Control_Before']:.3f} | "
            f"{r['SMD_Before']:.3f} | {r['Mean_Treated_After']:.3f} | {r['Mean_Control_After']:.3f} | "
            f"{r['SMD_After']:.3f} |\n"
        )
    
    # ----------------------------
    # Step 4: Covariate adjustment
    # outcome ~ treatment + propensity_score
    # ----------------------------
    X_adj = df[["treatment", "propensity_score"]]
    y_out = df["outcome"]

    lin_model = LinearRegression()
    lin_model.fit(X_adj, y_out)

    # Coefficients: intercept + beta_treatment + beta_ps
    intercept = lin_model.intercept_
    beta_treat = lin_model.coef_[0]
    beta_ps = lin_model.coef_[1]

    # Summaries
    avg_ps_treated = treated["propensity_score"].mean()
    avg_ps_control = control["propensity_score"].mean()
    avg_iptw_treated = treated["iptw_weight"].mean()
    avg_iptw_control = control["iptw_weight"].mean()

    n_treated = treated.shape[0]
    n_control = control.shape[0]

    text = f"""
# Propensity Score Covariate Adjustment – Drug Development Example

## 1. Data Summary

- Number of patients: **{df.shape[0]}**
- Treated (new drug): **{n_treated}**
- Control (standard of care): **{n_control}**

Outcome is interpreted as:
- 1 = event of interest (e.g., progression-free at 12 months)
- 0 = no event (e.g., progressed or not progression-free)

---

## 2. Crude (Unadjusted) Treatment Effect

Unadjusted difference in mean outcome:

- Mean outcome (treated): **{treated["outcome"].mean():.3f}**
- Mean outcome (control): **{control["outcome"].mean():.3f}**

**Crude effect (treated - control):** **{crude_effect:.3f}**

This ignores all baseline differences between the two groups.

---

## 3. Propensity Score Model

We fit a logistic regression to estimate the probability of receiving the new drug:

**P(treatment=1 | age, sex, baseline_risk_score, comorbidity_index)**

Average estimated propensity scores:

- Treated group: **{avg_ps_treated:.3f}**
- Control group: **{avg_ps_control:.3f}**

A big difference here indicates some baseline imbalance in who gets treated.

---

## 4. Standardized Mean Differences (Balance Table)

Standardized Mean Differences (SMD) measure the balance of covariates between treated and control groups. 
SMD < 0.1 is generally considered well-balanced. SMD < 0.25 is often acceptable.

**Balance Before vs After IPTW Weighting:**

{balance_table}

**Interpretation:**
- SMD values closer to 0 indicate better balance
- After IPTW weighting, SMDs should be reduced, indicating improved balance
- The propensity score itself is included as a check on the propensity model

---

## 5. IPTW (Inverse Probability of Treatment Weighting)

We calculate IPTW weights as:
- **Treated subjects:** w = 1 / propensity_score
- **Control subjects:** w = 1 / (1 - propensity_score)

Average IPTW weights:
- Treated group: **{avg_iptw_treated:.3f}**
- Control group: **{avg_iptw_control:.3f}**

### Weighted Outcome Means

- Weighted mean outcome (treated): **{weighted_mean_outcome_treated:.3f}**
- Weighted mean outcome (control): **{weighted_mean_outcome_control:.3f}**

**IPTW-adjusted effect (treated - control):** **{iptw_effect:.3f}**

This is the treatment effect estimated using IPTW weighting to balance the groups.

---

## 6. Covariate Adjustment Using Propensity Scores

We also fit a linear regression:

**outcome ~ treatment + propensity_score**

- Intercept: **{intercept:.3f}**
- Coefficient on treatment (adjusted effect): **{beta_treat:.3f}**
- Coefficient on propensity score: **{beta_ps:.3f}**

**Interpretation:**

- The **crude effect** shows what happens if we just compare treated vs control.
- The **IPTW-adjusted effect** uses weighting to create a pseudo-population with balanced covariates.
- The **regression-adjusted effect** (coefficient on treatment) estimates the treatment effect
  **after controlling for baseline covariates via the propensity score** in a regression model.

Both methods (IPTW and regression adjustment) should give similar results if the model is correctly specified.

---

## Summary of Treatment Effects

| Method | Treatment Effect |
|--------|------------------|
| Crude (unadjusted) | **{crude_effect:.3f}** |
| IPTW-weighted | **{iptw_effect:.3f}** |
| Regression-adjusted | **{beta_treat:.3f}** |

In a real drug development / RWE setting, you might:
- Use more covariates (labs, performance status, biomarkers)
- Use logistic or survival models for the outcome
- Compute confidence intervals and p-values
- Combine IPTW with regression adjustment (doubly robust estimation)

This app demonstrates **propensity score-based covariate adjustment** and **IPTW weighting**.
"""

    return text


with gr.Blocks() as demo:
    gr.Markdown(
        """
# Propensity Score Covariate Adjustment – Drug Development (Demo)

Upload a CSV file with observational data comparing a **new drug** vs **standard of care**.

### Required columns:
- `treatment` (0 = control, 1 = new drug)  
- `outcome` (0/1 or continuous outcome)  
- `age`  
- `sex` (0/1 or M/F)  
- `baseline_risk_score`  
- `comorbidity_index`  

The app will:
1. Estimate **propensity scores** with logistic regression  
2. Compute the **crude (unadjusted)** treatment effect  
3. Calculate **IPTW (Inverse Probability of Treatment Weighting)** and weighted means
4. Compute **Standardized Mean Differences (SMD)** before vs after adjustment
5. Fit an **outcome model** with outcome ~ treatment + propensity_score  
6. Report **propensity-adjusted treatment effect** and **IPTW-adjusted effect**
"""
    )

    file_input = gr.File(label="Upload CSV")
    run_button = gr.Button("Run Propensity Score Adjustment")
    output_md = gr.Markdown()

    run_button.click(
        propensity_covariate_adjustment,
        inputs=[file_input],
        outputs=[output_md],
    )

if __name__ == "__main__":
    demo.launch(share=True)