app.py

import math
import gradio as gr


def ab_test_analysis(
    n_a,
    resp_a,
    n_b,
    resp_b,
):
    # Basic validation
    try:
        n_a = int(n_a)
        resp_a = int(resp_a)
        n_b = int(n_b)
        resp_b = int(resp_b)
    except (TypeError, ValueError):
        return "❌ Please enter valid integers for all inputs."

    if n_a <= 0 or n_b <= 0:
        return "❌ Group sizes must be positive."
    if not (0 <= resp_a <= n_a and 0 <= resp_b <= n_b):
        return "❌ Responders must be between 0 and group size for each group."

    # Conversion rates
    rate_a = resp_a / n_a
    rate_b = resp_b / n_b

    abs_lift = rate_b - rate_a
    rel_lift = abs_lift / rate_a if rate_a > 0 else float("nan")

    # Two-sample z-test for proportions
    # H0: rate_a == rate_b ; H1: rate_a != rate_b
    pooled = (resp_a + resp_b) / (n_a + n_b)
    se = math.sqrt(pooled * (1 - pooled) * (1 / n_a + 1 / n_b))

    if se == 0:
        z_val = 0.0
        p_val = 1.0
    else:
        z_val = (rate_b - rate_a) / se
        # normal CDF using erf
        cdf = 0.5 * (1 + math.erf(z_val / math.sqrt(2)))
        # two-sided p-value
        p_val = 2 * (1 - cdf) if z_val >= 0 else 2 * cdf

    # Interpretation
    significance = "statistically significant at α = 0.05" if p_val < 0.05 else "NOT statistically significant at α = 0.05"

    summary = f"""
## A/B Test Results – Pharmaceutical Refill Campaign

**Input**
- Group A (Control): **{n_a}** patients, **{resp_a}** refilled
- Group B (Treatment): **{n_b}** patients, **{resp_b}** refilled

**Conversion Rates**
- Group A: **{rate_a:.2%}**
- Group B: **{rate_b:.2%}**

**Lift**
- Absolute lift: **{abs_lift:.2%}** (percentage point difference)
- Relative lift: **{rel_lift:.2%}** vs Group A

**Statistical Test (Two-sample z-test for proportions)**
- z-value: **{z_val:.3f}**
- p-value: **{p_val:.4f}**
- Result: The difference is **{significance}**.

---

### Interpretation

In this example, Group B represents a **new intervention** (for example, a more personalized SMS reminder or additional pharmacist counseling), and Group A is the **current standard approach**.

- If the p-value is below 0.05, you can say that the new strategy (Group B) **meaningfully improves** refill adherence.
- If the p-value is above 0.05, the observed difference could be due to random variation, and you may need **more data** or a **different intervention**.

You can update the inputs above with any A/B test from:
- Different patient education materials
- Different reminder frequency
- Portal vs. SMS outreach
- Pharmacist counseling vs. usual care
"""
    return summary


with gr.Blocks() as demo:
    gr.Markdown(
        """
# Pharmaceutical A/B Testing – Refill Adherence Example

This demo compares two groups of patients:

- **Group A (Control)** – current reminder process  
- **Group B (Treatment)** – new intervention (e.g., personalized SMS, educational content, or pharmacist counseling)

Enter the number of patients and the number who refilled their prescription in each group.
The app will calculate conversion rates, lift, and a simple statistical test.
"""
    )

    with gr.Row():
        with gr.Column():
            gr.Markdown("### Group A – Control")
            n_a = gr.Number(label="Number of patients in Group A", value=2000)
            resp_a = gr.Number(label="Number of refills in Group A", value=520)

        with gr.Column():
            gr.Markdown("### Group B – Treatment")
            n_b = gr.Number(label="Number of patients in Group B", value=1900)
            resp_b = gr.Number(label="Number of refills in Group B", value=608)

    run_button = gr.Button("Run A/B Analysis")

    output = gr.Markdown()

    run_button.click(
        fn=ab_test_analysis,
        inputs=[n_a, resp_a, n_b, resp_b],
        outputs=output,
    )

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