Upload 2 files

app.py

@@ -0,0 +1,292 @@
+import gradio as gr
+import shap
+import numpy as np
+import torch
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
+
+# ── Model Setup ──────────────────────────────────────────────────────────────
+MODEL_ID = "rayshunp/ADR2026Team5"
+
+tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+model     = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
+model.eval()
+
+# Hugging-Face pipeline used by SHAP
+clf_pipeline = pipeline(
+    "text-classification",
+    model=model,
+    tokenizer=tokenizer,
+    return_all_scores=True,   # needed so SHAP sees both class probabilities
+    device=0 if torch.cuda.is_available() else -1,
+)
+
+# Label map – adjust if your model uses different label names
+LABELS = {0: "Non-Severe", 1: "Severe"}
+
+# SHAP explainer (uses the HF pipeline directly)
+explainer = shap.Explainer(clf_pipeline)
+
+# --- Medical Information Extraction Model (NER) ---
+# Token classification model for extracting medical entities
+# (e.g. drugs, symptoms, body sites, dosages) from ADR text.
+# Swap NER_MODEL_ID for any token-classification model on the Hub,
+# for example:
+#   "allenai/scibert_scivocab_cased"           – general biomedical
+#   "d4data/biomedical-ner-all"                – multi-entity biomedical NER
+#   "pruas/BENT-PubMedBERT-NER-Disease"        – disease/symptom focused
+
+NER_MODEL_ID = "d4data/biomedical-ner-all"   # ← replace with your chosen model
+
+ner_pipe = pipeline(
+    "ner",
+    model=NER_MODEL_ID,
+    aggregation_strategy="simple",   # merges subword tokens → whole words
+    device=0 if torch.cuda.is_available() else -1,
+)
+
+
+# ── Core prediction function ─────────────────────────────────────────────────
+def predict(text: str):
+    if not text.strip():
+        return "⚠️ Please enter a reaction description.", None, None
+
+    # ── 1. Severity prediction ──
+    results   = clf_pipeline(text)[0]           # list of {label, score}
+    scores    = {r["label"]: r["score"] for r in results}
+
+    # Normalise to expected label keys
+    # The model may emit "LABEL_0"/"LABEL_1" or "Non-Severe"/"Severe"
+    def get_score(idx):
+        for key in (LABELS[idx], f"LABEL_{idx}"):
+            if key in scores:
+                return scores[key]
+        # fallback: pick by position
+        return results[idx]["score"]
+
+    score_non_severe = get_score(0)
+    score_severe     = get_score(1)
+    predicted_idx    = int(score_severe > score_non_severe)
+    label            = LABELS[predicted_idx]
+    confidence       = score_severe if predicted_idx == 1 else score_non_severe
+
+    verdict_html = f"""
+    <div style="
+        font-family: 'Courier New', monospace;
+        border: 2px solid {'#e74c3c' if predicted_idx == 1 else '#2ecc71'};
+        border-radius: 10px;
+        padding: 18px 24px;
+        background: {'#2c0a0a' if predicted_idx == 1 else '#0a2c12'};
+        color: {'#ff6b6b' if predicted_idx == 1 else '#6bffaa'};
+        text-align: center;
+    ">
+        <div style="font-size:2rem; font-weight:900; letter-spacing:2px;">
+            {'🚨 SEVERE' if predicted_idx == 1 else '✅ NON-SEVERE'}
+        </div>
+        <div style="font-size:1rem; margin-top:8px; opacity:0.85;">
+            Confidence: <strong>{confidence:.1%}</strong>
+        </div>
+        <div style="margin-top:12px; display:flex; gap:8px; justify-content:center; flex-wrap:wrap;">
+            <span style="background:#333; border-radius:6px; padding:4px 12px;">
+                Non-Severe: {score_non_severe:.1%}
+            </span>
+            <span style="background:#333; border-radius:6px; padding:4px 12px;">
+                Severe: {score_severe:.1%}
+            </span>
+        </div>
+    </div>
+    """
+
+    # ── 2. SHAP values ──
+    shap_values = explainer([text])
+
+    # shap_values.values shape: (1, n_tokens, n_classes)
+    # We want the SHAP values for the SEVERE class (index 1)
+    tokens      = shap_values.data[0]          # list of token strings
+    vals_severe = shap_values.values[0, :, 1]  # shape (n_tokens,)
+
+    # ── 3. Inline word-highlight HTML ──
+    word_html = _build_highlight_html(tokens, vals_severe)
+
+    # ── 4. SHAP bar/force plot image ──
+    fig = _build_shap_plot(tokens, vals_severe, text)
+
+    return verdict_html, word_html, fig
+
+
+# ── HTML word highlights ──────────────────────────────────────────────────────
+def _build_highlight_html(tokens, values):
+    max_abs = max(abs(values).max(), 1e-8)
+
+    parts = []
+    for token, val in zip(tokens, values):
+        # Skip special tokens
+        if token in ("[CLS]", "[SEP]", "<s>", "</s>", "<pad>"):
+            continue
+
+        intensity = abs(val) / max_abs          # 0-1
+        alpha     = 0.15 + 0.75 * intensity     # 0.15-0.90
+
+        if val > 0:                             # pushes toward SEVERE → red
+            bg = f"rgba(231, 76, 60, {alpha:.2f})"
+            fg = "#fff" if intensity > 0.4 else "#111"
+        else:                                   # pushes toward NON-SEVERE → green
+            bg = f"rgba(46, 204, 113, {alpha:.2f})"
+            fg = "#fff" if intensity > 0.4 else "#111"
+
+        # Strip leading ## from subword tokens for readability
+        display = token.lstrip("#")
+
+        parts.append(
+            f'<span style="background:{bg}; color:{fg}; '
+            f'padding:3px 5px; border-radius:4px; margin:2px; '
+            f'font-weight:{"700" if intensity > 0.5 else "400"}; '
+            f'display:inline-block;" '
+            f'title="SHAP: {val:+.4f}">{display}</span>'
+        )
+
+    html = f"""
+    <div style="
+        font-family: Georgia, serif;
+        font-size: 1.05rem;
+        line-height: 2.2;
+        padding: 16px;
+        background: #1a1a2e;
+        border-radius: 10px;
+        border: 1px solid #333;
+        color: #eee;
+    ">
+        <div style="font-size:0.78rem; color:#aaa; margin-bottom:10px;">
+            🔴 Red = pushes toward <strong>Severe</strong> &nbsp;|&nbsp;
+            🟢 Green = pushes toward <strong>Non-Severe</strong> &nbsp;|&nbsp;
+            Darker = higher importance
+        </div>
+        {''.join(parts)}
+    </div>
+    """
+    return html
+
+
+# ── SHAP bar plot ─────────────────────────────────────────────────────────────
+def _build_shap_plot(tokens, values, text):
+    # Filter special tokens
+    pairs = [
+        (t.lstrip("#"), v)
+        for t, v in zip(tokens, values)
+        if t not in ("[CLS]", "[SEP]", "<s>", "</s>", "<pad>")
+    ]
+    if not pairs:
+        return None
+
+    labels_plot, vals = zip(*pairs)
+
+    # Sort by absolute value descending, take top 20
+    order  = np.argsort(np.abs(vals))[::-1][:20]
+    labels_plot = [labels_plot[i] for i in order]
+    vals   = [vals[i] for i in order]
+
+    # Re-sort for display: positive first, then negative
+    combined = sorted(zip(vals, labels_plot), key=lambda x: x[0])
+    vals, labels_plot = zip(*combined)
+
+    colors = ["#e74c3c" if v > 0 else "#2ecc71" for v in vals]
+
+    fig, ax = plt.subplots(figsize=(8, max(4, len(vals) * 0.38)))
+    fig.patch.set_facecolor("#1a1a2e")
+    ax.set_facecolor("#1a1a2e")
+
+    bars = ax.barh(labels_plot, vals, color=colors, edgecolor="none", height=0.65)
+
+    ax.axvline(0, color="#555", linewidth=1)
+    ax.set_xlabel("SHAP value (impact on Severe class)", color="#ccc", fontsize=10)
+    ax.set_title(
+        f"Token SHAP Breakdown\n\"{text[:60]}{'…' if len(text)>60 else ''}\"",
+        color="#eee", fontsize=11, pad=12
+    )
+    ax.tick_params(colors="#ccc")
+    for spine in ax.spines.values():
+        spine.set_edgecolor("#444")
+
+    # Value labels on bars
+    for bar, val in zip(bars, vals):
+        ax.text(
+            val + (0.002 if val >= 0 else -0.002),
+            bar.get_y() + bar.get_height() / 2,
+            f"{val:+.3f}",
+            va="center",
+            ha="left" if val >= 0 else "right",
+            color="#eee",
+            fontsize=8,
+        )
+
+    plt.tight_layout()
+    return fig
+
+
+# ── Gradio UI ─────────────────────────────────────────────────────────────────
+EXAMPLES = [
+    ["I felt groggy and dizzy after taking Tylenol."],
+    ["I had a mild headache after the medication."],
+    ["I experienced severe chest pain and difficulty breathing after the injection."],
+    ["My stomach felt slightly upset after the pill."],
+    ["I had an anaphylactic reaction with hives and throat swelling after penicillin."],
+]
+
+CSS = """
+body { background: #0d0d1a !important; }
+#title {
+    text-align: center;
+    font-family: 'Courier New', monospace;
+    color: #7eb8f7;
+    letter-spacing: 3px;
+    text-transform: uppercase;
+    margin-bottom: 4px;
+}
+#subtitle {
+    text-align: center;
+    color: #888;
+    font-size: 0.9rem;
+    margin-bottom: 20px;
+    font-family: Georgia, serif;
+}
+.gr-button-primary {
+    background: #3a5fc8 !important;
+    border: none !important;
+    font-weight: 700 !important;
+    letter-spacing: 1px !important;
+}
+"""
+
+with gr.Blocks(css=CSS, theme=gr.themes.Base()) as demo:
+    gr.HTML('<h1 id="title">⚕ ADR Severity Analyzer</h1>')
+    gr.HTML('<p id="subtitle">Adverse Drug Reaction · Severity Classification · SHAP Explainability</p>')
+
+    with gr.Row():
+        with gr.Column(scale=1):
+            text_input = gr.Textbox(
+                label="Describe your reaction",
+                placeholder="e.g. I felt groggy and dizzy after taking Tylenol.",
+                lines=3,
+            )
+            analyze_btn = gr.Button("🔍 Analyze Reaction", variant="primary")
+            gr.Examples(examples=EXAMPLES, inputs=text_input, label="Try an example")
+
+        with gr.Column(scale=1):
+            verdict_out = gr.HTML(label="Severity Verdict")
+
+    gr.Markdown("### 🎨 Word-Level SHAP Highlights")
+    highlight_out = gr.HTML(label="Token Highlights")
+
+    gr.Markdown("### 📊 SHAP Token Importance Plot")
+    plot_out = gr.Plot(label="SHAP Bar Chart")
+
+    analyze_btn.click(
+        fn=predict,
+        inputs=text_input,
+        outputs=[verdict_out, highlight_out, plot_out],
+    )
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,6 @@
+gradio>=4.0.0
+transformers>=4.38.0
+torch>=2.0.0
+shap>=0.44.0
+matplotlib>=3.7.0
+numpy>=1.24.0