// Still inside runAnalysis() in index.html const r = await fetch('/predict', { method: 'POST', headers: { 'Content-Type': 'application/json' }, body: JSON.stringify({ text }) // sends: { "text": "I feel hopeless..." } }); const d = await r.json(); // parse the JSON response render(d, text); // draw results on screen

# app.py from predict import predict_all @app.route('/predict', methods=['POST']) def predict(): data = request.get_json() text = data['text'].strip() # validation if not text: return jsonify({'error': 'Text cannot be empty'}), 400 if len(text) > 5000: return jsonify({'error': 'Too long'}), 400 t0 = time.time() result = predict_all(text) # ← the big function (next steps) result['processing_time_ms'] = round((time.time() - t0) * 1000) return jsonify(result) # sends JSON back to browser

# predict.py def clean_text(text): text = str(text).lower() # UPPERCASE → lowercase text = re.sub(r'http\S+|www\S+', '', text) # remove URLs text = re.sub(r'@\w+', '', text) # remove @mentions text = re.sub(r'#', '', text) # remove # (keep word) text = text.translate(str.maketrans('','',punctuation)) # remove !.,?etc text = re.sub(r'\s+', ' ', text).strip() # collapse spaces return text # Example: # IN: "@user I've been SO depressed https://t.co #mentalhealth 😢" # OUT: "ive been so depressed mentalhealth"

# predict.py — called 3× (once per dataset) def predict_classical(text_clean, ds): tfidf = _models[f'tfidf_{ds}'] le = _models[f'le_{ds}'] vec = tfidf.transform([text_clean]) # text → 50K-dim vector for model_name in ['logistic_regression', 'svm', 'xgboost']: model = _models[f'{model_name}_{ds}'] pred_idx = model.predict(vec)[0] # → e.g. 4 label = le.classes_[pred_idx] # 4 → "postpartum" # LR / XGBoost: direct probability if hasattr(model, 'predict_proba'): conf = model.predict_proba(vec)[0][pred_idx] # SVM: no predict_proba → use softmax of decision scores elif hasattr(model, 'decision_function'): scores = model.decision_function(vec)[0] e = np.exp(scores - scores.max()) conf = e[pred_idx] / e.sum() # normalise to 0–1

# predict.py — called 3× (once per dataset) def predict_xlmr(raw_text, xlmr_model, le, max_len=128): inputs = tokenizer( raw_text, return_tensors='pt', # PyTorch tensors max_length=max_len, # 128 for tweets, 256 for Reddit truncation=True, padding='max_length' ).to(device) # send to GPU if available with torch.no_grad(): # no_grad saves memory (not training) logits = xlmr_model(**inputs).logits probs = torch.softmax(logits, dim=1)[0] # → [0.91, 0.04, 0.02, ...] pred_idx = int(probs.argmax()) # index of highest label = le.classes_[pred_idx] all_probs = {le.classes_[i]: float(p) for i, p in enumerate(probs)} # all_probs = {"postpartum":0.913, "bipolar":0.041, ...} # only D1 uses this for the breakdown chart return {'label': label, 'confidence': float(probs[pred_idx]), 'all_probs': all_probs}

# predict.py — the main function Flask calls def predict_all(raw_text): clean = clean_text(raw_text) # Run all 4 models per dataset d1 = predict_classical(clean, 'd1') # → {LR:{}, SVM:{}, XGB:{}} d1['XLM-RoBERTa'] = predict_xlmr(raw_text, xlmr1, le1, 128) # same for d2, d3... # Winner = model with highest confidence d1_winner = max(d1.items(), key=lambda x: x[1]['confidence']) # → ('XGBoost', {'label': 'postpartum', 'confidence': 0.999}) # Suicide risk = majority vote across 4 D3 models suicide_count = sum( 1 for r in d3.values() if 'suicide' in r['label'] and 'non' not in r['label'] ) risk_flag = suicide_count >= 3 # ≥3 of 4 models → HIGH RISK return { 'dataset1': {'models': d1, 'winner_model': d1_winner[0], ...}, 'dataset2': {...}, 'dataset3': {...}, 'risk_flag': risk_flag, 'suicide_votes': f'{suicide_count}/4 models flagged' }

{ "dataset1": { "task": "Depression Type (6 Classes)", "models": { "Logistic Regression": { "label": "postpartum", "confidence": 0.958 }, "SVM": { "label": "postpartum", "confidence": 0.828 }, "XGBoost": { "label": "postpartum", "confidence": 0.999 }, "XLM-RoBERTa": { "label": "postpartum", "confidence": 0.997 } }, "winner_model": "XGBoost", "winner_prediction": "postpartum", "winner_confidence": 0.999, "class_probs": { "postpartum": 0.997, "bipolar": 0.001, ... } }, "dataset2": { ... }, "dataset3": { ... }, "risk_flag": false, "suicide_votes": "0/4 models flagged suicide risk", "processing_time_ms": 2341 }

// index.html — called after fetch() returns function render(d, text) { // 1. Fill winner cards document.getElementById('wpA').textContent = d.dataset1.winner_prediction; document.getElementById('wcA').textContent = (d.dataset1.winner_confidence * 100).toFixed(1) + '%'; // 2. Build per-model rows for each dataset buildPanel('p1', d.dataset1.models, d.dataset1.winner_model); buildPanel('p2', d.dataset2.models, d.dataset2.winner_model); buildPanel('p3', d.dataset3.models, d.dataset3.winner_model); // 3. Risk banner if (d.risk_flag) { riskBanner.className = 'risk-banner danger'; } else { riskBanner.className = 'risk-banner safe'; } // 4. Show results section document.getElementById('results').style.display = 'block'; } function buildPanel(panelId, models, winner) { let html = ''; Object.entries(models).forEach(([name, res]) => { html += `<div class="mr ${name===winner?'winner':''}"> <div class="mr-name">${name}</div> <div class="mr-pred">${res.label}</div> <div class="mr-fill" data-w="${(res.confidence*100).toFixed(1)}"></div> <div class="mr-pct">${(res.confidence*100).toFixed(1)}%</div> </div>`; }); panel.innerHTML = html; // inject HTML // bars animate next step ↓ }

/* CSS — transition defined in <style> */ .mr-fill { width: 0%; /* starts invisible */ transition: width 0.8s cubic-bezier(.4,0,.2,1); /* smooth ease-out */ } .mr.winner .mr-fill { background: var(--purple); } /* winner = purple */ // JavaScript — in buildPanel() setTimeout(() => { panel.querySelectorAll('.mr-fill').forEach(el => { el.style.width = el.getAttribute('data-w') + '%'; // sets e.g. "82.8%" → CSS transition plays automatically }); }, 80); // 80ms wait for DOM to paint first // The 6-class breakdown bars work the same way // but use 200ms delay and .cp-fill class