Dina-Raslan commited on
Commit
410c3e5
Β·
verified Β·
1 Parent(s): a2a3a9e

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +31 -7
app.py CHANGED
@@ -251,9 +251,14 @@ def prepare_clinical_features(
251
  def get_top_shap_features(clinical_feat, top_n=6):
252
  """
253
  Compute SHAP values for the clinical (GB) model's prediction and
254
- return the top_n features ranked by absolute impact.
 
 
255
  clinical_feat must be the scaled numpy array in ALL_COLS order.
256
- Returns a list of (feature_name, shap_value) tuples.
 
 
 
257
  """
258
  shap_values = shap_explainer.shap_values(clinical_feat)
259
 
@@ -263,8 +268,26 @@ def get_top_shap_features(clinical_feat, top_n=6):
263
  else:
264
  shap_row = shap_values[0]
265
 
266
- top_idx = np.argsort(np.abs(shap_row))[::-1][:top_n]
267
- return [(ALL_COLS[i], float(shap_row[i])) for i in top_idx]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
268
 
269
  # ── Grad-CAM++ ────────────────────────────────────────────────────────────────
270
  class GradCAMPlusPlus:
@@ -501,8 +524,8 @@ def predict_ckd_risk(
501
  # 5. GB clinical prediction
502
  clinical_prob = gb_model.predict_proba(clinical_feat)[0, 1]
503
 
504
- # 5.b Top-6 SHAP features driving the clinical prediction
505
- top_shap_features = get_top_shap_features(clinical_feat, top_n=6)
506
 
507
  # 6. Fusion model prediction
508
  retinal_tensor = torch.tensor(retinal_feat_pca, dtype=torch.float32)
@@ -592,7 +615,8 @@ def predict_ckd_risk(
592
  Risk Level: {risk_label}
593
  ╠══════════════════════════════════════╣
594
  TOP 6 CLINICAL FEATURES (SHAP)
595
- {chr(10).join(f" {i+1}. {name:<8s} {'+' if val >= 0 else ''}{val:.4f}" for i, (name, val) in enumerate(top_shap_features))}
 
596
  β•šβ•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•
597
  """
598
 
 
251
  def get_top_shap_features(clinical_feat, top_n=6):
252
  """
253
  Compute SHAP values for the clinical (GB) model's prediction and
254
+ return the top_n features ranked by absolute impact, together with
255
+ each feature's percentage share of total |SHAP| impact and the
256
+ combined percentage share of all remaining features.
257
  clinical_feat must be the scaled numpy array in ALL_COLS order.
258
+
259
+ Returns:
260
+ top_features: list of (feature_name, shap_value, pct_impact) tuples
261
+ other_pct: combined percentage impact of the remaining features
262
  """
263
  shap_values = shap_explainer.shap_values(clinical_feat)
264
 
 
268
  else:
269
  shap_row = shap_values[0]
270
 
271
+ abs_vals = np.abs(shap_row)
272
+ total_abs = abs_vals.sum()
273
+
274
+ # Avoid division by zero in the (unlikely) case all SHAP values are 0
275
+ if total_abs == 0:
276
+ pct_vals = np.zeros_like(abs_vals)
277
+ else:
278
+ pct_vals = (abs_vals / total_abs) * 100.0
279
+
280
+ sorted_idx = np.argsort(abs_vals)[::-1]
281
+ top_idx = sorted_idx[:top_n]
282
+ other_idx = sorted_idx[top_n:]
283
+
284
+ top_features = [
285
+ (ALL_COLS[i], float(shap_row[i]), float(pct_vals[i]))
286
+ for i in top_idx
287
+ ]
288
+ other_pct = float(pct_vals[other_idx].sum()) if len(other_idx) else 0.0
289
+
290
+ return top_features, other_pct
291
 
292
  # ── Grad-CAM++ ────────────────────────────────────────────────────────────────
293
  class GradCAMPlusPlus:
 
524
  # 5. GB clinical prediction
525
  clinical_prob = gb_model.predict_proba(clinical_feat)[0, 1]
526
 
527
+ # 5.b Top-6 SHAP features driving the clinical prediction (with % impact)
528
+ top_shap_features, other_shap_pct = get_top_shap_features(clinical_feat, top_n=6)
529
 
530
  # 6. Fusion model prediction
531
  retinal_tensor = torch.tensor(retinal_feat_pca, dtype=torch.float32)
 
615
  Risk Level: {risk_label}
616
  ╠══════════════════════════════════════╣
617
  TOP 6 CLINICAL FEATURES (SHAP)
618
+ {chr(10).join(f" {i+1}. {name:<8s} {'+' if val >= 0 else ''}{val:.4f} ({pct:.1f}% impact)" for i, (name, val, pct) in enumerate(top_shap_features))}
619
+ Other 18 factors combined: {other_shap_pct:.1f}% impact
620
  β•šβ•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•β•
621
  """
622