5x LLM improvements: saliency grounding, anti-hallucination system prompt, temp 0.1, n_subjects, per-network scores in prompt

app.py

@@ -141,10 +141,15 @@ _SYSTEM_PROMPT = (
     "You are a clinical AI assistant specializing in functional MRI brain "
     "connectivity analysis for autism spectrum disorder (ASD) diagnosis support. "
     "You interpret outputs from a validated graph neural network (GCN) trained on "
-    "the ABIDE I dataset and provide structured clinical summaries for neurologists "
-    "and psychiatrists. Your reports are informative and evidence-based but always "
-    "clarify that findings are AI-assisted and should be integrated with full "
-    "clinical assessment. You do not make a diagnosis."
+    "the ABIDE I dataset (1,102 subjects, 20 acquisition sites) and provide structured "
+    "clinical summaries for neurologists and psychiatrists. "
+    "CRITICAL RULES: (1) Only reference brain networks, connectivity patterns, and "
+    "statistics that are explicitly provided in the input report — do NOT invent or "
+    "hallucinate network names, connectivity findings, or numeric values. "
+    "(2) Base every observation directly on the per-network saliency scores and "
+    "ensemble probabilities given in the input. (3) If a network is not listed in the "
+    "input, do not mention it. (4) Always clarify findings are AI-assisted and require "
+    "full clinical assessment. You do not make a diagnosis."
 )
 _llm_cache = None
 
@@ -162,7 +167,7 @@ def get_llm():
     _llm_cache = (mdl, tok)
     return _llm_cache
 
-def _llm_report(p_mean: float, per_model: list) -> str:
+def _llm_report(p_mean: float, per_model: list, net_saliency: dict | None = None) -> str:
     consensus = sum(1 for _, p in per_model if p > 0.5)
     per_model_str = "\n".join(
         f"  {s}-blind: {'ASD' if v > 0.5 else 'TC'} (p={v:.3f})" for s, v in per_model
@@ -173,13 +178,29 @@ def _llm_report(p_mean: float, per_model: list) -> str:
         "LOW / UNCERTAIN" if p_mean >= 0.4 else
         "MODERATE (TC)" if p_mean >= 0.25 else "HIGH (TC)"
     )
+
+    sal_section = ""
+    if net_saliency:
+        sorted_nets = sorted(net_saliency.items(), key=lambda x: x[1], reverse=True)
+        sal_lines = "\n".join(
+            f"  {name}: {score:.5f}" for name, score in sorted_nets
+        )
+        sal_section = (
+            f"\nPer-Network Gradient Saliency (ranked high→low, actual GCN values):\n"
+            f"{sal_lines}\n"
+            f"[ONLY reference these networks with these exact values — no others.]\n"
+        )
+
     user_msg = (
         f"Brain Connectivity GCN Analysis Report\n{'='*40}\n"
-        f"p(ASD)           : {p_mean:.3f}\n"
-        f"Confidence Level : {conf_label}\n"
-        f"Model Consensus  : {consensus}/{len(per_model)} site-blind models predict ASD\n\n"
+        f"Dataset           : ABIDE I · 1,102 subjects · 20 acquisition sites\n"
+        f"p(ASD)            : {p_mean:.3f}\n"
+        f"Confidence Level  : {conf_label}\n"
+        f"Model Consensus   : {consensus}/{len(per_model)} site-blind models predict ASD\n"
+        f"{sal_section}\n"
         f"Per-Model Breakdown (LOSO ensemble):\n{per_model_str}\n\n"
-        f"Please provide a structured clinical interpretation of these findings."
+        f"Provide a structured clinical interpretation referencing ONLY the networks "
+        f"and values listed above. Do not mention any network not in this report."
     )
     try:
         mdl, tok = get_llm()
@@ -191,7 +212,7 @@ def _llm_report(p_mean: float, per_model: list) -> str:
         inputs = tok(text, return_tensors="pt").to(next(mdl.parameters()).device)
         with torch.no_grad():
             out = mdl.generate(
-                **inputs, max_new_tokens=512, temperature=0.3,
+                **inputs, max_new_tokens=512, temperature=0.1,
                 do_sample=True, pad_token_id=tok.eos_token_id,
             )
         generated = out[0][inputs["input_ids"].shape[1]:]
@@ -473,11 +494,21 @@ def run_gcn(file_path):
     consensus = sum(1 for _, p in per_model if p > 0.5)
     conf      = max(p_mean, 1 - p_mean) * 100
 
+    net_saliency = None
     try:
+        sal = _compute_saliency(bw_t, adj_t, models)
+        net_names  = atlas_cfg["net_names"]
+        net_bounds = atlas_cfg["net_bounds"]
+        # aggregate ROI-level saliency to network-level importance scores
+        net_imp = np.array([
+            sal[s:e, :].mean() + sal[:, s:e].mean()
+            for s, e in zip(net_bounds[:-1], net_bounds[1:])
+        ])
+        net_saliency = dict(zip(net_names, net_imp.tolist()))
         sal_img = _saliency_figure(
-            _compute_saliency(bw_t, adj_t, models), p_mean,
-            net_names=atlas_cfg["net_names"],
-            net_bounds=atlas_cfg["net_bounds"],
+            sal, p_mean,
+            net_names=net_names,
+            net_bounds=net_bounds,
             net_colors=atlas_cfg["net_colors"],
         )
     except Exception:
@@ -596,7 +627,7 @@ LOSO AUC = 0.7260 · 1,102 held-out subjects · 20 acquisition sites
 AI-assisted screening only · Not a clinical diagnosis · Findings must be integrated with ADOS-2, ADI-R, and full developmental history · Refer to licensed neuropsychologist for formal evaluation.</div></div>"""
 
     # LLM clinical interpretation
-    llm_text = _llm_report(p_mean, per_model)
+    llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
     report += f"""
 <div style="background:#0f1a1a;border:1px solid #1a3a3a;border-radius:8px;padding:18px 24px;margin-top:12px">
 <div style="color:#2dc653;font-size:0.68rem;text-transform:uppercase;letter-spacing:1.5px;margin-bottom:10px;font-weight:600">