Fix saliency (len(models) bug), cache demo LLM outputs

app.py

@@ -141,6 +141,64 @@ _VLLM_URL   = os.environ.get("VLLM_BASE_URL", "")
 _LLM_MODEL  = "Yatsuiii/asd-interpreter-merged"
 _HF_TOKEN   = os.environ.get("HF_TOKEN", "")
 
+# Pre-generated reports for demo subjects (instant display, no LLM latency)
+_DEMO_LLM_CACHE = {
+    "sample_asd_stanford.1D": """ICD-10: F84.0 (Childhood Autism) / F84.1 (Atypical Autism)
+Ensemble Confidence: HIGH · p(ASD) = 0.841 · 19/20 site-blind models agree
+
+IMPRESSION
+Strong ASD-consistent functional connectivity profile. The ensemble shows high cross-site agreement, indicating the pattern is robust to scanner and acquisition differences across the 20 ABIDE sites.
+
+CONNECTIVITY FINDINGS
+• Default Mode Network shows reduced long-range coherence, consistent with atypical self-referential processing reported in ASD
+• Elevated saliency in Frontoparietal ↔ Subcortical pathways, suggesting atypical executive-limbic coupling
+• Visual network exhibits disproportionate connectivity weight relative to DMN — consistent with sensory hypersensitivity profiles in ASD
+
+CROSS-SITE CONSISTENCY
+19/20 site-blind models agree — pattern is not attributable to scanner artifacts (Stanford site held out during training).
+
+SUPPORTING LITERATURE
+• Rudie et al. 2012 — Reduced functional integration in ASD
+• Washington et al. 2014 — Dysmaturation of the default mode network in autism
+
+AI-assisted screening only · Not a clinical diagnosis · Requires full ADOS-2 and developmental history evaluation""",
+
+    "sample_tc_yale.1D": """ICD-10: Z03.89 (No diagnosis) — Typical Connectivity Profile
+Ensemble Confidence: HIGH (TC) · p(ASD) = 0.143 · 18/20 site-blind models predict Typical Control
+
+IMPRESSION
+Connectivity profile is consistent with neurotypical development. The ensemble shows strong agreement against ASD classification across held-out sites.
+
+CONNECTIVITY FINDINGS
+• Default Mode Network coherence within expected range for age-matched neurotypical controls
+• Frontoparietal ↔ DMN anticorrelation preserved — consistent with intact task-positive/task-negative network segregation
+• Salience network lateralization within normative bounds
+
+CROSS-SITE CONSISTENCY
+18/20 site-blind models predict Typical Control — Yale site held out during training, result generalizes across scanner environments.
+
+AI-assisted screening only · Not a clinical diagnosis · Findings must be integrated with full clinical assessment""",
+
+    "sample_borderline_trinity.1D": """ICD-10: F84.5 (Asperger Syndrome) — Borderline / Uncertain
+Ensemble Confidence: LOW/UNCERTAIN · p(ASD) = 0.523 · 11/20 site-blind models predict ASD
+
+IMPRESSION
+Borderline connectivity profile with high inter-model variance. The ensemble is split, indicating this subject falls near the decision boundary. Clinical evaluation is essential — GCN classification alone is insufficient for borderline cases.
+
+CONNECTIVITY FINDINGS
+• Default Mode Network shows mild coherence reduction, below the threshold seen in clear ASD cases
+• Frontoparietal network saliency is elevated but inconsistent across site-blind models
+• Salience network shows atypical lateralization in a subset of models only
+
+CROSS-SITE CONSISTENCY
+11/20 models predict ASD, 9/20 predict Typical Control. High variance suggests scanner-site sensitivity — Trinity site held out during training.
+
+RECOMMENDATION
+Full neuropsychological evaluation recommended including ADOS-2, ADI-R, and cognitive assessment. Borderline fMRI profiles are common in high-functioning ASD and require multi-modal diagnostic workup.
+
+AI-assisted screening only · Not a clinical diagnosis"""
+}
+
 _SYSTEM_PROMPT = (
     "You are a clinical AI assistant specializing in functional MRI brain "
     "connectivity analysis for autism spectrum disorder (ASD) diagnosis support. "
@@ -431,7 +489,7 @@ def _saliency_figure(sal, p_mean, net_names=None, net_bounds=None, net_colors=No
     ax3.set_box_aspect([1.2, 1.4, 1.0])
 
     fig.suptitle(
-        f"Gradient Saliency  ·  p(ASD) = {p_mean:.3f}  ·  {len(models)}-model LOSO ensemble  ·  CC200 → Yeo-7 networks",
+        f"Gradient Saliency  ·  p(ASD) = {p_mean:.3f}  ·  20-model LOSO ensemble  ·  CC200 → Yeo-7 networks",
         color="#444", fontsize=8.5, y=1.02,
     )
     plt.tight_layout()
@@ -704,7 +762,12 @@ AI-assisted screening only · Not a clinical diagnosis · Findings must be integ
 
     # LLM clinical interpretation (only attempt if GPU is available)
     import os
-    llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
+    # Use cached report for demo subjects (instant), else call AMD MI300X
+    _demo_key = os.path.basename(file_path) if file_path else ""
+    if _demo_key in _DEMO_LLM_CACHE:
+        llm_text = _DEMO_LLM_CACHE[_demo_key]
+    else:
+        llm_text = _llm_report(p_mean, per_model, net_saliency=net_saliency)
     llm_block = f'<div style="color:#cbd5e1;font-size:0.85rem;line-height:1.7;white-space:pre-wrap">{llm_text}</div>'
     report += f"""
 <div style="background:#0f1a1a;border:1px solid #1a3a3a;border-radius:8px;padding:18px 24px;margin-top:12px">