Update app.py

app.py

@@ -139,66 +139,75 @@ def log_system_usage(tag=""):
 
 def clean_response(text: str) -> str:
     text = sanitize_utf8(text)
+    # Remove unwanted patterns and tool call artifacts
     text = re.sub(r"\[.*?\]|\bNone\b|To analyze the patient record excerpt.*?medications\.|Since the previous attempts.*?\.|I need to.*?medications\.|Retrieving tools.*?\.", "", text, flags=re.DOTALL)
-    text = re.sub(r"\n{3,}", "\n\n", text)
-    text = re.sub(r"[^\n#\-\*\w\s\.\,\:\(\)]+", "", text)
-
-    sections = {}
-    current_section = None
+    # Extract only missed diagnoses, ignoring other categories
+    diagnoses = []
     lines = text.splitlines()
+    in_diagnoses_section = False
     for line in lines:
         line = line.strip()
         if not line:
             continue
-        section_match = re.match(r"###\s*(Missed Diagnoses|Medication Conflicts|Incomplete Assessments|Urgent Follow-up)", line)
-        if section_match:
-            current_section = section_match.group(1)
-            if current_section not in sections:
-                sections[current_section] = []
+        if re.match(r"###\s*Missed Diagnoses", line):
+            in_diagnoses_section = True
             continue
-        finding_match = re.match(r"-\s*.+", line)
-        if finding_match and current_section and not re.match(r"-\s*No issues identified", line):
-            sections[current_section].append(line)
-
-    cleaned = []
-    for heading, findings in sections.items():
-        if findings:
-            cleaned.append(f"### {heading}\n" + "\n".join(findings))
-
-    text = "\n\n".join(cleaned).strip()
+        if re.match(r"###\s*(Medication Conflicts|Incomplete Assessments|Urgent Follow-up)", line):
+            in_diagnoses_section = False
+            continue
+        if in_diagnoses_section and re.match(r"-\s*.+", line):
+            diagnosis = re.sub(r"^\-\s*", "", line).strip()
+            if diagnosis and not re.match(r"No issues identified", diagnosis, re.IGNORECASE):
+                diagnoses.append(diagnosis)
+    # Join diagnoses into a plain text paragraph
+    text = " ".join(diagnoses)
+    # Clean up extra whitespace and punctuation
+    text = re.sub(r"\s+", " ", text).strip()
+    text = re.sub(r"[^\w\s\.\,\(\)\-]", "", text)
     return text if text else ""
 
 def summarize_findings(combined_response: str) -> str:
-    if not combined_response or all("No oversights identified" in chunk for chunk in combined_response.split("--- Analysis for Chunk")):
-        return "### Summary of Clinical Oversights\nNo critical oversights identified in the provided records."
-
-    sections = {}
-    lines = combined_response.splitlines()
-    current_section = None
-    for line in lines:
-        line = line.strip()
-        if not line:
-            continue
-        section_match = re.match(r"###\s*(Missed Diagnoses|Medication Conflicts|Incomplete Assessments|Urgent Follow-up)", line)
-        if section_match:
-            current_section = section_match.group(1)
-            if current_section not in sections:
-                sections[current_section] = []
+    # Split response by chunk analyses
+    chunks = combined_response.split("--- Analysis for Chunk")
+    diagnoses = []
+    for chunk in chunks:
+        chunk = chunk.strip()
+        if not chunk or "No oversights identified" in chunk:
             continue
-        finding_match = re.match(r"-\s*(.+)", line)
-        if finding_match and current_section:
-            sections[current_section].append(finding_match.group(1))
-
-    summary_lines = []
-    for heading, findings in sections.items():
-        if findings:
-            summary = f"- **{heading}**: {'; '.join(findings[:2])}. Risks: {heading.lower()} may lead to adverse outcomes. Recommend: urgent review and specialist referral."
-            summary_lines.append(summary)
-
-    if not summary_lines:
-        return "### Summary of Clinical Oversights\nNo critical oversights identified."
-
-    return "### Summary of Clinical Oversights\n" + "\n".join(summary_lines)
+        # Extract missed diagnoses from chunk
+        lines = chunk.splitlines()
+        in_diagnoses_section = False
+        for line in lines:
+            line = line.strip()
+            if not line:
+                continue
+            if re.match(r"###\s*Missed Diagnoses", line):
+                in_diagnoses_section = True
+                continue
+            if re.match(r"###\s*(Medication Conflicts|Incomplete Assessments|Urgent Follow-up)", line):
+                in_diagnoses_section = False
+                continue
+            if in_diagnoses_section and re.match(r"-\s*.+", line):
+                diagnosis = re.sub(r"^\-\s*", "", line).strip()
+                if diagnosis and not re.match(r"No issues identified", diagnosis, re.IGNORECASE):
+                    diagnoses.append(diagnosis)
+
+    # Remove duplicates while preserving order
+    seen = set()
+    unique_diagnoses = [d for d in diagnoses if not (d in seen or seen.add(d))]
+    
+    if not unique_diagnoses:
+        return "No missed diagnoses were identified in the provided records."
+
+    # Combine into a single paragraph
+    summary = "Missed diagnoses include " + ", ".join(unique_diagnoses[:-1])
+    if len(unique_diagnoses) > 1:
+        summary += f", and {unique_diagnoses[-1]}"
+    elif len(unique_diagnoses) == 1:
+        summary = "Missed diagnoses include " + unique_diagnoses[0]
+    summary += ", all of which require urgent clinical review to prevent potential adverse outcomes."
+    
+    return summary.strip()
 
 def init_agent():
     logger.info("Initializing model...")
@@ -227,7 +236,7 @@ def create_ui(agent):
     with gr.Blocks(theme=gr.themes.Soft()) as demo:
         gr.Markdown("<h1 style='text-align: center;'>🩺 Clinical Oversight Assistant</h1>")
         chatbot = gr.Chatbot(label="Detailed Analysis", height=600, type="messages")
-        final_summary = gr.Markdown(label="Summary of Clinical Oversights")
+        final_summary = gr.Markdown(label="Summary of Missed Diagnoses")
         file_upload = gr.File(file_types=[".pdf", ".csv", ".xls", ".xlsx"], file_count="multiple")
         msg_input = gr.Textbox(placeholder="Ask about potential oversights...", show_label=False)
         send_btn = gr.Button("Analyze", variant="primary")
@@ -235,7 +244,7 @@ def create_ui(agent):
         progress_bar = gr.Progress()
 
         prompt_template = """
-Analyze the patient record excerpt for clinical oversights. Provide a concise, evidence-based summary in markdown with findings grouped under headings (e.g., 'Missed Diagnoses'). For each finding, include clinical context, risks, and recommendations. Output only markdown bullet points under headings. If no issues, state "No issues identified".
+Analyze the patient record excerpt for missed diagnoses only. Provide a concise, evidence-based summary as a single paragraph without headings or bullet points. Include specific clinical findings (e.g., 'elevated blood pressure (160/95) on page 10'), their potential implications (e.g., 'may indicate untreated hypertension'), and a recommendation for urgent review. Do not include other oversight categories like medication conflicts. If no missed diagnoses are found, state 'No missed diagnoses identified' in a single sentence.
 Patient Record Excerpt (Chunk {0} of {1}):
 {chunk}
 """
@@ -284,13 +293,13 @@ Patient Record Excerpt (Chunk {0} of {1}):
                                     for m in chunk_output:
                                         if hasattr(m, 'content') and m.content:
                                             cleaned = clean_response(m.content)
-                                            if cleaned and re.search(r"###\s*\w+", cleaned):
-                                                chunk_response += cleaned + "\n\n"
+                                            if cleaned:
+                                                chunk_response += cleaned + " "
                                 elif isinstance(chunk_output, str) and chunk_output.strip():
-                                    cleaned = clean_response(m.content)
-                                    if cleaned and re.search(r"###\s*\w+", cleaned):
-                                        chunk_response += cleaned + "\n\n"
-                            batch_responses.append(chunk_response)
+                                    cleaned = clean_response(chunk_output)
+                                    if cleaned:
+                                        chunk_response += cleaned + " "
+                            batch_responses.append(chunk_response.strip())
                             torch.cuda.empty_cache()
                             gc.collect()
 
@@ -298,14 +307,14 @@ Patient Record Excerpt (Chunk {0} of {1}):
                         if chunk_response:
                             combined_response += f"--- Analysis for Chunk {chunk_idx} ---\n{chunk_response}\n"
                         else:
-                            combined_response += f"--- Analysis for Chunk {chunk_idx} ---\nNo oversights identified for this chunk.\n\n"
+                            combined_response += f"--- Analysis for Chunk {chunk_idx} ---\nNo missed diagnoses identified.\n"
                         history[-1] = {"role": "assistant", "content": combined_response.strip()}
                         yield history, None, ""
 
-                if combined_response.strip() and not all("No oversights identified" in chunk for chunk in combined_response.split("--- Analysis for Chunk")):
+                if combined_response.strip() and not all("No missed diagnoses identified" in chunk for chunk in combined_response.split("--- Analysis for Chunk")):
                     history[-1]["content"] = combined_response.strip()
                 else:
-                    history.append({"role": "assistant", "content": "No oversights identified in the provided records."})
+                    history.append({"role": "assistant", "content": "No missed diagnoses identified in the provided records."})
 
                 summary = summarize_findings(combined_response)
                 report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt") if file_hash_value else None
@@ -317,7 +326,7 @@ Patient Record Excerpt (Chunk {0} of {1}):
             except Exception as e:
                 logger.error("Analysis error: %s", e)
                 history.append({"role": "assistant", "content": f"❌ Error occurred: {str(e)}"})
-                yield history, None, f"### Summary of Clinical Oversights\nError occurred during analysis: {str(e)}"
+                yield history, None, f"Error occurred during analysis: {str(e)}"
 
         send_btn.click(analyze, inputs=[msg_input, gr.State([]), file_upload], outputs=[chatbot, download_output, final_summary])
         msg_input.submit(analyze, inputs=[msg_input, gr.State([]), file_upload], outputs=[chatbot, download_output, final_summary])