Jesse Liu commited on
Commit
88cb0d9
·
1 Parent(s): a25d70c

Sync latest app + data (clinician groups, jsonl)

Browse files
Files changed (1) hide show
  1. chatgpt.py +84 -70
chatgpt.py CHANGED
@@ -1035,19 +1035,20 @@ with gr.Blocks() as app:
1035
  clinician_choices = sorted(eval_groups.keys())
1036
  eval_group_dropdown = gr.Dropdown(
1037
  choices=clinician_choices,
1038
- label="Select Your ID",
1039
  value=None,
1040
  interactive=True
1041
  )
1042
  group_info_display = gr.Markdown(
1043
- value="Select a group to see your assigned patients",
1044
  visible=True
1045
  )
1046
 
1047
  option_titles = [option["title"] for option in options]
1048
 
1049
  # Single patient selector (one admission per patient)
1050
- patient_ids = patient_ids_list
 
1051
  eval_patient_radio = gr.Radio(
1052
  choices=patient_ids,
1053
  label="Choose Patient",
@@ -1075,6 +1076,7 @@ with gr.Blocks() as app:
1075
  )
1076
 
1077
  # Patient Data Display (based on selected sample)
 
1078
  with gr.Accordion("Patient raw data from EHR", open=True):
1079
  patient_input_display = gr.HTML(
1080
  value="Select a patient sample to view data",
@@ -1083,20 +1085,21 @@ with gr.Blocks() as app:
1083
  )
1084
 
1085
  # AI Summary Display (scrollable)
 
1086
  with gr.Accordion("🤖 AI Generated Summary", open=True):
1087
  ai_summary_display = gr.Markdown(
1088
  value="Select a patient to view AI summary...",
1089
  label="AI Generated Summary",
1090
  elem_id="ai_summary_display"
1091
  )
1092
- timestamp_display = gr.Markdown(value="", label="Timestamp")
1093
 
1094
  # Overall Quality Rating
 
1095
  overall_rating = gr.Slider(
1096
  minimum=1,
1097
  maximum=10,
1098
  step=1,
1099
- label="Overall Quality (1=Poor, 10=Excellent)",
1100
  value=0
1101
  )
1102
  gr.Markdown("*A global assessment integrating all dimensions (accuracy, completeness, relevance...)*")
@@ -1104,95 +1107,107 @@ with gr.Blocks() as app:
1104
  # Core Evaluation Dimensions
1105
  gr.Markdown("### Core Evaluation Dimensions")
1106
  with gr.Row():
1107
- clinical_accuracy = gr.Slider(
1108
- minimum=1,
1109
- maximum=10,
1110
- step=1,
1111
- label="Clinical Accuracy (1=Poor, 10=Excellent)",
1112
- value=0
1113
- )
1114
- completeness_coverage = gr.Slider(
1115
- minimum=1,
1116
- maximum=10,
1117
- step=1,
1118
- label="Completeness/Coverage (1=Poor, 10=Excellent)",
1119
- value=0
1120
- )
 
 
 
 
1121
 
1122
  with gr.Row():
1123
- clinical_relevance = gr.Slider(
1124
- minimum=1,
1125
- maximum=10,
1126
- step=1,
1127
- label="Clinical Relevance (1=Poor, 10=Excellent)",
1128
- value=0
1129
- )
1130
- clarity_structure = gr.Slider(
1131
- minimum=1,
1132
- maximum=10,
1133
- step=1,
1134
- label="Clarity and Structure (1=Poor, 10=Excellent)",
1135
- value=0
1136
- )
 
 
 
 
1137
 
1138
  with gr.Row():
1139
- reasoning_risk = gr.Slider(
1140
- minimum=1,
1141
- maximum=10,
1142
- step=1,
1143
- label="Reasoning/Risk Stratification (1=Poor, 10=Excellent)",
1144
- value=0
1145
- )
1146
- actionability = gr.Slider(
1147
- minimum=1,
1148
- maximum=10,
1149
- step=1,
1150
- label="Actionability (1=Poor, 10=Excellent)",
1151
- value=0
1152
- )
 
 
 
 
1153
 
1154
  # Critical Error Assessment
1155
  gr.Markdown("### Critical Error Assessment")
1156
- gr.Markdown("*Number of hallucinations and omission of critical information that could directly lead to adverse events*")
1157
  with gr.Row():
1158
  hallucination = gr.Slider(
1159
  minimum=1,
1160
  maximum=10,
1161
  step=1,
1162
- label="Hallucination (1=None, 10=Severe)",
1163
  value=0
1164
  )
1165
  critical_omission = gr.Slider(
1166
  minimum=1,
1167
  maximum=10,
1168
  step=1,
1169
- label="Critical Omission (1=None, 10=Severe)",
1170
  value=0
1171
  )
1172
 
1173
- # Detailed feedback
1174
- feedback_text = gr.Textbox(
1175
- label="Overall Feedback",
1176
- placeholder="Please provide specific feedback about the AI summary...",
1177
- lines=3
1178
- )
1179
-
1180
  # Hallucination and Critical Omission Comments
1181
- gr.Markdown("### Point Out Issues")
1182
  with gr.Row():
1183
  hallucination_comments = gr.Textbox(
1184
  label="Hallucination",
1185
- placeholder="Point out any hallucinations (incorrect or fabricated information) in the AI summary...",
1186
- lines=3,
1187
  value=""
1188
  )
1189
  critical_omission_comments = gr.Textbox(
1190
  label="Critical Omission",
1191
- placeholder="Point out any critical omissions (important missing information) in the AI summary...",
1192
- lines=3,
1193
  value=""
1194
  )
1195
 
 
 
 
 
 
 
 
 
1196
  # Submit evaluation button
1197
  submit_eval_button = gr.Button(
1198
  "Submit Patient Evaluation",
@@ -1391,22 +1406,21 @@ with gr.Blocks() as app:
1391
 
1392
  def update_patient_eval_display(patient_id, admission_label=None):
1393
  """
1394
- Return patient input HTML, summary markdown, and timestamp for the selected patient.
1395
  Single-admission mode: always uses the first (only) record for that patient.
1396
  """
1397
  try:
1398
  if not patient_id:
1399
- return "No patient selected", "Select a patient to view AI summary...", ""
1400
 
1401
  pid = str(patient_id)
1402
  if pid not in eval_data_by_patient or not eval_data_by_patient[pid]:
1403
- return "Patient data not found", "AI summary not found", ""
1404
 
1405
  # Single admission per patient
1406
  first_key = list(eval_data_by_patient[pid].keys())[0]
1407
  rec = eval_data_by_patient[pid][first_key]
1408
 
1409
- at = rec.get('admittime') or 'None'
1410
  summary_text = clean_summary_output(rec.get('summary', '') or 'No summary available')
1411
  input_text = rec.get('input_text', '') or 'No input data available'
1412
 
@@ -1426,12 +1440,12 @@ with gr.Blocks() as app:
1426
  '</pre>'
1427
  )
1428
  # Return raw markdown text; scrolling controlled by CSS on elem_id
1429
- return input_html, summary_text, f"Admit time: {at}"
1430
  except Exception as e:
1431
  import traceback
1432
  error_msg = f"Error loading patient data: {str(e)}\n{traceback.format_exc()}"
1433
  print(f"[ERROR] {error_msg}")
1434
- return "Error loading data", f"Error: {str(e)}", ""
1435
 
1436
  # Handle clinician ID selection to filter patient list
1437
  if eval_groups:
@@ -1439,7 +1453,7 @@ with gr.Blocks() as app:
1439
  """Filter patient list based on selected clinician ID and auto-fill clinician ID field"""
1440
  if not selected_clinician:
1441
  filtered_patients = []
1442
- info_text = "Please select your clinician ID"
1443
  clinician_id_update = gr.update()
1444
  else:
1445
  # Directly use selected_clinician as key (no mapping needed)
@@ -1474,7 +1488,7 @@ with gr.Blocks() as app:
1474
  eval_patient_radio.change(
1475
  fn=on_patient_change,
1476
  inputs=[eval_patient_radio],
1477
- outputs=[patient_input_display, ai_summary_display, timestamp_display]
1478
  )
1479
 
1480
 
 
1035
  clinician_choices = sorted(eval_groups.keys())
1036
  eval_group_dropdown = gr.Dropdown(
1037
  choices=clinician_choices,
1038
+ label="Select your ID to see assigned patient cases",
1039
  value=None,
1040
  interactive=True
1041
  )
1042
  group_info_display = gr.Markdown(
1043
+ value="",
1044
  visible=True
1045
  )
1046
 
1047
  option_titles = [option["title"] for option in options]
1048
 
1049
  # Single patient selector (one admission per patient)
1050
+ # Default to empty list - patients will be shown only after clinician selects their ID
1051
+ patient_ids = [] if eval_groups else patient_ids_list
1052
  eval_patient_radio = gr.Radio(
1053
  choices=patient_ids,
1054
  label="Choose Patient",
 
1076
  )
1077
 
1078
  # Patient Data Display (based on selected sample)
1079
+ gr.Markdown("### Patient Raw Data from EHR")
1080
  with gr.Accordion("Patient raw data from EHR", open=True):
1081
  patient_input_display = gr.HTML(
1082
  value="Select a patient sample to view data",
 
1085
  )
1086
 
1087
  # AI Summary Display (scrollable)
1088
+ gr.Markdown("### AI Generated Summary")
1089
  with gr.Accordion("🤖 AI Generated Summary", open=True):
1090
  ai_summary_display = gr.Markdown(
1091
  value="Select a patient to view AI summary...",
1092
  label="AI Generated Summary",
1093
  elem_id="ai_summary_display"
1094
  )
 
1095
 
1096
  # Overall Quality Rating
1097
+ gr.Markdown("### Overall evaluation")
1098
  overall_rating = gr.Slider(
1099
  minimum=1,
1100
  maximum=10,
1101
  step=1,
1102
+ label="Overall Quality",
1103
  value=0
1104
  )
1105
  gr.Markdown("*A global assessment integrating all dimensions (accuracy, completeness, relevance...)*")
 
1107
  # Core Evaluation Dimensions
1108
  gr.Markdown("### Core Evaluation Dimensions")
1109
  with gr.Row():
1110
+ with gr.Column():
1111
+ clinical_accuracy = gr.Slider(
1112
+ minimum=1,
1113
+ maximum=10,
1114
+ step=1,
1115
+ label="Clinical Accuracy",
1116
+ value=0
1117
+ )
1118
+ gr.Markdown("*The extent to which the content is factually correct and consistent with the source EHR.*")
1119
+ with gr.Column():
1120
+ completeness_coverage = gr.Slider(
1121
+ minimum=1,
1122
+ maximum=10,
1123
+ step=1,
1124
+ label="Completeness / Coverage",
1125
+ value=0
1126
+ )
1127
+ gr.Markdown("*Whether all information required for a comprehensive preoperative evaluation is included.*")
1128
 
1129
  with gr.Row():
1130
+ with gr.Column():
1131
+ clinical_relevance = gr.Slider(
1132
+ minimum=1,
1133
+ maximum=10,
1134
+ step=1,
1135
+ label="Clinical Relevance",
1136
+ value=0
1137
+ )
1138
+ gr.Markdown("*The degree to which the summary prioritizes information most important for preoperative decision-making.*")
1139
+ with gr.Column():
1140
+ clarity_structure = gr.Slider(
1141
+ minimum=1,
1142
+ maximum=10,
1143
+ step=1,
1144
+ label="Clarity and Structure",
1145
+ value=0
1146
+ )
1147
+ gr.Markdown("*How easily clinicians can understand, navigate, and apply the summary in practice.*")
1148
 
1149
  with gr.Row():
1150
+ with gr.Column():
1151
+ reasoning_risk = gr.Slider(
1152
+ minimum=1,
1153
+ maximum=10,
1154
+ step=1,
1155
+ label="Reasoning / Risk Stratification",
1156
+ value=0
1157
+ )
1158
+ gr.Markdown("*Accurate identification of perioperative risk factors, appropriate interpretation, and required follow-up actions.*")
1159
+ with gr.Column():
1160
+ actionability = gr.Slider(
1161
+ minimum=1,
1162
+ maximum=10,
1163
+ step=1,
1164
+ label="Actionability",
1165
+ value=0
1166
+ )
1167
+ gr.Markdown("*Whether the summary supports clear clinical decisions and next steps.*")
1168
 
1169
  # Critical Error Assessment
1170
  gr.Markdown("### Critical Error Assessment")
1171
+ gr.Markdown("*If critical errors were identified, please provide detail information below*")
1172
  with gr.Row():
1173
  hallucination = gr.Slider(
1174
  minimum=1,
1175
  maximum=10,
1176
  step=1,
1177
+ label="Hallucination",
1178
  value=0
1179
  )
1180
  critical_omission = gr.Slider(
1181
  minimum=1,
1182
  maximum=10,
1183
  step=1,
1184
+ label="Critical Omission",
1185
  value=0
1186
  )
1187
 
 
 
 
 
 
 
 
1188
  # Hallucination and Critical Omission Comments
 
1189
  with gr.Row():
1190
  hallucination_comments = gr.Textbox(
1191
  label="Hallucination",
1192
+ placeholder="Copy the relevant portion of the summary and provide your feedback directly underneath",
1193
+ lines=5,
1194
  value=""
1195
  )
1196
  critical_omission_comments = gr.Textbox(
1197
  label="Critical Omission",
1198
+ placeholder="Copy the relevant portion of the summary and provide your feedback directly underneath",
1199
+ lines=5,
1200
  value=""
1201
  )
1202
 
1203
+ # Detailed feedback
1204
+ gr.Markdown("### Other Feedback and Comment")
1205
+ feedback_text = gr.Textbox(
1206
+ label="Other Feedback and Comment",
1207
+ placeholder="Please provide any additional feedback or comments...",
1208
+ lines=3
1209
+ )
1210
+
1211
  # Submit evaluation button
1212
  submit_eval_button = gr.Button(
1213
  "Submit Patient Evaluation",
 
1406
 
1407
  def update_patient_eval_display(patient_id, admission_label=None):
1408
  """
1409
+ Return patient input HTML and summary markdown for the selected patient.
1410
  Single-admission mode: always uses the first (only) record for that patient.
1411
  """
1412
  try:
1413
  if not patient_id:
1414
+ return "No patient selected", "Select a patient to view AI summary..."
1415
 
1416
  pid = str(patient_id)
1417
  if pid not in eval_data_by_patient or not eval_data_by_patient[pid]:
1418
+ return "Patient data not found", "AI summary not found"
1419
 
1420
  # Single admission per patient
1421
  first_key = list(eval_data_by_patient[pid].keys())[0]
1422
  rec = eval_data_by_patient[pid][first_key]
1423
 
 
1424
  summary_text = clean_summary_output(rec.get('summary', '') or 'No summary available')
1425
  input_text = rec.get('input_text', '') or 'No input data available'
1426
 
 
1440
  '</pre>'
1441
  )
1442
  # Return raw markdown text; scrolling controlled by CSS on elem_id
1443
+ return input_html, summary_text
1444
  except Exception as e:
1445
  import traceback
1446
  error_msg = f"Error loading patient data: {str(e)}\n{traceback.format_exc()}"
1447
  print(f"[ERROR] {error_msg}")
1448
+ return "Error loading data", f"Error: {str(e)}"
1449
 
1450
  # Handle clinician ID selection to filter patient list
1451
  if eval_groups:
 
1453
  """Filter patient list based on selected clinician ID and auto-fill clinician ID field"""
1454
  if not selected_clinician:
1455
  filtered_patients = []
1456
+ info_text = ""
1457
  clinician_id_update = gr.update()
1458
  else:
1459
  # Directly use selected_clinician as key (no mapping needed)
 
1488
  eval_patient_radio.change(
1489
  fn=on_patient_change,
1490
  inputs=[eval_patient_radio],
1491
+ outputs=[patient_input_display, ai_summary_display]
1492
  )
1493
 
1494