Spaces:
Running
Running
Commit
·
0e43c47
1
Parent(s):
075fc40
Update
Browse files- app.py +354 -94
- demo_assets.py +200 -0
app.py
CHANGED
@@ -1,4 +1,7 @@
|
|
|
|
1 |
import re
|
|
|
|
|
2 |
import argparse
|
3 |
import torch
|
4 |
import gradio as gr
|
@@ -10,7 +13,9 @@ from model import load_model
|
|
10 |
from datetime import datetime
|
11 |
from dateutil import parser
|
12 |
from demo_assets import *
|
13 |
-
from typing import List, Dict, Any
|
|
|
|
|
14 |
|
15 |
def get_args():
|
16 |
parser = argparse.ArgumentParser()
|
@@ -82,23 +87,235 @@ elif args.task == 'token':
|
|
82 |
elif args.label_encoding == 'boe':
|
83 |
args.num_labels *= 3
|
84 |
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
"
|
99 |
-
|
100 |
-
|
101 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
102 |
def postprocess_labels(text, logits, t2c):
|
103 |
tags = [None for _ in text]
|
104 |
labels = logits.argmax(-1)
|
@@ -217,6 +434,7 @@ def get_structured_data(*inputs):
|
|
217 |
data.append({
|
218 |
'date': date,
|
219 |
'timestamp': parser.parse(date),
|
|
|
220 |
'decision_type': categories[c], 'details': text[s:e]})
|
221 |
return data
|
222 |
|
@@ -279,7 +497,7 @@ def create_timeline_plot(data: List[Dict[str, Any]]):
|
|
279 |
fig.update_layout(height=600)
|
280 |
return fig
|
281 |
|
282 |
-
def filter_timeline(
|
283 |
global structured_data
|
284 |
filtered_data = structured_data
|
285 |
if 'All' not in decision_types:
|
@@ -294,59 +512,33 @@ def filter_timeline(decision_type: str, start_date: str, end_date: str) -> px.sc
|
|
294 |
def generate_summary(*inputs) -> str:
|
295 |
global structured_data
|
296 |
structured_data = get_structured_data(*inputs)
|
297 |
-
decision_types = {}
|
298 |
-
for event in structured_data:
|
299 |
-
decision_type = event['decision_type']
|
300 |
-
decision_types[decision_type] = decision_types.get(decision_type, 0) + 1
|
301 |
|
302 |
-
|
303 |
-
for
|
304 |
-
|
305 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
306 |
|
307 |
global sum_c
|
308 |
sum_c = 1
|
309 |
-
SUM_INPUTS = 20
|
310 |
structured_data = []
|
311 |
|
312 |
device = model.backbone.device
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
det_desc = ['Admit, discharge, follow-up, referral',
|
320 |
-
'Ordering test, consulting colleague, seeking external information',
|
321 |
-
'Diagnostic conclusion, evaluation of health state, etiological inference, prognostic judgment',
|
322 |
-
'Quantitative or qualitative',
|
323 |
-
'Start, stop, alter, maintain, refrain',
|
324 |
-
'Start, stop, alter, maintain, refrain',
|
325 |
-
'Positive, negative, ambiguous test results',
|
326 |
-
'Transfer responsibility, wait and see, change subject',
|
327 |
-
'Advice or precaution',
|
328 |
-
'Sick leave, drug refund, insurance, disability']
|
329 |
-
|
330 |
-
desc = '### Zones (categories)\n'
|
331 |
-
desc += '| | |\n| --- | --- |\n'
|
332 |
-
for i,cat in enumerate(categories):
|
333 |
-
desc += f'| {unicode_symbols[i]} **{cat}** | {det_desc[i]}|\n'
|
334 |
-
|
335 |
-
#colors
|
336 |
-
#markdown labels
|
337 |
-
#legend and desc
|
338 |
-
#css font-size
|
339 |
-
css = '.category-legend {border:1px dashed black;}'\
|
340 |
-
'.text-sm {font-size: 1.5rem; line-height: 200%;}'\
|
341 |
-
'.gr-sample-textbox {width: 1000px; white-space: nowrap; overflow: hidden; text-overflow: ellipsis;}'\
|
342 |
-
'.text-limit label textarea {height: 150px !important; overflow: scroll; }'\
|
343 |
-
'.text-gray-500 {color: #111827; font-weight: 600; font-size: 1.25em; margin-top: 1.6em; margin-bottom: 0.6em;'\
|
344 |
-
'line-height: 1.6;}'\
|
345 |
-
'#sum-out {border: 2px solid #007bff; padding: 20px; border-radius: 10px; box-shadow: 0 0 10px rgba(0, 0, 0, 0.1);'
|
346 |
-
title='Clinical Decision Zoning'
|
347 |
-
with gr.Blocks(title=title, css=css) as demo:
|
348 |
-
gr.Markdown(f'# {title}')
|
349 |
with gr.Tab("Label a Clinical Note"):
|
|
|
350 |
with gr.Row():
|
351 |
with gr.Column():
|
352 |
gr.Markdown("## Enter a Discharge Summary or Clinical Note"),
|
@@ -359,33 +551,18 @@ with gr.Blocks(title=title, css=css) as demo:
|
|
359 |
gr.Markdown("## Labeled Summary or Note"),
|
360 |
text_out = gr.Highlight(label="", combine_adjacent=True, show_legend=False, color_map=color_map)
|
361 |
gr.Examples(text_examples, inputs=text_input)
|
362 |
-
with gr.Tab("
|
363 |
-
|
364 |
-
with gr.Column():
|
365 |
-
sum_inputs = [gr.Text(label='Clinical Note 1', elem_classes='text-limit')]
|
366 |
-
sum_inputs.extend([gr.Text(label='Clinical Note %d'%i, visible=False, elem_classes='text-limit')
|
367 |
-
for i in range(2, SUM_INPUTS + 1)])
|
368 |
-
sum_btn = gr.Button('Run')
|
369 |
-
with gr.Row():
|
370 |
-
ex_add = gr.Button("+")
|
371 |
-
ex_sub = gr.Button("-")
|
372 |
-
upload = gr.File(label='Upload clinical notes', file_types=['text'], file_count='multiple')
|
373 |
-
gr.Examples(sum_examples, inputs=upload,
|
374 |
-
fn = update_inputs, outputs=sum_inputs, run_on_click=True)
|
375 |
-
with gr.Column():
|
376 |
-
gr.Markdown("## Summarized Clinical Decision History")
|
377 |
-
sum_out = gr.Markdown(elem_id='sum-out')
|
378 |
-
with gr.Tab("Timeline Visualization Tool"):
|
379 |
with gr.Column():
|
380 |
-
|
381 |
-
|
382 |
for i in range(2, SUM_INPUTS + 1)])
|
383 |
with gr.Row():
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
gr.Examples(sum_examples, inputs=
|
388 |
-
fn = update_inputs, outputs=
|
389 |
with gr.Column():
|
390 |
with gr.Row():
|
391 |
decision_type = gr.Dropdown(["All"] + categories,
|
@@ -399,19 +576,102 @@ with gr.Blocks(title=title, css=css) as demo:
|
|
399 |
timeline_plot = gr.Plot()
|
400 |
|
401 |
summary_button = gr.Button("Generate Summary")
|
402 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
403 |
gr.Markdown(desc)
|
404 |
|
405 |
# Functions
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
406 |
text_input.submit(process, inputs=text_input, outputs=text_out)
|
407 |
text_btn.click(process, inputs=text_input, outputs=text_out)
|
408 |
upload.change(update_inputs, inputs=upload, outputs=sum_inputs)
|
409 |
-
upload2.change(update_inputs, inputs=upload2, outputs=sum_inputs2)
|
410 |
ex_add.click(add_ex, inputs=sum_inputs, outputs=sum_inputs)
|
411 |
ex_sub.click(sub_ex, inputs=sum_inputs, outputs=sum_inputs)
|
412 |
-
ex_add2.click(add_ex, inputs=sum_inputs2, outputs=sum_inputs2)
|
413 |
-
ex_sub2.click(sub_ex, inputs=sum_inputs2, outputs=sum_inputs2)
|
414 |
-
sum_btn.click(process_sum, inputs=sum_inputs, outputs=sum_out)
|
415 |
filter_button.click(filter_timeline, inputs=[decision_type, start_date, end_date], outputs=timeline_plot)
|
416 |
-
summary_button.click(generate_summary, inputs=
|
417 |
demo.launch(share=True)
|
|
|
1 |
+
import os
|
2 |
import re
|
3 |
+
import uuid
|
4 |
+
import json
|
5 |
import argparse
|
6 |
import torch
|
7 |
import gradio as gr
|
|
|
13 |
from datetime import datetime
|
14 |
from dateutil import parser
|
15 |
from demo_assets import *
|
16 |
+
from typing import List, Dict, Any, Optional, Tuple
|
17 |
+
from dataclasses import dataclass
|
18 |
+
from collections import defaultdict
|
19 |
|
20 |
def get_args():
|
21 |
parser = argparse.ArgumentParser()
|
|
|
87 |
elif args.label_encoding == 'boe':
|
88 |
args.num_labels *= 3
|
89 |
|
90 |
+
@dataclass
|
91 |
+
class KeyDef:
|
92 |
+
key: str
|
93 |
+
name: str
|
94 |
+
desc: str = ''
|
95 |
+
color: str = 'lightblue'
|
96 |
+
symbol: str = ''
|
97 |
+
|
98 |
+
class AnnotationState:
|
99 |
+
def __init__(self):
|
100 |
+
self.entity_regex = r'\[\@.*?\#.*?\*\](?!\#)'
|
101 |
+
self.recommend_regex = r'\[\$.*?\#.*?\*\](?!\#)'
|
102 |
+
self.history = []
|
103 |
+
self.config_file = "configs/default.config"
|
104 |
+
self.press_commands = self.read_config()
|
105 |
+
# Internal state holds the actual annotations
|
106 |
+
self.annotations = []
|
107 |
+
self.raw_text = ""
|
108 |
+
|
109 |
+
def read_config(self) -> List[KeyDef]:
|
110 |
+
if not os.path.exists(self.config_file):
|
111 |
+
default_config = [{
|
112 |
+
'key': key,
|
113 |
+
'name': name,
|
114 |
+
'color': color,
|
115 |
+
'symbol': symbol
|
116 |
+
}
|
117 |
+
for key, name, color, symbol in zip(keys, categories, colors, unicode_symbols)
|
118 |
+
]
|
119 |
+
os.makedirs("configs", exist_ok=True)
|
120 |
+
with open(self.config_file, 'w') as fp:
|
121 |
+
json.dump(default_config, fp)
|
122 |
+
|
123 |
+
with open(self.config_file, 'r') as fp:
|
124 |
+
config_dict = json.load(fp)
|
125 |
+
|
126 |
+
return [KeyDef(**entry) for entry in config_dict]
|
127 |
+
|
128 |
+
def get_cmd_by_key(self, key: str) -> Optional[KeyDef]:
|
129 |
+
return next((cmd for cmd in self.press_commands if cmd.key == key), None)
|
130 |
+
|
131 |
+
def set_text(self, text: str):
|
132 |
+
"""Initialize with new text, clearing annotations"""
|
133 |
+
self.raw_text = text
|
134 |
+
self.annotations = []
|
135 |
+
self.history = []
|
136 |
+
|
137 |
+
def add_annotation(self, start: int, end: int, entity_type: str) -> str:
|
138 |
+
"""Add new annotation and return display text"""
|
139 |
+
# Save current state to history
|
140 |
+
self.history.append((self.raw_text, list(self.annotations)))
|
141 |
+
if len(self.history) > 20:
|
142 |
+
self.history.pop(0)
|
143 |
+
|
144 |
+
# Add new annotation
|
145 |
+
self.annotations.append((start, end, entity_type))
|
146 |
+
return self.get_display_text()
|
147 |
+
|
148 |
+
def remove_annotation(self, start: int, end: int) -> str:
|
149 |
+
"""Remove annotation at position if it exists, splitting spans if needed"""
|
150 |
+
self.history.append((self.raw_text, list(self.annotations)))
|
151 |
+
|
152 |
+
new_annotations = []
|
153 |
+
|
154 |
+
for a in self.annotations:
|
155 |
+
annotation_start, annotation_end, entity_type = a
|
156 |
+
|
157 |
+
# If the current annotation does not overlap, keep it as is
|
158 |
+
if annotation_end < start or annotation_start > end:
|
159 |
+
new_annotations.append(a)
|
160 |
+
|
161 |
+
# If the removed span is a proper subset, split the annotation
|
162 |
+
elif annotation_start < start and annotation_end > end:
|
163 |
+
new_annotations.append((annotation_start, start - 1, entity_type))
|
164 |
+
new_annotations.append((end + 1, annotation_end, entity_type))
|
165 |
+
|
166 |
+
# If there's overlap with the start, but not the end
|
167 |
+
elif annotation_start < start <= annotation_end:
|
168 |
+
new_annotations.append((annotation_start, start - 1, entity_type))
|
169 |
+
|
170 |
+
# If there's overlap with the end, but not the start
|
171 |
+
elif annotation_start <= end < annotation_end:
|
172 |
+
new_annotations.append((end + 1, annotation_end, entity_type))
|
173 |
+
|
174 |
+
self.annotations = new_annotations
|
175 |
+
return self.get_display_text()
|
176 |
+
|
177 |
+
def undo(self) -> str:
|
178 |
+
"""Undo last annotation action"""
|
179 |
+
if not self.history:
|
180 |
+
return self.get_display_text()
|
181 |
+
|
182 |
+
self.raw_text, self.annotations = self.history.pop()
|
183 |
+
return self.get_display_text()
|
184 |
+
|
185 |
+
def get_display_text(self) -> str:
|
186 |
+
"""Generate display text with HTML formatting for annotations"""
|
187 |
+
if not self.annotations:
|
188 |
+
return f'<div id="annotated-text">{self.raw_text}</div> <div id="legend"></div>'
|
189 |
+
|
190 |
+
# Sort annotations by start position
|
191 |
+
sorted_annotations = sorted(self.annotations, key=lambda x: (x[0], -x[1]))
|
192 |
+
|
193 |
+
# Build display text with HTML spans
|
194 |
+
result = ['<div id="annotated-text">']
|
195 |
+
last_end = 0
|
196 |
+
|
197 |
+
for start, end, entity_type in sorted_annotations:
|
198 |
+
if start < last_end and end > last_end:
|
199 |
+
start = last_end
|
200 |
+
elif start < last_end:
|
201 |
+
continue
|
202 |
+
|
203 |
+
|
204 |
+
# Add text before annotation
|
205 |
+
result.append(self.raw_text[last_end:start])
|
206 |
+
|
207 |
+
# Add annotated text with highlighting
|
208 |
+
text = self.raw_text[start:end]
|
209 |
+
cmd = self.get_cmd_by_key(entity_type)
|
210 |
+
color = cmd.color
|
211 |
+
result.append(f'<span style="background-color: {color};" title="{cmd.name}">{text}</span>') # Nicer tooltip
|
212 |
+
|
213 |
+
last_end = end
|
214 |
+
|
215 |
+
# Add remaining text
|
216 |
+
result.append(self.raw_text[last_end:])
|
217 |
+
result.append('</div>')
|
218 |
+
|
219 |
+
# Generate legend
|
220 |
+
legend = ['<div id="legend" style="margin-top: 10px;"><span style="font-weight: bold;">Legend:</span > '] # Margin and bold legend title
|
221 |
+
used_categories = sorted(list(set([a[2] for a in self.annotations])))
|
222 |
+
for cat in used_categories:
|
223 |
+
cmd = self.get_cmd_by_key(cat)
|
224 |
+
legend.append(f'<span style="background-color: {cmd.color}; padding: 3px 5px; border-radius: 3px; margin-right: 5px; font-size:0.9em; display: inline-block; vertical-align: middle; color: black; font-family: sans-serif;">{cmd.name}</span>') # Improved legend item styling
|
225 |
+
legend.append('</div>')
|
226 |
+
result.extend(legend)
|
227 |
+
|
228 |
+
return "".join(result)
|
229 |
+
|
230 |
+
|
231 |
+
def get_annotated_text(self, annotator_id=None, discharge_summary_id=None) -> dict:
|
232 |
+
"""Generate a dictionary containing annotation data."""
|
233 |
+
unique_id = str(uuid.uuid4())[:8]
|
234 |
+
annotations = []
|
235 |
+
if self.annotations:
|
236 |
+
sorted_annotations = sorted(self.annotations, key=lambda x: (x[0], -x[1]))
|
237 |
+
for idx, (start, end, entity_type) in enumerate(sorted_annotations):
|
238 |
+
cmd = self.get_cmd_by_key(entity_type)
|
239 |
+
annotations.append({
|
240 |
+
"decision": self.raw_text[start:end],
|
241 |
+
"category": f'Category {categories.index(cmd.name) + 1}: {cmd.name}',
|
242 |
+
"start_offset": start,
|
243 |
+
"end_offset": end,
|
244 |
+
"annotation_id": f'{unique_id}_{idx}'
|
245 |
+
})
|
246 |
+
|
247 |
+
return {
|
248 |
+
"annotator_id": annotator_id if annotator_id else None,
|
249 |
+
"discharge_summary_id": discharge_summary_id if discharge_summary_id else None,
|
250 |
+
"annotations": annotations
|
251 |
+
}
|
252 |
+
|
253 |
+
def init_text(text):
|
254 |
+
if text:
|
255 |
+
state.set_text(text)
|
256 |
+
return state.get_display_text()
|
257 |
+
return "<div id='annotated-text'>Enter text to begin...</div>"
|
258 |
+
|
259 |
+
def add_entity(cmd_key, start: int, end: int):
|
260 |
+
"""Handle adding new entity annotations"""
|
261 |
+
if start == end:
|
262 |
+
return state.get_display_text(), "No text selected"
|
263 |
+
|
264 |
+
cmd = state.get_cmd_by_key(cmd_key)
|
265 |
+
if not cmd:
|
266 |
+
return state.get_display_text(), "Invalid command"
|
267 |
+
|
268 |
+
new_text = state.add_annotation(start, end, cmd.key)
|
269 |
+
return new_text, f"Added {cmd.name} entity"
|
270 |
+
|
271 |
+
def remove_entity(start: int, end: int):
|
272 |
+
"""Handle removal of annotations"""
|
273 |
+
if start == end:
|
274 |
+
return state.get_display_text(), "No text selected"
|
275 |
+
return state.remove_annotation(start, end), "Removed annotation"
|
276 |
+
|
277 |
+
def undo():
|
278 |
+
"""Handle undoing the last action"""
|
279 |
+
return state.undo(), "Undid last action"
|
280 |
+
|
281 |
+
def download_annotations(annotator_id, discharge_summary_id):
|
282 |
+
"""Generates and provides annotation data for download."""
|
283 |
+
annotation_data = state.get_annotated_text(annotator_id, discharge_summary_id)
|
284 |
+
with open(OUTPUT_PATH, 'w') as f:
|
285 |
+
json.dump(annotation_data, f, indent=4)
|
286 |
+
return OUTPUT_PATH
|
287 |
+
|
288 |
+
|
289 |
+
|
290 |
+
def refresh_annotations(annotator_id, discharge_summary_id):
|
291 |
+
"""Refreshes the displayed annotation JSON."""
|
292 |
+
return state.get_annotated_text(annotator_id, discharge_summary_id)
|
293 |
+
|
294 |
+
|
295 |
+
def clear_annotations():
|
296 |
+
state.set_text(state.raw_text) # Clears annotations by setting empty text
|
297 |
+
return gr.update(interactive=True, elem_classes=[]), state.get_display_text() # added value
|
298 |
+
|
299 |
+
def model_predict(text):
|
300 |
+
"""Placeholder for model prediction logic"""
|
301 |
+
output, t2c = predict(text)
|
302 |
+
spans = indicators_to_spans(output.argmax(-1), t2c)
|
303 |
+
spans = [(s, e, keys[c]) for c, s, e in spans]
|
304 |
+
return spans
|
305 |
+
|
306 |
+
def apply_predictions(text):
|
307 |
+
predictions = model_predict(text)
|
308 |
+
state.set_text(text)
|
309 |
+
for start, end, entity_type in predictions:
|
310 |
+
state.add_annotation(start, end, entity_type)
|
311 |
+
return state.get_display_text()
|
312 |
+
|
313 |
+
state = AnnotationState()
|
314 |
+
all_keys = [f'"{cmd.key}"' for cmd in state.press_commands]
|
315 |
+
key_list_str = f'[{", ".join(all_keys)}]'
|
316 |
+
shortcut_js = shortcut_js_template%key_list_str
|
317 |
+
|
318 |
+
|
319 |
def postprocess_labels(text, logits, t2c):
|
320 |
tags = [None for _ in text]
|
321 |
labels = logits.argmax(-1)
|
|
|
434 |
data.append({
|
435 |
'date': date,
|
436 |
'timestamp': parser.parse(date),
|
437 |
+
'decision_cat': c,
|
438 |
'decision_type': categories[c], 'details': text[s:e]})
|
439 |
return data
|
440 |
|
|
|
497 |
fig.update_layout(height=600)
|
498 |
return fig
|
499 |
|
500 |
+
def filter_timeline(decision_types: str, start_date: str, end_date: str) -> px.scatter:
|
501 |
global structured_data
|
502 |
filtered_data = structured_data
|
503 |
if 'All' not in decision_types:
|
|
|
512 |
def generate_summary(*inputs) -> str:
|
513 |
global structured_data
|
514 |
structured_data = get_structured_data(*inputs)
|
|
|
|
|
|
|
|
|
515 |
|
516 |
+
dates = defaultdict(lambda: defaultdict(list))
|
517 |
+
for event in structured_data:
|
518 |
+
dates[event['date']][event['decision_cat']].append(event['details'])
|
519 |
+
|
520 |
+
out = ""
|
521 |
+
for date in sorted(dates.keys(), key = lambda x: parser.parse(x)):
|
522 |
+
out += f'## **[{date}]**\n\n'
|
523 |
+
decs = dates[date]
|
524 |
+
for c in decs:
|
525 |
+
out += f'### {unicode_symbols[c]} ***{categories[c]}***\n\n'
|
526 |
+
for dec in decs[c]:
|
527 |
+
out += f'{dec}\n\n'
|
528 |
+
return out, create_timeline_plot(structured_data)
|
529 |
|
530 |
global sum_c
|
531 |
sum_c = 1
|
|
|
532 |
structured_data = []
|
533 |
|
534 |
device = model.backbone.device
|
535 |
+
|
536 |
+
with gr.Blocks(head=shortcut_js,
|
537 |
+
title='MedDecXtract', css=css) as demo:
|
538 |
+
gr.Image('assets/logo.png', height=100, container=False, show_download_button=False)
|
539 |
+
gr.Markdown(title)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
540 |
with gr.Tab("Label a Clinical Note"):
|
541 |
+
gr.Markdown(label_desc)
|
542 |
with gr.Row():
|
543 |
with gr.Column():
|
544 |
gr.Markdown("## Enter a Discharge Summary or Clinical Note"),
|
|
|
551 |
gr.Markdown("## Labeled Summary or Note"),
|
552 |
text_out = gr.Highlight(label="", combine_adjacent=True, show_legend=False, color_map=color_map)
|
553 |
gr.Examples(text_examples, inputs=text_input)
|
554 |
+
with gr.Tab("Patient History Visualization"):
|
555 |
+
gr.Markdown(vis_desc)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
556 |
with gr.Column():
|
557 |
+
sum_inputs = [gr.Text(label='Clinical Note 1', elem_classes='text-limit')]
|
558 |
+
sum_inputs.extend([gr.Text(label='Clinical Note %d'%i, visible=False, elem_classes='text-limit')
|
559 |
for i in range(2, SUM_INPUTS + 1)])
|
560 |
with gr.Row():
|
561 |
+
ex_add = gr.Button("+")
|
562 |
+
ex_sub = gr.Button("-")
|
563 |
+
upload = gr.File(label='Upload clinical notes', file_types=['text'], file_count='multiple')
|
564 |
+
gr.Examples(sum_examples, inputs=upload,
|
565 |
+
fn = update_inputs, outputs=sum_inputs, run_on_click=True)
|
566 |
with gr.Column():
|
567 |
with gr.Row():
|
568 |
decision_type = gr.Dropdown(["All"] + categories,
|
|
|
576 |
timeline_plot = gr.Plot()
|
577 |
|
578 |
summary_button = gr.Button("Generate Summary")
|
579 |
+
with gr.Accordion('Summary'):
|
580 |
+
summary_output = gr.Markdown(elem_id='sum-out') #gr.Textbox(label="Summary")
|
581 |
+
with gr.Tab("Interactive Text Annotator"):
|
582 |
+
gr.Markdown(annotator_desc)
|
583 |
+
with gr.Row():
|
584 |
+
with gr.Column():
|
585 |
+
annot_text_input = gr.Textbox(
|
586 |
+
label="Enter Text to Annotate",
|
587 |
+
placeholder="Enter or paste text here...",
|
588 |
+
lines=5,
|
589 |
+
elem_id='annot_text_input'
|
590 |
+
)
|
591 |
+
gr.Examples(text_examples, inputs=annot_text_input)
|
592 |
+
msg_output = gr.Textbox(label="Status Messages", interactive=False)
|
593 |
+
display_area = gr.HTML(
|
594 |
+
label="Annotated Text",
|
595 |
+
value="<div id='annotated-text'><i>Output box</i></div>"
|
596 |
+
)
|
597 |
+
|
598 |
+
k = 3 # Set the maximum number of buttons per row
|
599 |
+
num_buttons = len(state.press_commands)
|
600 |
+
rows = (num_buttons + k - 1) // k
|
601 |
+
entity_buttons = []
|
602 |
+
with gr.Group():
|
603 |
+
predict_btn = gr.Button("Generate Predictions", size='lg', variant='primary')
|
604 |
+
for i in range(rows):
|
605 |
+
with gr.Row():
|
606 |
+
for j in range(min(k, num_buttons - i * k)):
|
607 |
+
real_idx = i * k + j
|
608 |
+
cmd = state.press_commands[real_idx]
|
609 |
+
entity_buttons.append(
|
610 |
+
gr.Button(f"{cmd.symbol} {cmd.name} ({cmd.key})",
|
611 |
+
elem_id=f'btn_{cmd.key}',
|
612 |
+
size='sm'))
|
613 |
+
if i == (rows - 1):
|
614 |
+
remove_btn = gr.Button("Remove (q)", size='sm', variant='secondary', elem_id='btn_q')
|
615 |
+
undo_btn = gr.Button("Undo (z)", size='sm', elem_id='btn_z')
|
616 |
+
clear_btn = gr.Button("Clear Annotations", size='lg', variant='stop')
|
617 |
+
|
618 |
+
|
619 |
+
with gr.Accordion("Download/View Annotations \U0001F4BE", open=False): # Combined Accordion
|
620 |
+
with gr.Row():
|
621 |
+
annotator_id = gr.Textbox(label="Annotator ID", placeholder="Enter your annotator ID")
|
622 |
+
discharge_summary_id = gr.Textbox(label="Discharge Summary ID", placeholder="Enter the discharge summary ID")
|
623 |
+
|
624 |
+
with gr.Row():
|
625 |
+
download_file = gr.File(interactive=False, visible=True, label="Download") # download_btn renamed and made into gr.File
|
626 |
+
annotations_json = gr.JSON(label="Annotations JSON")
|
627 |
+
|
628 |
+
refresh_btn = gr.Button("🔄 Refresh Annotations", elem_id="refresh_btn") # Renamed for clarity
|
629 |
+
download_btn = gr.Button("Download Annotated Text", elem_id="download_btn") # Added a button to trigger download
|
630 |
+
|
631 |
+
|
632 |
+
# Hidden state components for selection
|
633 |
+
selection_start = gr.Number(value=0, visible=False)
|
634 |
+
selection_end = gr.Number(value=0, visible=False)
|
635 |
+
|
636 |
gr.Markdown(desc)
|
637 |
|
638 |
# Functions
|
639 |
+
# Wire up event handlers
|
640 |
+
annot_text_input.change(init_text, annot_text_input, display_area)
|
641 |
+
|
642 |
+
# Wire up the buttons with the selection JavaScript
|
643 |
+
for btn, cmd in zip(entity_buttons, state.press_commands):
|
644 |
+
btn.click(lambda s=None, e=None, c=cmd.key: add_entity(c, s, e),[selection_start, selection_end], [display_area, msg_output], js=select_js).then(
|
645 |
+
lambda: gr.update(interactive=state.annotations == [], elem_classes=[] if state.annotations == [] else ['locked-input']), # Disable input if annotations exist
|
646 |
+
outputs=annot_text_input
|
647 |
+
)
|
648 |
+
|
649 |
+
remove_btn.click( remove_entity, [selection_start, selection_end], [display_area, msg_output], js=select_js).then(
|
650 |
+
lambda: gr.update(interactive=state.annotations == [], elem_classes=[] if state.annotations == [] else ['locked-input']),
|
651 |
+
outputs=annot_text_input
|
652 |
+
)
|
653 |
+
|
654 |
+
undo_btn.click(undo, None, [display_area, msg_output]).then(
|
655 |
+
lambda: gr.update(interactive=state.annotations == [], elem_classes=[] if state.annotations == [] else ['locked-input']),
|
656 |
+
outputs=annot_text_input
|
657 |
+
)
|
658 |
+
|
659 |
+
download_btn.click(download_annotations, [annotator_id, discharge_summary_id], download_file) # Output to download_file
|
660 |
+
refresh_btn.click(refresh_annotations, [annotator_id, discharge_summary_id], annotations_json) # No change in functionality
|
661 |
+
|
662 |
+
|
663 |
+
clear_btn.click(clear_annotations, outputs=[annot_text_input, display_area])
|
664 |
+
|
665 |
+
predict_btn.click(apply_predictions, annot_text_input, display_area).then(
|
666 |
+
lambda: gr.update(interactive=state.annotations == [], elem_classes=[] if state.annotations == [] else ['locked-input']),
|
667 |
+
outputs=text_input
|
668 |
+
)
|
669 |
+
|
670 |
text_input.submit(process, inputs=text_input, outputs=text_out)
|
671 |
text_btn.click(process, inputs=text_input, outputs=text_out)
|
672 |
upload.change(update_inputs, inputs=upload, outputs=sum_inputs)
|
|
|
673 |
ex_add.click(add_ex, inputs=sum_inputs, outputs=sum_inputs)
|
674 |
ex_sub.click(sub_ex, inputs=sum_inputs, outputs=sum_inputs)
|
|
|
|
|
|
|
675 |
filter_button.click(filter_timeline, inputs=[decision_type, start_date, end_date], outputs=timeline_plot)
|
676 |
+
summary_button.click(generate_summary, inputs=sum_inputs, outputs=[summary_output, timeline_plot])
|
677 |
demo.launch(share=True)
|
demo_assets.py
CHANGED
@@ -1,3 +1,7 @@
|
|
|
|
|
|
|
|
|
|
1 |
sum_examples = [
|
2 |
[['examples/note%d.txt'%i for i in range(1,n)]]
|
3 |
for n in range(5,1, -1)
|
@@ -18,3 +22,199 @@ text_examples = [
|
|
18 |
"..."
|
19 |
"Call your PCP or return to the ED for fevers/chills/shakes, chest pain, shortness of breath, pain at the site of your dialysis catheter, nausea, vomiting, or swelling in your legs/feet. "]
|
20 |
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
OTHERS_ID = 18
|
2 |
+
SUM_INPUTS = 20
|
3 |
+
OUTPUT_PATH = 'meddec_annots.json'
|
4 |
+
|
5 |
sum_examples = [
|
6 |
[['examples/note%d.txt'%i for i in range(1,n)]]
|
7 |
for n in range(5,1, -1)
|
|
|
22 |
"..."
|
23 |
"Call your PCP or return to the ED for fevers/chills/shakes, chest pain, shortness of breath, pain at the site of your dialysis catheter, nausea, vomiting, or swelling in your legs/feet. "]
|
24 |
]
|
25 |
+
|
26 |
+
keys = ['c', 'g', 'p', 't', 'd', 'p', 'e', 'f', 'a', 'l']
|
27 |
+
categories = ['Contact related', 'Gathering additional information', 'Defining problem',
|
28 |
+
'Treatment goal', 'Drug related', 'Therapeutic procedure related', 'Evaluating test result',
|
29 |
+
'Deferment', 'Advice and precaution', 'Legal and insurance related']
|
30 |
+
det_desc = ['Admit, discharge, follow-up, referral',
|
31 |
+
'Ordering test, consulting colleague, seeking external information',
|
32 |
+
'Diagnostic conclusion, evaluation of health state, etiological inference, prognostic judgment',
|
33 |
+
'Quantitative or qualitative',
|
34 |
+
'Start, stop, alter, maintain, refrain',
|
35 |
+
'Start, stop, alter, maintain, refrain',
|
36 |
+
'Positive, negative, ambiguous test results',
|
37 |
+
'Transfer responsibility, wait and see, change subject',
|
38 |
+
'Advice or precaution',
|
39 |
+
'Sick leave, drug refund, insurance, disability']
|
40 |
+
unicode_symbols = [
|
41 |
+
"\U0001F91D", # Handshake
|
42 |
+
"\U0001F50D", # Magnifying glass
|
43 |
+
"\U0001F9E9", # Puzzle piece
|
44 |
+
"\U0001F3AF", # Target
|
45 |
+
"\U0001F48A", # Pill
|
46 |
+
"\U00002702", # Surgical scissors
|
47 |
+
"\U0001F9EA", # Test tube
|
48 |
+
"\U000023F0", # Alarm clock
|
49 |
+
"\U000026A0", # Warning sign
|
50 |
+
"\U0001F4C4" # Document
|
51 |
+
]
|
52 |
+
|
53 |
+
colors = ['#8dd3c7', '#ffffb3', '#bebada', '#fb8072', '#80b1d3', '#fdb462', '#b3de69', '#fccde5', '#d9d9d9', '#bc80bd']
|
54 |
+
|
55 |
+
color_map = {cat: colors[i] for i,cat in enumerate(categories)}
|
56 |
+
|
57 |
+
|
58 |
+
annotator_desc = """
|
59 |
+
## Interactive Text Annotator
|
60 |
+
|
61 |
+
This tool allows you to manually annotate medical text for detailed analysis and model training. Follow these steps:
|
62 |
+
|
63 |
+
1. **Input Text:** Enter or paste the text you want to annotate into the text box. You can also use the provided examples.
|
64 |
+
2. **Generate Predictions (Optional):** Click "Generate Predictions" to pre-annotate the text using a machine learning model. This can serve as a starting point, which you can then review and modify. Note: Generating predictions will lock the input textbox to prevent accidental edits.
|
65 |
+
3. **Select Text:** In the **output box below**, highlight the portion of the text you wish to annotate.
|
66 |
+
4. **Annotate:** Click one of the category buttons to apply the annotation to the selected text. The selected text will be highlighted with the corresponding color.
|
67 |
+
5. **Remove/Undo:**
|
68 |
+
* **Remove (q):** Click "Remove" or press 'q' to remove the annotation from the selected text.
|
69 |
+
* **Undo (z):** Click "Undo" or press 'z' to revert the last annotation action.
|
70 |
+
6. **Download/View Annotations:**
|
71 |
+
* **Annotator ID & Discharge Summary ID:** Enter an identifier for the annotator and an optional ID for the discharge summary. These IDs will be included in the downloaded JSON data.
|
72 |
+
* **Refresh Annotations:** Click "Refresh Annotations" to display the current annotations in JSON format.
|
73 |
+
* **Download Annotated Text:** Click "Download Annotated Text" to save your annotations as a JSON file.
|
74 |
+
7. **Clear Annotations:** Click "Clear Annotations" to remove all annotations and start over. This will also unlock the input textbox.
|
75 |
+
|
76 |
+
|
77 |
+
**Keyboard Shortcuts:**
|
78 |
+
|
79 |
+
* Use the keys indicated on the buttons to annotate selections.
|
80 |
+
* `q`: Remove annotation
|
81 |
+
* `z`: Undo last action
|
82 |
+
"""
|
83 |
+
|
84 |
+
label_desc = """
|
85 |
+
## Label a Clinical Note
|
86 |
+
|
87 |
+
This tool allows you to quickly identify and categorize key clinical decisions within a single clinical note.
|
88 |
+
|
89 |
+
1. **Input Text:** Enter or paste the clinical note into the text box. You can also select from the provided examples.
|
90 |
+
2. **Run:** Click the "Run" button to process the text.
|
91 |
+
3. **View Results:** The labeled output will be displayed, highlighting the identified clinical decisions with color-coded categories.
|
92 |
+
|
93 |
+
This provides a simplified overview of the patient's history and key decisions at a glance.
|
94 |
+
"""
|
95 |
+
vis_desc = """
|
96 |
+
## Patient History Visualization
|
97 |
+
|
98 |
+
This tool allows you to visualize the timeline of clinical decisions across multiple clinical notes for a single patient.
|
99 |
+
|
100 |
+
1. **Upload Clinical Notes:** Upload multiple clinical notes in text format. You can also select from the provided examples.
|
101 |
+
2. **Add/Remove Notes:** Use the "+" and "-" buttons to add or remove input fields for additional notes.
|
102 |
+
3. **Filter Timeline:**
|
103 |
+
* **Decision Type:** Select one or more decision types from the dropdown menu to filter the timeline. Select "All" to display all decision types.
|
104 |
+
* **Start/End Date:** Enter a start and end date (MM/DD/YYYY) to filter the timeline by date range.
|
105 |
+
* **Filter:** Click the "Filter Timeline" button to apply the selected filters.
|
106 |
+
4. **Generate Summary:** Click the "Generate Summary" button to generate a summarized view of the clinical decisions organized by date and category. This will also update the timeline plot.
|
107 |
+
5. **View Timeline:** The interactive timeline plot displays the clinical decisions over time. Hover over each data point to view details.
|
108 |
+
|
109 |
+
This visualization helps to understand the progression of clinical decisions and identify trends in patient care.
|
110 |
+
"""
|
111 |
+
|
112 |
+
|
113 |
+
title = """
|
114 |
+
<h1 style="text-align: center;">Medical Decisions Extraction, Visualization, and Annotation</h1>
|
115 |
+
|
116 |
+
<p style="font-size:1.2em;">This application offers three interactive tools for working with clinical text data:</p>
|
117 |
+
<p style="font-size:1.2em;">1. <b>Label a Clinical Note:</b> Process individual notes and receive highlighted key clinical decisions.</br>
|
118 |
+
2. <b>Patient History Visualization:</b> Upload multiple notes to visualize the timeline of decisions.</br>
|
119 |
+
3. <b>Interactive Text Annotator:</b> Manually annotate text for detailed analysis and model training.</p>
|
120 |
+
"""
|
121 |
+
|
122 |
+
desc = '### Decision Categories\n'
|
123 |
+
desc += '| | |\n| --- | --- |\n'
|
124 |
+
for i,cat in enumerate(categories):
|
125 |
+
desc += f'| {unicode_symbols[i]} **{cat}** | {det_desc[i]}|\n'
|
126 |
+
|
127 |
+
shortcut_js_template = """
|
128 |
+
<script>
|
129 |
+
function shortcuts(e) {
|
130 |
+
switch (e.target.tagName.toLowerCase()) {
|
131 |
+
case "input":
|
132 |
+
case "textarea":
|
133 |
+
break;
|
134 |
+
default:
|
135 |
+
if (e.key.toLowerCase() === 'q') {
|
136 |
+
document.getElementById('btn_q').click();
|
137 |
+
} else if (e.key.toLowerCase() === 'z') {
|
138 |
+
document.getElementById('btn_z').click();
|
139 |
+
} else {
|
140 |
+
const buttonKeys = %s;
|
141 |
+
for (const key of buttonKeys) {
|
142 |
+
if (e.key.toLowerCase() === key) {
|
143 |
+
document.getElementById(`btn_${key}`).click();
|
144 |
+
break;
|
145 |
+
}
|
146 |
+
}
|
147 |
+
}
|
148 |
+
}
|
149 |
+
}
|
150 |
+
document.addEventListener('keypress', shortcuts, false);
|
151 |
+
</script>
|
152 |
+
"""
|
153 |
+
|
154 |
+
# JavaScript for handling text selection
|
155 |
+
select_js = """
|
156 |
+
function(x) {
|
157 |
+
const element = document.getElementById('annotated-text');
|
158 |
+
if (!element) {
|
159 |
+
console.log("Element not found");
|
160 |
+
return [0, 0];
|
161 |
+
}
|
162 |
+
|
163 |
+
const selection = window.getSelection();
|
164 |
+
if (!selection.rangeCount) {
|
165 |
+
console.log("No selection found");
|
166 |
+
return [0, 0];
|
167 |
+
}
|
168 |
+
|
169 |
+
const range = selection.getRangeAt(0);
|
170 |
+
|
171 |
+
if (!element.contains(range.commonAncestorContainer)) {
|
172 |
+
console.log("Selection not within element");
|
173 |
+
return [0, 0];
|
174 |
+
}
|
175 |
+
|
176 |
+
const start = getCharacterOffset(element, range.startContainer, range.startOffset);
|
177 |
+
const end = getCharacterOffset(element, range.endContainer, range.endOffset);
|
178 |
+
|
179 |
+
|
180 |
+
console.log("Selection:", start, end);
|
181 |
+
return [start, end];
|
182 |
+
|
183 |
+
|
184 |
+
function getCharacterOffset(root, node, offset) {
|
185 |
+
let currentOffset = 0;
|
186 |
+
|
187 |
+
const iterator = document.createNodeIterator(root, NodeFilter.SHOW_TEXT);
|
188 |
+
let currentNode;
|
189 |
+
|
190 |
+
while (currentNode = iterator.nextNode()) {
|
191 |
+
if (currentNode === node) {
|
192 |
+
return currentOffset + offset;
|
193 |
+
}
|
194 |
+
currentOffset += currentNode.textContent.length;
|
195 |
+
}
|
196 |
+
return currentOffset;
|
197 |
+
}
|
198 |
+
}
|
199 |
+
"""
|
200 |
+
|
201 |
+
css="""
|
202 |
+
.category-legend {border:1px dashed black;}
|
203 |
+
.text-sm {font-size: 1.5rem; line-height: 200%;}
|
204 |
+
.gr-sample-textbox {width: 1000px; white-space: nowrap; overflow: hidden; text-overflow: ellipsis;}
|
205 |
+
.gallery {width: 1000px; white-space: nowrap; overflow: hidden; text-overflow: ellipsis;}
|
206 |
+
.text-limit label textarea {height: 150px !important; overflow: scroll; }
|
207 |
+
.text-gray-500 {color: #111827; font-weight: 600; font-size: 1.25em; margin-top: 1.6em; margin-bottom: 0.6em; 'line-height: 1.6;}
|
208 |
+
#sum-out {border: 2px solid #007bff; padding: 20px; border-radius: 10px; box-shadow: 0 0 10px rgba(0, 0, 0, 0.1); max-height: 300px; overflow-y: auto;}
|
209 |
+
#annotated-text { font-family: monospace; padding: 10px; border: 1px solid #ccc; }
|
210 |
+
#refresh_btn { max-width: 20px; max-height: 40px;}
|
211 |
+
.locked-input { background-color: #f0f0f0; border: 1px solid #ccc; pointer-events: none; }
|
212 |
+
body {
|
213 |
+
--text-sm: 12px;
|
214 |
+
--text-md: 16px;
|
215 |
+
--text-lg: 18px;
|
216 |
+
--input-text-size: 16px;
|
217 |
+
--section-text-size: 16px;
|
218 |
+
--input-background: --neutral-50;
|
219 |
+
}
|
220 |
+
"""
|