Spaces:
Running
Running
File size: 10,047 Bytes
79b0bef | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 | """
src/etl/transform.py
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
Data transformation layer β the "T" in ETL.
Responsibilities
ββββββββββββββββ
- Clean and validate raw clinical notes
- Derive severity labels using weak supervision
- Engineer features needed by the NLP models
- Produce the final, analysis-ready DataFrame
Weak supervision for severity labels
βββββββββββββββββββββββββββββββββββββ
MTSamples has no severity labels, so we derive them using a set
of keyword rules applied to the transcription text. This is
"weak supervision" β the labels are noisy but good enough to
train a classifier that learns to generalise beyond the rules.
The three severity tiers:
critical β ICU, ventilator, cardiac arrest, septic shock,
stroke, code, emergent, life-threatening
urgent β emergency, acute, admitted, inpatient, fever,
infection, pain scale > 6, SOB, unstable
routine β elective, outpatient, follow-up, stable, chronic
management, wellness, screening
Notes that match no keywords default to "routine".
ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
"""
from __future__ import annotations
import re
import pandas as pd
from src.utils.logger import get_logger
from src.utils.text_utils import clean_clinical_text, word_count
logger = get_logger(__name__)
# ββ Severity keyword rules ββββββββββββββββββββββββββββββββββββββββ
# Evaluated in order: critical β urgent β routine.
# The first tier that matches wins.
_SEVERITY_RULES: dict[str, list[str]] = {
"critical": [
r"\bicu\b", r"\bintensive care\b", r"\bventilat",
r"\bcardiac arrest\b", r"\bseptic shock\b",
r"\brespiratory failure\b", r"\bmulti.?organ failure\b",
r"\bcode blue\b", r"\bemergent intubat",
r"\blife.?threatening\b", r"\bcritically ill\b",
r"\bstroke\b", r"\bcva\b", r"\binfarction\b",
r"\bpulseless\b", r"\bunresponsive\b",
],
"urgent": [
r"\bemergency\b", r"\bemer\b", r"(?<!no )\bacute\b",
r"\badmitted\b", r"\binpatient\b",
r"\bhigh fever\b", r"\btemperature of [3-4]\d\.",
r"\bsevere pain\b", r"\bpain.{0,20}[7-9][/\s]10\b",
r"\bshortness of breath\b", r"\bsob\b",
r"\bunstable\b", r"\brapidly\b",
r"\binfection\b", r"\bsepsis\b",
r"\btrauma\b", r"\bfracture\b",
r"\bhemorrhage\b", r"\bbleeding\b",
r"\bchest pain\b",
],
"routine": [
r"\belective\b", r"\boutpatient\b",
r"\bfollow.?up\b", r"\bstable\b",
r"\bchronic\b", r"\bwellness\b",
r"\bscreening\b", r"\bpreventive\b",
r"\bmanagement of\b", r"\broutine\b",
r"\bno acute\b", r"\bnormal\b",
],
}
def clean_notes(df: pd.DataFrame) -> pd.DataFrame:
"""Clean and validate a DataFrame of clinical notes.
Applies the following steps:
1. Drop rows with missing or empty transcription text
2. Normalise the transcription through the text cleaner
3. Drop duplicates (same transcription appearing twice)
4. Reset the index
Args:
df: Raw DataFrame from the extraction layer. Must contain
a ``transcription`` column.
Returns:
Cleaned DataFrame. Row count will be <= input row count.
Raises:
KeyError: If ``transcription`` column is absent.
"""
if "transcription" not in df.columns:
raise KeyError(
"DataFrame must have a 'transcription' column. "
f"Got: {list(df.columns)}"
)
original_count = len(df)
# Drop rows where the note text is missing or whitespace-only
df = df.dropna(subset=["transcription"]).copy()
df = df[df["transcription"].str.strip().astype(bool)].copy()
after_drop = len(df)
if after_drop < original_count:
logger.info(
"Dropped %d rows with missing transcription text",
original_count - after_drop,
)
# Apply clinical text cleaning
df["transcription"] = df["transcription"].apply(clean_clinical_text)
# Remove duplicates β same note appearing more than once
before_dedup = len(df)
df = df.drop_duplicates(subset=["transcription"]).copy()
if len(df) < before_dedup:
logger.info(
"Removed %d duplicate notes", before_dedup - len(df)
)
df = df.reset_index(drop=True)
logger.info("Clean notes: %d rows remaining", len(df))
return df
def derive_severity_labels(df: pd.DataFrame) -> pd.DataFrame:
"""Assign a severity label to each clinical note.
Uses weak supervision (keyword matching on transcription text)
to derive one of three labels: ``routine``, ``urgent``,
``critical``.
Labels are noisy β they are intended as training signal, not
ground truth. The classifier learns to generalise beyond the
keyword patterns.
Args:
df: DataFrame with a ``transcription`` column.
Returns:
DataFrame with a new ``severity`` column appended.
Distribution is logged at INFO level.
Note:
Notes that match no keywords are labelled ``"routine"``
(the most common clinical scenario).
"""
def _assign_label(text: str) -> str:
"""Classify one note by matching against keyword tiers."""
if not isinstance(text, str):
return "routine"
lower = text.lower()
for tier in ("critical", "urgent", "routine"):
patterns = _SEVERITY_RULES[tier]
if any(re.search(p, lower) for p in patterns):
return tier
# Default: routine
return "routine"
logger.info("Deriving severity labels via weak supervision...")
df = df.copy()
df["severity"] = df["transcription"].apply(_assign_label)
# Log the distribution so the user can spot class imbalance
dist = df["severity"].value_counts()
for label, count in dist.items():
pct = count / len(df) * 100
logger.info(" %-10s %4d (%.1f%%)", label, count, pct)
return df
def engineer_features(df: pd.DataFrame) -> pd.DataFrame:
"""Add derived columns used by downstream models and analysis.
Features added:
- ``word_count`` : number of words in the transcription
- ``char_count`` : character count
- ``has_keywords`` : whether the keywords column is populated
- ``note_type`` : normalised note type (lowercase)
- ``specialty_clean``: specialty with whitespace normalised
Args:
df: DataFrame with at minimum a ``transcription`` column.
Returns:
DataFrame with additional feature columns.
"""
df = df.copy()
df["word_count"] = df["transcription"].apply(word_count)
df["char_count"] = df["transcription"].str.len().fillna(0).astype(int)
if "keywords" in df.columns:
df["has_keywords"] = (
df["keywords"].notna() & df["keywords"].str.strip().astype(bool)
)
else:
df["has_keywords"] = False
if "sample_name" in df.columns:
df["note_type"] = (
df["sample_name"]
.fillna("unknown")
.str.lower()
.str.strip()
)
if "specialty" in df.columns:
df["specialty_clean"] = (
df["specialty"]
.fillna("unknown")
.str.strip()
.str.title()
)
return df
def filter_minimum_length(
df: pd.DataFrame,
min_words: int = 30,
) -> pd.DataFrame:
"""Remove notes that are too short to be clinically useful.
Very short notes (< 30 words by default) tend to be incomplete
transcriptions or metadata entries rather than real clinical notes.
Including them degrades model training.
Args:
df: DataFrame with a ``word_count`` column. Call
:func:`engineer_features` first if needed.
min_words: Minimum word count to retain a note.
Returns:
Filtered DataFrame.
"""
if "word_count" not in df.columns:
df = engineer_features(df)
before = len(df)
df = df[df["word_count"] >= min_words].copy()
dropped = before - len(df)
if dropped:
logger.info(
"Filtered %d notes shorter than %d words", dropped, min_words
)
return df.reset_index(drop=True)
def prepare_for_training(df: pd.DataFrame) -> pd.DataFrame:
"""Run the full transformation pipeline for model training.
Convenience wrapper that runs all transformation steps in the
correct order.
Steps:
1. :func:`clean_notes`
2. :func:`engineer_features`
3. :func:`filter_minimum_length`
4. :func:`derive_severity_labels`
Args:
df: Raw DataFrame from the extraction layer.
Returns:
Fully transformed, analysis-ready DataFrame.
"""
logger.info("Running full transformation pipeline...")
df = clean_notes(df)
df = engineer_features(df)
df = filter_minimum_length(df)
df = derive_severity_labels(df)
logger.info(
"Transformation complete: %d notes ready for training", len(df)
)
return df
def prepare_for_inference(text: str) -> str:
"""Prepare a single raw text string for model inference.
Used by the API and Streamlit demo when a user pastes clinical
text. Applies the same cleaning as training data without the
label derivation.
Args:
text: Raw clinical note text from the user.
Returns:
Cleaned text string ready for NER and classification.
"""
return clean_clinical_text(text, expand_abbreviations=True)
|