Spaces:
Sleeping
Sleeping
File size: 18,283 Bytes
9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 8514865 9d2afb8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 | """
Baseline classifiers for screening: ASD vs TD vs DD.
Two tasks are run:
(A) Binary: ASD vs non-ASD (TD + DD) -> screening use-case
(B) Multi-class: ASD vs DD vs TD -> differential
Models:
- Logistic Regression
- Random Forest
- Support Vector Machine (RBF)
Evaluation: stratified 5-fold cross-validation.
Outputs:
reports/metrics/classification_results.csv
reports/figures/confusion_matrix_<task>_<model>.png
reports/figures/feature_importance.png
reports/figures/roc_curve_binary.png
"""
from __future__ import annotations
import hashlib
import json
from datetime import date
from pathlib import Path
import joblib
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from sklearn.ensemble import RandomForestClassifier
from sklearn.impute import SimpleImputer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import (
average_precision_score,
brier_score_loss,
ConfusionMatrixDisplay,
RocCurveDisplay,
accuracy_score,
classification_report,
confusion_matrix,
f1_score,
precision_score,
recall_score,
roc_auc_score,
)
from sklearn.model_selection import LeaveOneGroupOut, StratifiedKFold, cross_val_predict
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.svm import SVC
try:
from src.feature_schema import (
FEATURES,
UNCERTAIN_HIGH,
UNCERTAIN_LOW,
feature_schema_rows,
)
except ModuleNotFoundError: # running as `python src/classifier.py`
from feature_schema import (
FEATURES,
UNCERTAIN_HIGH,
UNCERTAIN_LOW,
feature_schema_rows,
)
PROJECT_ROOT = Path(__file__).resolve().parent.parent
DATA_DIR = PROJECT_ROOT / "data"
FIG_DIR = PROJECT_ROOT / "reports" / "figures"
METRIC_DIR = PROJECT_ROOT / "reports" / "metrics"
ARTIFACT_DIR = PROJECT_ROOT / "artifacts"
FIG_DIR.mkdir(parents=True, exist_ok=True)
METRIC_DIR.mkdir(parents=True, exist_ok=True)
ARTIFACT_DIR.mkdir(parents=True, exist_ok=True)
sns.set_theme(style="whitegrid", context="talk")
RANDOM_STATE = 42
MODEL_VERSION = "v0.17.0-trust-dashboard"
def _build_models():
return {
"LogReg": Pipeline([
("imp", SimpleImputer(strategy="median")),
("sc", StandardScaler()),
("clf", LogisticRegression(max_iter=2000,
class_weight="balanced",
random_state=RANDOM_STATE)),
]),
"RandomForest": Pipeline([
("imp", SimpleImputer(strategy="median")),
("clf", RandomForestClassifier(
n_estimators=300,
class_weight="balanced",
random_state=RANDOM_STATE,
)),
]),
"SVM": Pipeline([
("imp", SimpleImputer(strategy="median")),
("sc", StandardScaler()),
("clf", SVC(kernel="rbf", probability=True,
class_weight="balanced",
random_state=RANDOM_STATE)),
]),
}
def _safe_div(num: float, den: float) -> float:
return float(num / den) if den else 0.0
def _binary_metric_row(
y_true: np.ndarray,
y_pred: np.ndarray,
y_proba: np.ndarray,
*,
threshold: float,
) -> dict:
tn, fp, fn, tp = confusion_matrix(y_true, y_pred, labels=[0, 1]).ravel()
uncertain = (y_proba >= UNCERTAIN_LOW) & (y_proba < UNCERTAIN_HIGH)
return {
"accuracy": accuracy_score(y_true, y_pred),
"f1_macro": f1_score(y_true, y_pred, average="macro"),
"roc_auc": roc_auc_score(y_true, y_proba),
"pr_auc": average_precision_score(y_true, y_proba),
"sensitivity": recall_score(y_true, y_pred, pos_label=1, zero_division=0),
"specificity": _safe_div(tn, tn + fp),
"ppv": precision_score(y_true, y_pred, pos_label=1, zero_division=0),
"npv": _safe_div(tn, tn + fn),
"brier_score": brier_score_loss(y_true, y_proba),
"threshold": threshold,
"tp": int(tp),
"fp": int(fp),
"tn": int(tn),
"fn": int(fn),
"uncertain_count": int(uncertain.sum()),
"uncertain_rate": float(uncertain.mean()),
}
def _round_metric_row(row: dict) -> dict:
rounded = {}
for key, value in row.items():
if isinstance(value, (float, np.floating)):
rounded[key] = round(float(value), 4)
else:
rounded[key] = value
return rounded
def _cv_evaluate(X, y, models, task: str, class_order, display_labels):
"""Run 5-fold CV for each model.
class_order: labels as they appear in y (e.g. [0, 1] or ['ASD', 'DD', 'TD'])
display_labels: human-readable names in the same order.
"""
skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=RANDOM_STATE)
rows = []
preds = {}
probs = {}
for name, pipe in models.items():
y_pred = cross_val_predict(pipe, X, y, cv=skf, n_jobs=-1)
preds[name] = y_pred
acc = accuracy_score(y, y_pred)
f1_macro = f1_score(y, y_pred, average="macro")
row = {"task": task, "model": name}
if task == "binary":
y_proba = cross_val_predict(
pipe, X, y, cv=skf, method="predict_proba", n_jobs=-1
)[:, 1]
probs[name] = y_proba
row.update(_binary_metric_row(
np.asarray(y), np.asarray(y_pred), np.asarray(y_proba),
threshold=0.5,
))
else:
row.update({
"accuracy": acc,
"f1_macro": f1_macro,
})
rows.append(_round_metric_row(row))
cm = confusion_matrix(y, y_pred, labels=class_order)
fig, ax = plt.subplots(figsize=(6, 5))
ConfusionMatrixDisplay(cm, display_labels=display_labels).plot(
ax=ax, cmap="Blues", values_format="d", colorbar=False,
)
ax.set_title(f"{task} | {name}\nacc={acc:.3f} f1={f1_macro:.3f}")
fig.tight_layout()
out = FIG_DIR / f"confusion_matrix_{task}_{name}.png"
fig.savefig(out, dpi=150, bbox_inches="tight")
plt.close(fig)
print(f" saved {out.relative_to(PROJECT_ROOT)}")
print(f"\n[{task} / {name}]")
print(classification_report(y, y_pred, labels=class_order,
target_names=display_labels, digits=3))
return rows, preds, probs
def _plot_feature_importance(X, y):
pipe = Pipeline([
("imp", SimpleImputer(strategy="median")),
("clf", RandomForestClassifier(n_estimators=500,
class_weight="balanced",
random_state=RANDOM_STATE)),
])
pipe.fit(X, y)
imp = pipe.named_steps["clf"].feature_importances_
order = np.argsort(imp)[::-1]
feats = np.array(FEATURES)[order]
vals = imp[order]
fig, ax = plt.subplots(figsize=(9, 6))
sns.barplot(x=vals, y=feats, ax=ax, color="#4C72B0")
ax.set_title("Random Forest feature importance (multi-class)")
ax.set_xlabel("Importance")
fig.tight_layout()
out = FIG_DIR / "feature_importance.png"
fig.savefig(out, dpi=150, bbox_inches="tight")
plt.close(fig)
print(f" saved {out.relative_to(PROJECT_ROOT)}")
def _plot_roc_curves(X, y, probs):
fig, ax = plt.subplots(figsize=(7, 6))
for name, p in probs.items():
RocCurveDisplay.from_predictions(y, p, name=name, ax=ax)
ax.plot([0, 1], [0, 1], "k--", alpha=0.4)
ax.set_title("ROC curves - ASD vs non-ASD (5-fold CV)")
fig.tight_layout()
out = FIG_DIR / "roc_curve_binary.png"
fig.savefig(out, dpi=150, bbox_inches="tight")
plt.close(fig)
print(f" saved {out.relative_to(PROJECT_ROOT)}")
def _data_hash(df: pd.DataFrame) -> str:
payload = df.sort_values(["corpus", "participant_id"]).to_csv(index=False)
return hashlib.sha256(payload.encode("utf-8")).hexdigest()[:16]
def _threshold_table(y_true: np.ndarray, y_proba: np.ndarray) -> pd.DataFrame:
rows = []
for threshold in np.round(np.arange(0.05, 0.96, 0.05), 2):
y_pred = (y_proba >= threshold).astype(int)
row = _binary_metric_row(y_true, y_pred, y_proba, threshold=float(threshold))
rows.append(_round_metric_row(row))
return pd.DataFrame(rows)
def _calibration_bins(y_true: np.ndarray, y_proba: np.ndarray, n_bins: int = 10) -> pd.DataFrame:
bins = np.linspace(0, 1, n_bins + 1)
labels = [f"{bins[i]:.1f}-{bins[i + 1]:.1f}" for i in range(n_bins)]
df = pd.DataFrame({"y_true": y_true, "prob_asd": y_proba})
df["bin"] = pd.cut(df["prob_asd"], bins=bins, labels=labels,
include_lowest=True, right=True)
rows = []
for label, group in df.groupby("bin", observed=False):
if group.empty:
continue
rows.append({
"bin": str(label),
"n": int(len(group)),
"predicted_mean": round(float(group["prob_asd"].mean()), 4),
"observed_rate": round(float(group["y_true"].mean()), 4),
})
return pd.DataFrame(rows)
def _decision_curve(y_true: np.ndarray, y_proba: np.ndarray) -> pd.DataFrame:
n = len(y_true)
prevalence = float(np.mean(y_true))
rows = []
for threshold in np.round(np.arange(0.05, 0.96, 0.05), 2):
y_pred = y_proba >= threshold
tp = int(((y_true == 1) & y_pred).sum())
fp = int(((y_true == 0) & y_pred).sum())
odds = threshold / (1 - threshold)
rows.append({
"threshold": float(threshold),
"model_net_benefit": round(tp / n - fp / n * odds, 4),
"treat_all_net_benefit": round(prevalence - (1 - prevalence) * odds, 4),
"treat_none_net_benefit": 0.0,
})
return pd.DataFrame(rows)
def _subgroup_performance(df: pd.DataFrame, y_true: np.ndarray, y_proba: np.ndarray) -> pd.DataFrame:
rows = []
eval_df = df.copy()
eval_df["y_true"] = y_true
eval_df["prob_asd"] = y_proba
eval_df["pred"] = (y_proba >= 0.5).astype(int)
eval_df["age_band"] = pd.cut(
eval_df["age_months"],
bins=[0, 36, 48, 60, 72, 200],
labels=["<36", "36-47", "48-59", "60-71", "72+"],
include_lowest=True,
).astype(str)
for dimension in ["corpus", "sex", "age_band"]:
for value, sub in eval_df.groupby(dimension, dropna=False):
if len(sub) < 5 or sub["y_true"].nunique() < 2:
continue
metrics = _binary_metric_row(
sub["y_true"].to_numpy(),
sub["pred"].to_numpy(),
sub["prob_asd"].to_numpy(),
threshold=0.5,
)
rows.append(_round_metric_row({
"dimension": dimension,
"value": str(value),
"n": len(sub),
**metrics,
}))
return pd.DataFrame(rows)
def _leave_one_corpus_out(df: pd.DataFrame) -> pd.DataFrame:
X = df[FEATURES].values
y = (df["group"] == "ASD").astype(int).values
groups = df["corpus"].values
rows = []
for train_idx, test_idx in LeaveOneGroupOut().split(X, y, groups):
test_corpus = str(groups[test_idx][0])
if len(np.unique(y[test_idx])) < 2:
rows.append({
"held_out_corpus": test_corpus,
"n_test": int(len(test_idx)),
"status": "skipped_single_class",
})
continue
pipe = _build_models()["LogReg"]
pipe.fit(X[train_idx], y[train_idx])
proba = pipe.predict_proba(X[test_idx])[:, 1]
pred = (proba >= 0.5).astype(int)
rows.append(_round_metric_row({
"held_out_corpus": test_corpus,
"n_test": int(len(test_idx)),
"status": "evaluated",
**_binary_metric_row(y[test_idx], pred, proba, threshold=0.5),
}))
return pd.DataFrame(rows)
def _write_model_bundle(df: pd.DataFrame) -> dict:
X = df[FEATURES].values
y = (df["group"] == "ASD").astype(int).values
model = _build_models()["LogReg"]
model.fit(X, y)
bundle = {
"model": model,
"model_version": MODEL_VERSION,
"features": FEATURES,
"thresholds": {
"uncertain_low": UNCERTAIN_LOW,
"uncertain_high": UNCERTAIN_HIGH,
"default_binary": 0.5,
},
"training_metadata": {
"trained_on": date.today().isoformat(),
"n_rows": int(len(df)),
"n_asd": int(y.sum()),
"n_non_asd": int((1 - y).sum()),
"corpora": sorted(df["corpus"].dropna().unique().tolist()),
"data_hash": _data_hash(df),
},
}
out = ARTIFACT_DIR / "screening_model.joblib"
joblib.dump(bundle, out)
print(f" saved {out.relative_to(PROJECT_ROOT)}")
return bundle
def _write_json(path: Path, payload: dict) -> None:
path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
print(f" saved {path.relative_to(PROJECT_ROOT)}")
def main() -> None:
csv_path = DATA_DIR / "combined_features.csv"
df = pd.read_csv(csv_path)
df = df.dropna(subset=["group"])
print(f"Loaded {len(df)} rows. Group counts:")
print(df["group"].value_counts().to_string())
X = df[FEATURES].values
all_rows = []
# ---------------- Binary: ASD vs non-ASD ----------------
print("\n" + "=" * 70)
print("TASK A: Binary ASD (1) vs non-ASD (0)")
print("=" * 70)
y_bin = (df["group"] == "ASD").astype(int).values
rows, preds, probs = _cv_evaluate(
X, y_bin, _build_models(),
task="binary",
class_order=[0, 1],
display_labels=["non-ASD", "ASD"],
)
all_rows.extend(rows)
if probs:
_plot_roc_curves(X, y_bin, probs)
# ---------------- Multi-class: ASD / DD / TD ----------------
print("\n" + "=" * 70)
print("TASK B: Multi-class ASD vs DD vs TD")
print("=" * 70)
multi_df = df[df["group"].isin(["ASD", "DD", "TD"])]
X_m = multi_df[FEATURES].values
y_m = multi_df["group"].astype(str).to_numpy()
rows, _, _ = _cv_evaluate(
X_m, y_m, _build_models(),
task="multiclass",
class_order=["ASD", "DD", "TD"],
display_labels=["ASD", "DD", "TD"],
)
all_rows.extend(rows)
_plot_feature_importance(X_m, y_m)
# Save results
results_df = pd.DataFrame(all_rows)
out = METRIC_DIR / "classification_results.csv"
results_df.to_csv(out, index=False)
print(f"\n[saved] {out.relative_to(PROJECT_ROOT)}")
# Model Trust Dashboard inputs focus on LogReg, the selected interpretable
# screening model. These CSVs are static assets for project_dashboard/.
logreg_prob = probs.get("LogReg")
logreg_pred = preds.get("LogReg")
if logreg_prob is not None and logreg_pred is not None:
pred_df = df[[
"participant_id", "corpus", "group", "sex", "age_months",
]].copy()
pred_df["y_true"] = y_bin
pred_df["prob_asd"] = np.round(logreg_prob, 6)
pred_df["pred_050"] = logreg_pred
pred_df["uncertainty_zone"] = np.select(
[
logreg_prob < UNCERTAIN_LOW,
(logreg_prob >= UNCERTAIN_LOW) & (logreg_prob < UNCERTAIN_HIGH),
logreg_prob >= UNCERTAIN_HIGH,
],
["low", "uncertain", "high"],
default="unknown",
)
pred_out = METRIC_DIR / "binary_oof_predictions.csv"
pred_df.to_csv(pred_out, index=False)
print(f"[saved] {pred_out.relative_to(PROJECT_ROOT)}")
threshold_out = METRIC_DIR / "threshold_metrics.csv"
_threshold_table(y_bin, logreg_prob).to_csv(threshold_out, index=False)
print(f"[saved] {threshold_out.relative_to(PROJECT_ROOT)}")
calibration_out = METRIC_DIR / "calibration_bins.csv"
_calibration_bins(y_bin, logreg_prob).to_csv(calibration_out, index=False)
print(f"[saved] {calibration_out.relative_to(PROJECT_ROOT)}")
dca_out = METRIC_DIR / "decision_curve.csv"
_decision_curve(y_bin, logreg_prob).to_csv(dca_out, index=False)
print(f"[saved] {dca_out.relative_to(PROJECT_ROOT)}")
subgroup_out = METRIC_DIR / "subgroup_performance.csv"
_subgroup_performance(df, y_bin, logreg_prob).to_csv(subgroup_out, index=False)
print(f"[saved] {subgroup_out.relative_to(PROJECT_ROOT)}")
loco_out = METRIC_DIR / "leave_one_corpus_out.csv"
_leave_one_corpus_out(df).to_csv(loco_out, index=False)
print(f"[saved] {loco_out.relative_to(PROJECT_ROOT)}")
feature_schema_out = ARTIFACT_DIR / "feature_schema.json"
_write_json(feature_schema_out, {
"features": FEATURES,
"feature_docs": feature_schema_rows(),
"thresholds": {
"uncertain_low": UNCERTAIN_LOW,
"uncertain_high": UNCERTAIN_HIGH,
},
})
bundle = _write_model_bundle(df)
model_card_out = ARTIFACT_DIR / "model_card.json"
_write_json(model_card_out, {
"model_version": MODEL_VERSION,
"model_type": "Logistic Regression with median imputation and standard scaling",
"intended_use": "ASD screening support and research demo; not diagnostic.",
"not_intended_use": "Autonomous diagnosis, emergency triage, or replacement for clinician assessment.",
"inputs": FEATURES,
"training_metadata": bundle["training_metadata"],
"thresholds": bundle["thresholds"],
"reporting_guidance": [
"TRIPOD+AI prediction model reporting",
"DECIDE-AI early clinical decision-support evaluation",
"Model card and dataset-card transparency",
],
"clinical_caveats": [
"TalkBank/ASDBank cohorts are not a Thai external validation set.",
"Audio-derived predictions require transcript QA and feature-drift checks.",
"Probability estimates require calibration review before clinical use.",
],
})
print("\n=== SUMMARY ===")
print(results_df.to_string(index=False))
if __name__ == "__main__":
main()
|