Datasets:
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
acupuncture
traditional-chinese-medicine
multi-label-classification
conformal-prediction
knowledge-graph
health
License:
File size: 14,184 Bytes
2803982 | 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 | #!/usr/bin/env python3
"""AcuBench evaluation harness (self-contained).
Ports the metric definitions from the research repo's `acubench/src/metrics.py`
with NO external import beyond numpy/scikit-learn. Scores a predictions file
against a gold file over the fixed 361-point WHO label space.
METRIC SUITE
------------
* jaccard_mean -- mean per-sample Jaccard(pred_set, gold_set)
* f1_macro -- mean per-sample (example-based) F1
* f1_micro -- pooled TP/FP/FN across all samples, one global F1
* precision_at_k / recall_at_k / ndcg_at_k for k in --ks (default 5,10,20)
* prauc_mean -- mean per-sample average precision (needs scores)
* invalid_combination_rate -- structural meridian-scatter PROXY (needs --who)
INPUT FORMATS
-------------
predictions file (JSONL, one object per line). Each object needs an "id" and
at least one of:
{"id": 12, "acupoints": ["LU1","LU7"]} # a predicted set
{"id": 12, "scores": {"LU1": 0.9, "LU7": 0.8, ...}} # per-point scores
{"id": 12, "acupoints": [...], "scores": {...}} # both (recommended)
- "acupoints" drives the SET metrics (Jaccard / F1-micro / F1-macro) and the
meridian-validity proxy.
- "scores" drives the RANKING metrics (P@k / R@k / NDCG@k / PRAUC). If a
row has no "scores", ranking falls back to its "acupoints" ranked
alphabetically (a weak but deterministic proxy -- provide real scores for
meaningful ranking metrics). PRAUC is only reported when at least one row
supplies "scores".
- If a row has "scores" but no "acupoints", the predicted set is taken as
every point with score > --score-threshold (default 0.5).
gold file (JSONL). Each object needs "id" and "acupoints". The AcuBench
`acubench.jsonl` produced by build_acubench.py works directly as the gold
file (it also carries "split", so pass --split test to score only the test
rows). The shipped sample_labels.jsonl uses "symptoms" instead of an "id";
pass --gold-key symptoms to key rows by their symptom string, and key your
predictions the same way (use the symptom string as "id").
USAGE
-----
python3 eval.py --pred preds.jsonl --gold acubench.jsonl --who who_acupoints.csv
python3 eval.py --pred preds.jsonl --gold acubench.jsonl --split test --who who_acupoints.csv
python3 eval.py --pred preds.jsonl --gold sample_labels.jsonl --gold-key symptoms
Prints a JSON metric dict to stdout (use --out to also write it to a file).
TINY EXAMPLE (predictions JSONL)
{"id": 0, "acupoints": ["CV15"], "scores": {"CV15": 0.9, "LU1": 0.1}}
{"id": 1, "acupoints": ["LU1", "LU2"], "scores": {"LU1": 0.8, "LU2": 0.7}}
"""
import argparse
import csv
import json
import math
from collections import defaultdict
from typing import Dict, Iterable, List, Optional, Sequence, Tuple
import numpy as np
from sklearn.metrics import average_precision_score
# --------------------------------------------------------------------------
# Set metrics (ported verbatim from src/metrics.py).
# --------------------------------------------------------------------------
def jaccard(pred: Iterable[str], gold: Iterable[str]) -> float:
pred, gold = set(pred), set(gold)
if not pred and not gold:
return 1.0
union = pred | gold
if not union:
return 1.0
return len(pred & gold) / len(union)
def precision_recall_f1(pred: Iterable[str], gold: Iterable[str]) -> Tuple[float, float, float]:
pred, gold = set(pred), set(gold)
tp = len(pred & gold)
precision = tp / len(pred) if pred else (1.0 if not gold else 0.0)
recall = tp / len(gold) if gold else (1.0 if not pred else 0.0)
f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
return precision, recall, f1
def f1_macro(pred_sets: Sequence[Iterable[str]], gold_sets: Sequence[Iterable[str]]) -> float:
if not pred_sets:
return 0.0
scores = [precision_recall_f1(p, g)[2] for p, g in zip(pred_sets, gold_sets)]
return sum(scores) / len(scores)
def f1_micro(pred_sets: Sequence[Iterable[str]], gold_sets: Sequence[Iterable[str]]) -> float:
tp = fp = fn = 0
for pred, gold in zip(pred_sets, gold_sets):
pred, gold = set(pred), set(gold)
tp += len(pred & gold)
fp += len(pred - gold)
fn += len(gold - pred)
precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
return 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0.0
# --------------------------------------------------------------------------
# Ranking metrics.
# --------------------------------------------------------------------------
def precision_at_k(ranked_points: Sequence[str], gold: Iterable[str], k: int) -> float:
gold = set(gold)
top_k = ranked_points[:k]
if not top_k:
return 0.0
hits = sum(1 for p in top_k if p in gold)
return hits / len(top_k)
def recall_at_k(ranked_points: Sequence[str], gold: Iterable[str], k: int) -> float:
gold = set(gold)
if not gold:
return 1.0
top_k = ranked_points[:k]
hits = sum(1 for p in top_k if p in gold)
return hits / len(gold)
def ndcg_at_k(ranked_points: Sequence[str], gold: Iterable[str], k: int) -> float:
gold = set(gold)
top_k = ranked_points[:k]
dcg = sum((1.0 if p in gold else 0.0) / math.log2(i + 2) for i, p in enumerate(top_k))
ideal_hits = min(len(gold), k)
idcg = sum(1.0 / math.log2(i + 2) for i in range(ideal_hits))
return dcg / idcg if idcg > 0 else 0.0
def average_precision(scores: Dict[str, float], gold: Iterable[str]) -> float:
gold = set(gold)
if not scores:
return 0.0
points = list(scores.keys())
y_true = [1 if p in gold else 0 for p in points]
if sum(y_true) == 0:
return 0.0
y_score = [scores[p] for p in points]
return float(average_precision_score(y_true, y_score))
def _ranked_from_scores(scores: Dict[str, float]) -> List[str]:
"""Rank points by score descending; deterministic alphabetical tie-break."""
return [p for p, _ in sorted(scores.items(), key=lambda kv: (-kv[1], kv[0]))]
# --------------------------------------------------------------------------
# Meridian-validity heuristic (structural PROXY, NOT a clinical judgment).
# --------------------------------------------------------------------------
def is_valid_combination(
pred: Iterable[str],
point_meridian_map: Dict[str, str],
max_distinct_meridian_ratio: float = 0.8,
min_size_for_check: int = 3,
) -> bool:
"""Documented, rule-based PROXY for prescription plausibility -- NOT a
clinical validity check. Flags a set of size >= min_size_for_check as
'implausible' (structurally scattered) when distinct_meridians / n_points
exceeds max_distinct_meridian_ratio. See src/metrics.py for full rationale.
"""
pred = list(pred)
n = len(pred)
if n < min_size_for_check:
return True
meridians = [point_meridian_map[p] for p in pred if p in point_meridian_map]
if not meridians:
return True
distinct = len(set(meridians))
ratio = distinct / n
return ratio <= max_distinct_meridian_ratio
def invalid_combination_rate(
pred_sets: Sequence[Iterable[str]],
point_meridian_map: Dict[str, str],
max_distinct_meridian_ratio: float = 0.8,
min_size_for_check: int = 3,
) -> float:
if not pred_sets:
return 0.0
flags = [
not is_valid_combination(p, point_meridian_map, max_distinct_meridian_ratio, min_size_for_check)
for p in pred_sets
]
return sum(flags) / len(flags)
# --------------------------------------------------------------------------
# Batch summary.
# --------------------------------------------------------------------------
def summarize(
pred_sets: Sequence[Iterable[str]],
gold_sets: Sequence[Iterable[str]],
scores: Optional[Sequence[Optional[Dict[str, float]]]] = None,
point_meridian_map: Optional[Dict[str, str]] = None,
ks: Sequence[int] = (5, 10, 20),
) -> Dict[str, float]:
pred_sets = [set(p) for p in pred_sets]
gold_sets = [set(g) for g in gold_sets]
n = len(pred_sets)
result: Dict[str, float] = {"n_samples": n}
if n == 0:
return result
result["jaccard_mean"] = sum(jaccard(p, g) for p, g in zip(pred_sets, gold_sets)) / n
result["f1_macro"] = f1_macro(pred_sets, gold_sets)
result["f1_micro"] = f1_micro(pred_sets, gold_sets)
# Ranked list per sample: real scores where given, else alphabetical
# fallback over the predicted set (documented weak proxy).
any_scores = scores is not None and any(s for s in scores)
if scores is not None:
ranked_lists = [
_ranked_from_scores(s) if s else sorted(pred_sets[i])
for i, s in enumerate(scores)
]
else:
ranked_lists = [sorted(p) for p in pred_sets]
for k in ks:
result[f"precision_at_{k}"] = sum(
precision_at_k(r, g, k) for r, g in zip(ranked_lists, gold_sets)
) / n
result[f"recall_at_{k}"] = sum(
recall_at_k(r, g, k) for r, g in zip(ranked_lists, gold_sets)
) / n
result[f"ndcg_at_{k}"] = sum(
ndcg_at_k(r, g, k) for r, g in zip(ranked_lists, gold_sets)
) / n
if any_scores:
result["prauc_mean"] = sum(
average_precision(s, g) for s, g in zip(scores, gold_sets) if s
) / sum(1 for s in scores if s)
if point_meridian_map is not None:
result["invalid_combination_rate"] = invalid_combination_rate(pred_sets, point_meridian_map)
return result
# --------------------------------------------------------------------------
# I/O helpers.
# --------------------------------------------------------------------------
def load_jsonl(path: str) -> List[dict]:
rows = []
with open(path, encoding="utf-8") as f:
for line in f:
line = line.strip()
if line:
rows.append(json.loads(line))
return rows
def load_point_meridian_map(who_csv: str) -> Dict[str, str]:
mapping = {}
with open(who_csv, newline="") as f:
for row in csv.DictReader(f):
mapping[row["point_code"]] = row["meridian"]
return mapping
def _row_key(row: dict, key: str):
"""Key a row: by 'id' (default), or by the join column (e.g. 'symptoms').
'symptoms' may be a list -> use its single element / joined string."""
if key == "id":
return row["id"]
val = row.get(key)
if isinstance(val, list):
return val[0] if len(val) == 1 else "|".join(map(str, val))
return val
def align(
preds: List[dict],
golds: List[dict],
gold_key: str,
score_threshold: float,
) -> Tuple[List[List[str]], List[List[str]], List[Optional[Dict[str, float]]]]:
"""Join predictions to gold rows on the key column. Returns aligned
(pred_sets, gold_sets, scores_per_row) over the intersection of keys."""
pred_by_key = {}
for r in preds:
# Predictions key on "id" by default; when gold is keyed by another
# column, predictions should carry that value in "id".
k = r.get("id")
if k is None and gold_key != "id":
k = _row_key(r, gold_key)
pred_by_key[k] = r
pred_sets: List[List[str]] = []
gold_sets: List[List[str]] = []
scores: List[Optional[Dict[str, float]]] = []
matched = 0
for g in golds:
gk = _row_key(g, gold_key)
if gk not in pred_by_key:
# Missing prediction = empty predicted set (penalized, not skipped).
pred_sets.append([])
gold_sets.append(list(g["acupoints"]))
scores.append(None)
continue
matched += 1
pr = pred_by_key[gk]
sc = pr.get("scores")
if pr.get("acupoints") is not None:
pset = list(pr["acupoints"])
elif sc:
pset = [p for p, v in sc.items() if v > score_threshold]
else:
pset = []
pred_sets.append(pset)
gold_sets.append(list(g["acupoints"]))
scores.append(sc if sc else None)
print(f"Matched {matched}/{len(golds)} gold rows to predictions "
f"({len(golds) - matched} gold rows had no prediction -> scored as empty set).")
return pred_sets, gold_sets, scores
def main() -> None:
ap = argparse.ArgumentParser(description="Evaluate AcuBench predictions against gold labels.")
ap.add_argument("--pred", required=True, help="Predictions JSONL file.")
ap.add_argument("--gold", required=True, help="Gold JSONL file (e.g. acubench.jsonl).")
ap.add_argument("--who", default=None, help="who_acupoints.csv (enables invalid_combination_rate).")
ap.add_argument("--split", default=None, help="Only score gold rows whose 'split' == this value.")
ap.add_argument("--gold-key", default="id", help="Join column (default 'id'; use 'symptoms' for sample_labels.jsonl).")
ap.add_argument("--score-threshold", type=float, default=0.5, help="Threshold to derive a set from scores when 'acupoints' absent.")
ap.add_argument("--ks", default="5,10,20", help="Comma-separated k values for @k metrics.")
ap.add_argument("--out", default=None, help="Optional path to also write the metric dict as JSON.")
args = ap.parse_args()
preds = load_jsonl(args.pred)
golds = load_jsonl(args.gold)
if args.split is not None:
golds = [g for g in golds if g.get("split") == args.split]
print(f"Filtered gold to split={args.split!r}: {len(golds)} rows.")
ks = tuple(int(x) for x in args.ks.split(",") if x.strip())
pmm = load_point_meridian_map(args.who) if args.who else None
pred_sets, gold_sets, scores = align(preds, golds, args.gold_key, args.score_threshold)
result = summarize(pred_sets, gold_sets, scores=scores, point_meridian_map=pmm, ks=ks)
print(json.dumps(result, indent=2))
if args.out:
with open(args.out, "w") as f:
json.dump(result, f, indent=2)
print(f"Wrote {args.out}")
if __name__ == "__main__":
main()
|