fryabcfry's picture
Publish AMP-PD checkpoints, portable loaders and verified metrics (part 3)
04ae9bb verified
Raw History Blame Contribute Delete
3.55 kB
"""Patient-level and cohort-held-out splits."""
from __future__ import annotations
from pathlib import Path
from typing import Any, Dict, List
import numpy as np
import pandas as pd
from sklearn.model_selection import KFold, StratifiedKFold
def make_folds(patients: pd.DataFrame, cfg: Dict[str, Any]) -> pd.DataFrame:
scheme = cfg.get("scheme", "patient_kfold")
seed = int(cfg.get("seed", 42))
df = patients.copy()
train_patients = df[df["is_supplemental"] == 0].reset_index(drop=True)
supp_patients = df[df["is_supplemental"] == 1].reset_index(drop=True)
if scheme == "patient_kfold":
n_splits = int(cfg.get("n_splits", 5))
strat = cfg.get("stratify_by", "cohort")
y = train_patients[strat].astype(str).values if strat in train_patients.columns else None
counts = pd.Series(y).value_counts() if y is not None else None
if y is not None and counts is not None and counts.min() >= n_splits:
splitter = StratifiedKFold(n_splits=n_splits, shuffle=True, random_state=seed)
splits = list(splitter.split(train_patients, y))
else:
splitter = KFold(n_splits=n_splits, shuffle=True, random_state=seed)
splits = list(splitter.split(train_patients))
train_patients["fold"] = -1
for fold, (_, val_idx) in enumerate(splits):
train_patients.loc[val_idx, "fold"] = fold
elif scheme == "cohort_heldout":
cohorts = sorted(train_patients["cohort"].unique().tolist())
holdout = cfg.get("holdout_cohorts")
if holdout:
holdout = [c for c in cohorts if c in set(holdout)]
else:
holdout = cohorts
mapping = {c: i for i, c in enumerate(holdout)}
train_patients["fold"] = train_patients["cohort"].map(mapping).fillna(-1).astype(int)
elif scheme == "single":
frac = float(cfg.get("single_val_fraction", 0.2))
rng = np.random.default_rng(seed)
idx = np.arange(len(train_patients))
rng.shuffle(idx)
n_val = max(1, int(round(len(idx) * frac)))
train_patients["fold"] = 1
train_patients.loc[idx[:n_val], "fold"] = 0
else:
raise ValueError(f"Unknown cv scheme: {scheme}")
out = train_patients[["uid", "patient_id", "cohort", "fold", "is_supplemental"]].copy()
if len(supp_patients):
supp_patients = supp_patients.copy()
supp_patients["fold"] = -2
out = pd.concat(
[out, supp_patients[["uid", "patient_id", "cohort", "fold", "is_supplemental"]]],
ignore_index=True,
)
out = out.sort_values(["is_supplemental", "cohort", "uid"]).reset_index(drop=True)
return out
def fold_ids(folds: pd.DataFrame) -> List[int]:
return sorted(int(x) for x in folds.loc[folds["fold"] >= 0, "fold"].unique())
def save_folds(folds: pd.DataFrame, path: str | Path) -> None:
p = Path(path)
p.parent.mkdir(parents=True, exist_ok=True)
folds.to_csv(p, index=False)
def fold_summary(folds: pd.DataFrame) -> pd.DataFrame:
rows = []
for fold, g in folds[folds["fold"] >= 0].groupby("fold"):
rows.append(
{
"fold": int(fold),
"n_patients": int(len(g)),
"n_control": int((g["cohort"] == "control").sum()),
"n_frequent": int((g["cohort"] == "frequent").sum()),
"n_annual": int((g["cohort"] == "annual").sum()),
}
)
return pd.DataFrame(rows).sort_values("fold").reset_index(drop=True)