"""
TIER 2 — sensitivity analyses on the FULL UNMATCHED sample (manuscript section 3.D)
and bootstrap confidence intervals.

Manuscript reports for the unmatched sample (23 cases, 290 controls), SMOTE inside
training folds only:
    Random Forest : accuracy 0.948, recall 0.422, F1 0.515
    Logistic Regr.: recall 0.706
    + socioeconomic features -> accuracy 0.983, F1 0.870
"""
import json
import numpy as np
import pandas as pd
from sklearn.base import clone
from sklearn.ensemble import RandomForestClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import RepeatedStratifiedKFold
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import accuracy_score, recall_score, f1_score, precision_score
from imblearn.pipeline import Pipeline as ImbPipeline
from imblearn.over_sampling import SMOTE

from common import build_arms, FEATURES, RISK_FEATURES, SES_FEATURES

RNG = 42
N_SPLITS, N_REPEATS = 5, 5
OUT = "/home/hermes/journal-revision/analysis"

df, prov = build_arms()

# unmatched sample: cases aged <= 72 months + all community controls (per manuscript)
cases = df[(df["case"] == 1) & (df["age_months"] <= 72)]
controls = df[df["case"] == 0]
unm = pd.concat([cases, controls], ignore_index=True)
print(f"unmatched: {len(cases)} cases, {len(controls)} controls  (manuscript: 23 cases, 290 controls)")


def evaluate(est, frame, cols, label, smote=True):
    X = frame[cols].fillna(frame[cols].median(numeric_only=True))
    y = frame["case"].astype(int)
    cv = RepeatedStratifiedKFold(n_splits=N_SPLITS, n_repeats=N_REPEATS, random_state=RNG)
    rows = []
    for tr, te in cv.split(X, y):
        m = clone(est)
        m.fit(X.iloc[tr], y.iloc[tr])
        pred = m.predict(X.iloc[te])
        rows.append({
            "accuracy": accuracy_score(y.iloc[te], pred),
            "precision": precision_score(y.iloc[te], pred, zero_division=0),
            "recall": recall_score(y.iloc[te], pred, zero_division=0),
            "f1": f1_score(y.iloc[te], pred, zero_division=0),
        })
    a = {k: np.array([r[k] for r in rows]) for k in rows[0]}
    return {"analysis": label, "n_features": len(cols),
            **{f"{k}_mean_sd": f"{v.mean():.3f} +- {v.std(ddof=1):.3f}" for k, v in a.items()},
            **{f"{k}_mean": round(float(v.mean()), 4) for k, v in a.items()}}


def mk_rf():
    return ImbPipeline([("smote", SMOTE(random_state=RNG, k_neighbors=4)), ("clf", RandomForestClassifier(
        n_estimators=500, min_samples_leaf=2, max_features="sqrt", class_weight="balanced",
        random_state=RNG, n_jobs=1))])


def mk_lr():
    return ImbPipeline([("smote", SMOTE(random_state=RNG, k_neighbors=4)),
                        ("clf", Pipeline([("sc", StandardScaler()),
                                          ("lr", LogisticRegression(max_iter=5000, class_weight="balanced"))]))])


results = [
    evaluate(mk_rf(), unm, RISK_FEATURES, "RF, risk factors only (SES excluded)"),
    evaluate(mk_lr(), unm, RISK_FEATURES, "LR, risk factors only (SES excluded)"),
    evaluate(mk_rf(), unm, FEATURES, "RF, risk factors + socioeconomic"),
    evaluate(mk_lr(), unm, FEATURES, "LR, risk factors + socioeconomic"),
]
for r in results:
    print(json.dumps(r, indent=2, ensure_ascii=False))

# ---------------------------------------------------------------- bootstrap CIs (matched set)
matched = pd.read_csv(f"{OUT}/matched_set.csv")
Xm = matched[FEATURES].fillna(matched[FEATURES].median(numeric_only=True))
ym = matched["case"].astype(int)
probs = np.load(f"{OUT}/oof_prob_rf.npy")
rng = np.random.default_rng(RNG)
boot = {"accuracy": [], "recall": [], "f1": [], "auc": []}
from sklearn.metrics import roc_auc_score
for _ in range(2000):
    idx = rng.integers(0, len(ym), len(ym))
    yb, pb = ym.values[idx], probs[idx]
    if len(np.unique(yb)) < 2:
        continue
    pred = (pb >= 0.5).astype(int)
    boot["accuracy"].append(accuracy_score(yb, pred))
    boot["recall"].append(recall_score(yb, pred, zero_division=0))
    boot["f1"].append(f1_score(yb, pred, zero_division=0))
    boot["auc"].append(roc_auc_score(yb, pb))

boot_ci = {k: {"mean": round(float(np.mean(v)), 4),
               "ci95": [round(float(np.percentile(v, 2.5)), 4), round(float(np.percentile(v, 97.5)), 4)]}
           for k, v in boot.items()}

payload = {
    "provenance": prov,
    "unmatched_n": {"cases": int(len(cases)), "controls": int(len(controls))},
    "unmatched_results": results,
    "manuscript_unmatched": {"RF": {"accuracy": 0.948, "recall": 0.422, "f1": 0.515},
                             "LR_recall": 0.706,
                             "RF_plus_SES": {"accuracy": 0.983, "f1": 0.870}},
    "bootstrap_ci_matched_rf_2000": boot_ci,
}
with open(f"{OUT}/tier2_unmatched.json", "w") as fh:
    json.dump(payload, fh, indent=2, ensure_ascii=False)
print("\n[bootstrap CIs]", json.dumps(boot_ci, indent=2))
print("[saved] tier2_unmatched.json")
