rcb/env.py

"""Environments: synthetic stochastic-covariate linear bandit (Appendix F.1
Settings 1-4) and the PharmGKB / IWPC warfarin dosing environment (Section 5.1)."""

import numpy as np
import pandas as pd


class SyntheticEnv:
    """Stochastic covariates x_t ~ P_X (uniform on the unit sphere, so ||x||=1
    as Corollary 1 assumes and Sigma = I/d, phi0 = 1/d), arm parameters drawn
    from the public Gaussian prior beta_i ~ N(beta_{i,0}, Sigma_{i,0}).

    Mean rewards mu(x,i) = offset + x'beta_i are kept inside [0,1] (the paper's
    model assumes mu(x_t) in [0,1]^K) via offset = 0.5.
    """

    def __init__(self, K, d, sigma, beta0, Sigma0, offset=0.5, rng=None):
        self.K, self.d, self.sigma, self.offset = K, d, float(sigma), float(offset)
        self.rng = rng if rng is not None else np.random.default_rng(0)
        self.beta = np.array([
            self.rng.multivariate_normal(np.asarray(beta0[i], float),
                                         np.asarray(Sigma0[i], float))
            for i in range(K)])
        self.phi0 = 1.0 / d          # lambda_min(E[xx']) for the unit sphere

    def context(self):
        z = self.rng.standard_normal(self.d)
        return z / np.linalg.norm(z)

    def mean_rewards(self, x):
        return self.offset + self.beta @ x

    def pull(self, x, arm):
        return float(self.mean_rewards(x)[arm] + self.sigma * self.rng.standard_normal())


# ------------------------------------------------------------------ warfarin

DOSE_COL = "Therapeutic Dose of Warfarin"


def build_warfarin(csv_path):
    """Feature construction per Appendix F.4: demographics, diagnosis, pre-existing
    diagnoses, three medications, CYP2C9 + VKORC1 genotypes; categoricals -> dummies;
    all missing values -> 0. Target: dose bucket Low <3, Medium 3-7, High >7 mg/day."""
    df = pd.read_csv(csv_path, low_memory=False)
    df = df[df[DOSE_COL].notna()].reset_index(drop=True)

    feats = {}

    # -- Demographics ------------------------------------------------------
    g = df.get("Gender")
    feats["male"] = (g == "male").astype(float) if g is not None else 0.0
    feats["female"] = (g == "female").astype(float) if g is not None else 0.0
    for r in ["White", "Asian", "Black or African American", "Unknown"]:
        feats[f"race_{r}"] = (df["Race (OMB)"] == r).astype(float)
    for e in df["Ethnicity (OMB)"].dropna().unique():
        feats[f"eth_{e}"] = (df["Ethnicity (OMB)"] == e).astype(float)
    age = df["Age"].astype(str).str.extract(r"(\d+)")[0].astype(float) / 10.0
    feats["age"] = age.fillna(0.0)
    for c in ["Height (cm)", "Weight (kg)"]:
        v = pd.to_numeric(df[c], errors="coerce").fillna(0.0)
        feats[c] = v / (v.max() if v.max() > 0 else 1.0)

    # -- Diagnosis (reason for treatment): 1 if a reason is recorded --------
    ind = df["Indication for Warfarin Treatment"].astype(str)
    for k, code in enumerate(["1", "2", "3", "4", "5", "6", "7", "8"]):
        feats[f"indication_{code}"] = ind.str.contains(code, regex=False).astype(float)
    feats["indication_known"] = (df["Indication for Warfarin Treatment"].notna()).astype(float)

    # -- Pre-existing diagnoses -------------------------------------------
    for c in ["Diabetes", "Congestive Heart Failure and/or Cardiomyopathy",
              "Valve Replacement", "Current Smoker"]:
        feats[c] = pd.to_numeric(df[c], errors="coerce").fillna(0.0).clip(0, 1)

    # -- Medications: aspirin, Tylenol, Zocor only (rest set to 0) ---------
    for c in ["Aspirin", "Acetaminophen or Paracetamol (Tylenol)", "Simvastatin (Zocor)"]:
        feats[c] = pd.to_numeric(df[c], errors="coerce").fillna(0.0).clip(0, 1)

    # -- Genetics: CYP2C9 + VKORC1 consensus dummies -----------------------
    cyp = df["CYP2C9 consensus"].fillna("missing")
    for lv in ["*1/*1", "*1/*2", "*1/*3", "*2/*2", "*2/*3", "*3/*3", "*1/*5",
               "*1/*6", "*1/*11", "*1/*13", "*1/*14", "missing"]:
        feats[f"cyp_{lv}"] = (cyp == lv).astype(float)
    for snp in ["-1639", "497", "1173", "1542", "3730", "2255", "-4451"]:
        col = f"VKORC1 {snp} consensus"
        if col in df.columns:
            v = df[col].fillna("missing")
            for lv in sorted(v.unique()):
                feats[f"vk{snp}_{lv}"] = (v == lv).astype(float)
    feats["gender_missing"] = df["Gender"].isna().astype(float)

    X = pd.DataFrame(feats).astype(float).fillna(0.0)
    X = X.loc[:, X.std(axis=0) > 0]                 # drop all-constant columns
    X.insert(0, "intercept", 1.0)

    dose = df[DOSE_COL].astype(float) / 7.0         # mg/week -> mg/day
    arm = np.where(dose < 3.0, 0, np.where(dose <= 7.0, 1, 2))
    return X.values, arm, dose.values, list(X.columns)