model/train_and_save.py

# Import required libraries
from pathlib import Path
import pandas as pd
import numpy as np
import joblib

# Scikit-learn imports for building ML pipeline
from sklearn.compose import ColumnTransformer       
from sklearn.preprocessing import OneHotEncoder, StandardScaler  
from sklearn.impute import SimpleImputer              
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import Pipeline
from sklearn.model_selection import train_test_split  
from sklearn.metrics import classification_report, roc_auc_score

DATA_PATH = Path("C:\Users\wissa\Downloads\data\stroke-flask-docker\data\healthcare-dataset-stroke-data.csv")
OUT_PATH = Path("C:\Users\wissa\Downloads\data\stroke-flask-docker\model\stroke_pipeline.joblib")
OUT_PATH.parent.mkdir(parents=True, exist_ok=True)  # Make sure output folder exists

# Define feature groups
CATEGORICAL = ["gender","ever_married","work_type","Residence_type","smoking_status"]
NUMERIC = ["age","avg_glucose_level","bmi"]
BINARY_INT = ["hypertension","heart_disease"]  # Already numeric (0/1), but treated separately


def load_real_or_synthetic():
    if DATA_PATH.exists():
        # Load dataset from CSV
        df = pd.read_csv(DATA_PATH)

        # Define which columns we MUST have
        must_have = ["gender","age","hypertension","heart_disease","ever_married",
                     "work_type","Residence_type","avg_glucose_level","bmi",
                     "smoking_status","stroke"]

        # Check if any required columns are missing
        missing = set(must_have) - set(df.columns)
        if missing:
            raise ValueError(f"Dataset is missing columns: {missing}")

        # Drop extra columns like "id" if present, keep only required ones
        df = df[[c for c in df.columns if c in must_have]]
        return df
    else:
        # If dataset file is not found, generate synthetic (random but realistic) data
        rng = np.random.RandomState(42)  # Random seed for reproducibility
        N = 2000  # number of synthetic rows

        # Generate random values for each feature
        df = pd.DataFrame({
            "gender": rng.choice(["Male","Female","Other"], size=N, p=[0.49,0.50,0.01]),
            "age": rng.randint(1, 90, size=N),
            "hypertension": rng.binomial(1, 0.15, size=N),   # 15% chance of hypertension
            "heart_disease": rng.binomial(1, 0.08, size=N), # 8% chance of heart disease
            "ever_married": rng.choice(["Yes","No"], size=N, p=[0.7,0.3]),
            "work_type": rng.choice(["Private","Self-employed","Govt_job","children","Never_worked"], 
                                    size=N, p=[0.6,0.2,0.18,0.01,0.01]),
            "Residence_type": rng.choice(["Urban","Rural"], size=N, p=[0.55,0.45]),
            "avg_glucose_level": rng.normal(100, 30, size=N).clip(50, 300),  # realistic range
            "bmi": rng.normal(28, 6, size=N).clip(10, 60),
            "smoking_status": rng.choice(["formerly smoked","never smoked","smokes","Unknown"], 
                                         size=N, p=[0.2,0.6,0.15,0.05]),
        })

        # Define a "logit" (linear combination of features) that influences stroke probability
        logit = (
            0.03*df["age"] +
            0.02*(df["avg_glucose_level"]-100) +
            0.05*(df["bmi"]-28) +
            0.8*df["hypertension"] +
            0.9*df["heart_disease"] +
            0.3*(df["ever_married"]=="Yes").astype(int)
        )

        # Convert logit to probability using sigmoid function
        prob = 1/(1+np.exp(- (logit-4.0)))  # shift so stroke is rare (imbalanced dataset)

        # Assign stroke label (1 = stroke, 0 = no stroke) based on probability
        df["stroke"] = (rng.rand(len(df)) < prob).astype(int)
        return df


def build_pipeline():
    # For categorical features: fill missing with most frequent, then one-hot encode
    cat_proc = Pipeline(steps=[
        ("impute", SimpleImputer(strategy="most_frequent")),
        ("ohe", OneHotEncoder(handle_unknown="ignore"))
    ])

    # For numeric features: fill missing with median, then scale to mean=0, std=1
    num_proc = Pipeline(steps=[
        ("impute", SimpleImputer(strategy="median")),
        ("scale", StandardScaler())
    ])

    # For binary integer features: impute, then scale (optional but safe for pipeline)
    bin_proc = Pipeline(steps=[
        ("impute", SimpleImputer(strategy="most_frequent")),
        ("scale", StandardScaler(with_mean=False))  # keep sparse-friendly format
    ])

    # Combine all processors into one column transformer
    pre = ColumnTransformer(transformers=[
        ("cat", cat_proc, CATEGORICAL),
        ("num", num_proc, NUMERIC),
        ("bin", bin_proc, BINARY_INT),
    ])

    # Define classifier (logistic regression for binary classification)
    clf = LogisticRegression(max_iter=1000, n_jobs=None)

    # Final pipeline: preprocessing → model
    pipeline = Pipeline([("pre", pre), ("clf", clf)])
    return pipeline


def main():
    df = load_real_or_synthetic()

    # Split into features (X) and target (y = stroke)
    X = df.drop(columns=["stroke"])
    y = df["stroke"].astype(int)

    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42, stratify=y
    )

    pipeline = build_pipeline()
    pipeline.fit(X_train, y_train)

    y_prob = pipeline.predict_proba(X_test)[:,1]   # probability of stroke
    y_pred = (y_prob >= 0.3).astype(int)           # classify as 1 if prob ≥ 0.3

    print("AUC:", roc_auc_score(y_test, y_prob))  # area under ROC curve
    print("Report:\n", classification_report(y_test, y_pred))  # precision/recall/F1

    joblib.dump(pipeline, OUT_PATH)
    print(f"Saved pipeline to {OUT_PATH.resolve()}")


if __name__ == "__main__":
    main()