import pandas as pd import numpy as np from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler from sklearn.linear_model import LassoCV from scipy.stats import mannwhitneyu from imblearn.over_sampling import SMOTENC from collections import defaultdict from joblib import Parallel, delayed import warnings warnings.filterwarnings('ignore') np.random.seed(42) # ======================== 数据加载与划分 ======================== def load_and_split_data(): rad_features = pd.read_csv('D:/taskPONV/stable/rad_features2009320_96.csv') labels = pd.read_csv('D:/taskPONV/stable/label320_96.csv') data = pd.merge(rad_features, labels, on='ID', how='inner') categorical_features = ['gender', 'hypertension', 'diabetes', 'coronaryheartdisease', 'PONVhistory', 'smoke', 'antiemetic', 'dezosine', 'ASA', 'volatileanesthetic', 'LBP', 'postopioiduse'] mask = data['ID'].str.match(r'sub\d{3}\.nii\.gz') data_a = data[mask].reset_index(drop=True) data_b = data[~mask].reset_index(drop=True) X_train, X_test, y_train, y_test = train_test_split( data_a.drop(['ID', 'label'], axis=1), data_a['label'], test_size=0.3, stratify=data_a['label'], random_state=42 ) return X_train, X_test, y_train, y_test, data_b, categorical_features # # ======================== 预处理管道 ======================== class PreprocessingPipeline: def __init__(self, categorical_features): self.categorical_features = categorical_features self.scaler = None self.smote = None def fit_resample(self, X, y): cat_indices = [X.columns.get_loc(c) for c in self.categorical_features] self.smote = SMOTENC(categorical_features=cat_indices, random_state=42) X_res, y_res = self.smote.fit_resample(X, y) continuous_features = [c for c in X.columns if c not in self.categorical_features] self.scaler = StandardScaler().fit(X_res[continuous_features]) X_res_cont = self.scaler.transform(X_res[continuous_features]) return pd.DataFrame( np.hstack([X_res[self.categorical_features].values, X_res_cont]), columns=self.categorical_features + continuous_features ), y_res def transform(self, X): continuous_features = [c for c in X.columns if c not in self.categorical_features] X_cont = self.scaler.transform(X[continuous_features]) return pd.DataFrame( np.hstack([X[self.categorical_features].values, X_cont]), columns=self.categorical_features + continuous_features ) # ======================== 特征筛选流程 ======================== def bootstrap_feature_selection(X_train, y_train, preprocessor, n_iter=1000): feature_counts = defaultdict(int) results = Parallel(n_jobs=-1)( delayed(single_bootstrap_iteration)(X_train, y_train, preprocessor) for _ in range(n_iter) ) for features in results: for feat in features: feature_counts[feat] += 1 stability_df = pd.DataFrame.from_dict(feature_counts, orient='index', columns=['count']) stability_df['frequency'] = stability_df['count'] / n_iter return stability_df.sort_values('frequency', ascending=False) def single_bootstrap_iteration(X_train, y_train, preprocessor): # 1. Bootstrap采样 n_samples = len(X_train) bootstrap_indices = np.random.choice(n_samples, size=n_samples, replace=True) X_bs = X_train.iloc[bootstrap_indices] y_bs = y_train.iloc[bootstrap_indices] # 2. 预处理 X_processed, y_resampled = preprocessor.fit_resample(X_bs, y_bs) # 3. 单变量筛选 selected = univariate_selection(X_processed, y_resampled) # 4. 高相关过滤 selected = filter_high_correlation(X_processed[selected]) # 5. Lasso筛选 final_features = lasso_selection(X_processed[selected], y_resampled) return final_features def univariate_selection(X, y, alpha=0.05): """Mann-Whitney U检验筛选""" p_values = [] for col in X.columns: _, p = mannwhitneyu(X[col][y == 0], X[col][y == 1]) p_values.append(p) return X.columns[np.array(p_values) < alpha].tolist() def filter_high_correlation(X, threshold=0.9): """高相关特征过滤""" corr_matrix = X.corr().abs() upper = corr_matrix.where(np.triu(np.ones(corr_matrix.shape), k=1).astype(bool)) to_drop = [col for col in upper.columns if any(upper[col] > threshold)] return [col for col in X.columns if col not in to_drop] def lasso_selection(X, y): """Lasso特征筛选""" lasso = LassoCV(alphas=np.logspace(-3, 1, 50), cv=10, max_iter=5000) lasso.fit(X, y) return X.columns[np.abs(lasso.coef_) > 1e-6].tolist() # 使用lambda.min对应的系数 # ======================== 新增代码：特征数据提取 ======================== def export_top_features(): # 1. 读取特征稳定性结果 stability_df = pd.read_csv('D:/taskPONV/stable/feature_stability.csv', index_col=0) # 2. 获取前30个特征并加入gender top_features = stability_df.head(30).index.tolist() required_features = list(set(top_features + ['gender','PONVhistory'])) # 自动去重 # 3. 重新加载原始数据（确保获取所有样本） rad_features = pd.read_csv('D:/taskPONV/stable/rad_features2009320_96.csv') labels = pd.read_csv('D:/taskPONV/stable/label320_96.csv') full_data = pd.merge(rad_features, labels, on='ID', how='inner') # 合并全部数据 # 4. 验证特征存在性 valid_features = [f for f in required_features if f in full_data.columns] # 5. 创建包含样本名的新表格 result_df = full_data[['ID'] + valid_features] # 6. 保存结果 result_df.to_csv('D:/taskPONV/stable/stable_features_data.csv', index=False) print(f"生成表格包含 {len(valid_features)} 个特征，样本数：{len(result_df)}") # ======================== 主执行流程 ======================== if __name__ == "__main__": # 1. 数据加载与划分 X_train, X_test, y_train, y_test, data_b, categorical_features = load_and_split_data() # 2. 初始化预处理管道 preprocessor = PreprocessingPipeline(categorical_features) # 3. Bootstrap特征筛选 stability_df = bootstrap_feature_selection(X_train, y_train, preprocessor, n_iter=5000) # 4. 筛选稳定特征 stable_features = stability_df[stability_df['frequency'] > 0.5].index.tolist() print(f"最终稳定特征数: {len(stable_features)}") print("Top特征:", stable_features) # 5. 保存结果 stability_df.to_csv('D:/taskPONV/stable/feature_stability.csv') # 新增数据导出 export_top_features() stability_df = pd.read_csv('D:/taskPONV/stable/feature_stability.csv', index_col=0) # 2. 获取前20个特征并加入gender top_features = stability_df.head(30).index.tolist() print("top_features:",top_features)