代码和第一个文件发上来了import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import KFold
import warnings
warnings.filterwarnings("ignore")
# 文档中iCOVID关键参数（用于特征处理对齐）
CLINICAL_NUM_FEATURES = ["age", "albumin", "hemoglobin", "total_protein", "ldh"] # 数值型临床特征（参考🔶1 - 41、🔶1 - 94）
CLINICAL_CAT_FEATURES = ["gender", "fever", "cough", "diabetes", "ards", "shock"] # 分类型临床特征（参考🔶1 - 41、🔶1 - 93）
OUTCOME_DEF = {0: "censored", 1: "recovered", 2: "deceased"} # 结局类型定义（对应🔶1 - 51的σₙ）
MAX_RECOVER_DAY = 31 # 恢复时间>30天设为31天，死亡设为32天（参考🔶1 - 60）
# 更新文件路径
file_paths = {
"metainfo": "/mnt/patients_metainfo (TARGET GDC,2025).csv",
"mutation": "/mnt/patients_mutation (TARGET GDC,2025).csv",
"outcomes": "/mnt/patients_outcomes (TARGET GDC,2025).csv",
"mrna": "/mnt/mrna_zcores.xlsx",
"cna": "/mnt/cna.xlsx"
}
# 数据加载与一致性校验（文档核心要求）
# 加载已有的文件
df_metainfo = pd.read_csv(file_paths["metainfo"])
df_mutation = pd.read_csv(file_paths["mutation"])
df_outcomes = pd.read_csv(file_paths["outcomes"])
df_mrna = pd.ExcelFile(file_paths["mrna"]).parse() # 读取 Excel 文件
df_cna = pd.ExcelFile(file_paths["cna"]).parse() # 读取 Excel 文件
# 查看各文件基础信息
print("各文件患者数量与列名：")
for name, df in [("metainfo", df_metainfo), ("mutation", df_mutation), ("outcomes", df_outcomes), ("mrna", df_mrna), ("cna", df_cna)]:
print(f"- {name}文件: 患者数={len(df)}, 列数={len(df.columns)}, 列名={list(df.columns)[:5]}...")
def unify_patient_id(df_list, base_df, id_col="patient_id"):
"""
统一患者ID：以base_df（结局文件）的ID为基准，删除其他文件中不匹配的ID
df_list: 待校验的DataFrame列表，base_df: 基准DataFrame（结局文件）
"""
# 基准ID集合
base_ids = set(base_df[id_col].unique())
print(f"基准结局文件患者数：{len(base_ids)}")
# 校验并过滤每个文件
unified_dfs = []
for df in df_list:
# 检查ID列是否存在
if id_col not in df.columns:
raise ValueError(f"文件缺少{id_col}列，请检查格式！")
# 筛选匹配ID
df_filtered = df[df[id_col].isin(base_ids)].copy()
# 统计删除数量
deleted = len(df) - len(df_filtered)
print(f"筛选后{df.columns[0].split('_')[0]}文件：保留{len(df_filtered)}例，删除{deleted}例（ID不匹配）")
unified_dfs.append(df_filtered)
return unified_dfs + [base_df] # 返回所有筛选后的数据（含基准结局文件）
# 执行ID统一（待校验文件：metainfo、mutation、mrna、cna）
df_metainfo, df_mutation, df_mrna, df_cna, df_outcomes = unify_patient_id(
df_list=[df_metainfo, df_mutation, df_mrna, df_cna],
base_df=df_outcomes
)
# 再次校验所有文件ID数量是否一致
all_ids = [set(df["patient_id"].unique()) for df in [df_metainfo, df_mutation, df_mrna, df_cna, df_outcomes]]
assert len(set.intersection(*all_ids)) == len(all_ids[0]), "ID筛选后仍不匹配，请检查数据！"
print(f"\n最终统一患者数：{len(set.intersection(*all_ids))}")
def stat_missing_values(df_list, df_names):
"""统计每个文件的缺失值比例"""
missing_stats = []
for df, name in zip(df_list, df_names):
# 计算每列缺失率
missing_rate = (df.isnull().sum() / len(df) * 100).round(2)
# 筛选缺失率>0的列
missing_cols = missing_rate[missing_rate > 0]
if len(missing_cols) > 0:
missing_stats.append({
"file": name,
"missing_cols": dict(missing_cols),
"max_missing_rate": missing_cols.max(),
"min_missing_rate": missing_cols.min()
})
else:
missing_stats.append({
"file": name,
"missing_cols": "无",
"max_missing_rate": 0,
"min_missing_rate": 0
})
# 转为DataFrame展示
df_missing = pd.DataFrame(missing_stats)
return df_missing
# 统计缺失值
df_missing = stat_missing_values(
df_list=[df_metainfo, df_mutation, df_mrna, df_cna, df_outcomes],
df_names=["metainfo", "mutation", "mrna", "cna", "outcomes"]
)
print("各文件缺失值统计：")
print(df_missing.to_string(index=False))
# 分文件特征预处理（严格对齐文档逻辑）
def preprocess_metainfo(df_metainfo, id_col="patient_id"):
"""
处理临床基础信息：分类特征二值化，数值特征归一化
分类特征（如gender、fever）：0=无/女，1=有/男；数值特征（age）：归一化到[0,1]
"""
df = df_metainfo.copy()
# 1. 分类特征二值化（参考文档中症状/合并症的编码逻辑🔶1 - 41）
cat_features = ["gender", "fever", "cough", "expectoration", "diabetes", "ards", "shock"]
for feat in cat_features:
if feat in df.columns:
# 缺失值填0（无），非缺失值映射为0/1
df[feat] = df[feat].fillna(0)
# 若原始值为字符串（如"男"/"女"），转换为0/1
if df[feat].dtype == "object":
df[feat] = df[feat].map({"女": 0, "男": 1, "无": 0, "有": 1}).fillna(0)
df[feat] = df[feat].astype(int)
# 2. 数值特征归一化（年龄，参考文档🔶1 - 136的Min - Max归一化）
num_features = ["age"]
scaler = MinMaxScaler(feature_range=(0, 1))
for feat in num_features:
if feat in df.columns:
# 缺失值填均值
df[feat] = df[feat].fillna(df[feat].mean())
# 归一化
df[feat] = scaler.fit_transform(df[[feat]]).flatten()
# 3. 保留ID和处理后的特征
keep_cols = [id_col] + cat_features + num_features
keep_cols = [col for col in keep_cols if col in df.columns]
df_processed = df[keep_cols]
return df_processed, scaler
# 执行处理
df_metainfo_processed, age_scaler = preprocess_metainfo(df_metainfo)
print(f"\nmetainfo处理后：列数={len(df_metainfo_processed.columns)}，示例：")
print(df_metainfo_processed.head())
def preprocess_mutation(df_mutation, id_col="patient_id"):
"""
处理基因突变特征：二值化（0=无突变，1=有突变），缺失值填0
"""
df = df_mutation.copy()
# 所有基因列（排除ID列）
gene_cols = [col for col in df.columns if col != id_col]
# 1. 缺失值填0（默认无突变）
df[gene_cols] = df[gene_cols].fillna(0)
# 2. 二值化（确保所有值为0/1，如原始为"突变"/"野生型"则映射）
for col in gene_cols:
if df[col].dtype == "object":
df[col] = df[col].map({"野生型": 0, "突变": 1}).fillna(0)
# 若为数值（如突变频率），>0视为1（有突变）
else:
df[col] = (df[col] > 0).astype(int)
df_processed = df[[id_col] + gene_cols]
return df_processed
# 执行处理
df_mutation_processed = preprocess_mutation(df_mutation)
print(f"\nmutation处理后：基因特征数={len(df_mutation_processed.columns)-1}，示例：")
print(df_mutation_processed.head())
def preprocess_mrna(df_mrna, id_col="patient_id"):
"""
处理mRNA表达特征：z值已标准化，仅填充缺失值（填0，即与均值一致），转换为CSV格式
"""
df = df_mrna.copy()
# 1. 确保ID列存在且命名统一
if "sample_id" in df.columns and id_col not in df.columns:
df.rename(columns={"sample_id": id_col}, inplace=True)
# 2. 基因列（排除ID列）
gene_cols = [col for col in df.columns if col != id_col]
# 3. 缺失值填0（z值为0表示与均值一致，符合文档缺失值处理逻辑🔶1 - 136）
df[gene_cols] = df[gene_cols].fillna(0)
df_processed = df[[id_col] + gene_cols]
return df_processed
# 执行处理
df_mrna_processed = preprocess_mrna(df_mrna)
print(f"\nmrna处理后：基因特征数={len(df_mrna_processed.columns)-1}，示例：")
print(df_mrna_processed.head())
def preprocess_cna(df_cna, id_col="patient_id"):
"""
处理拷贝数变异特征：编码为-1（缺失）、0（正常）、1（扩增），缺失值填0
"""
df = df_cna.copy()
gene_cols = [col for col in df.columns if col != id_col]
# 1. 缺失值填0（默认正常）
df[gene_cols] = df[gene_cols].fillna(0)
# 2. 编码（原始值如"缺失"/"正常"/"扩增"映射为-1/0/1）
for col in gene_cols:
if df[col].dtype == "object":
df[col] = df[col].map({"缺失": -1, "正常": 0, "扩增": 1}).fillna(0)
# 若为数值（如拷贝数），<0=-1，=0=0，>0=1
else:
df[col] = np.where(df[col] < 0, -1, np.where(df[col] > 0, 1, 0))
df_processed = df[[id_col] + gene_cols]
return df_processed
# 执行处理
df_cna_processed = preprocess_cna(df_cna)
print(f"\ncna处理后：基因特征数={len(df_cna_processed.columns)-1}，示例：")
print(df_cna_processed.head())
def preprocess_outcomes(df_outcomes, id_col="patient_id", max_recover_day=MAX_RECOVER_DAY):
"""
处理结局变量：定义σₙ（0=截尾，1=恢复，2=死亡）和tₙ（结局时间）
恢复时间>30天设为31，死亡设为32（参考文档🔶1 - 60）
"""
df = df_outcomes.copy()
# 1. 结局类型编码（σₙ）
if "outcome_type" in df.columns:
# 若原始为字符串，映射为0/1/2
if df["outcome_type"].dtype == "object":
df["outcome_type"] = df["outcome_type"].map({
"censored": 0, "lost_to_follow_up": 0, # 截尾（失访）
"recovered": 1, "discharged": 1, # 恢复（出院）
"deceased": 2, "death": 2 # 死亡
}).fillna(0) # 缺失值视为截尾
else:
# 若无outcome_type，按结局时间推断（参考文档🔶1 - 51）
df["outcome_type"] = np.where(
df["outcome_days"].isnull(), 0, # 无时间→截尾
np.where(df["outcome_days"] <= 30, 1, 2) # ≤30天→恢复，>30→死亡（暂设）
)
# 2. 结局时间处理（tₙ）
if "outcome_days" in df.columns:
# 缺失值填7（截尾默认时间，参考文档🔶1 - 30）
df["outcome_days"] = df["outcome_days"].fillna(7).astype(int)
# 恢复患者：>30天设为31
df.loc[(df["outcome_type"] == 1) & (df["outcome_days"] > 30), "outcome_days"] = max_recover_day
# 死亡患者：统一设为32
df.loc[df["outcome_type"] == 2, "outcome_days"] = 32
# 截尾患者：时间≤10天（参考文档🔶1 - 30）
df.loc[df["outcome_type"] == 0, "outcome_days"] = np.clip(df.loc[df["outcome_type"] == 0, "outcome_days"], 3, 10)
# 保留关键列
df_processed = df[[id_col, "outcome_type", "outcome_days"]]
return df_processed
# 执行处理
df_outcomes_processed = preprocess_outcomes(df_outcomes)
print(f"\noutcomes处理后：结局分布={df_outcomes_processed['outcome_type'].value_counts().to_dict()}，示例：")
print(df_outcomes_processed.head())
# 多表整合为患者级多模态矩阵
def integrate_multimodal_data(df_list, id_col="patient_id"):
"""
多表整合：以patient_id为键，左连接所有处理后的DataFrame
df_list: 处理后的DataFrame列表（含metainfo、mutation、cna、mrna、blood_cell、outcomes）
"""
# 以第一个DataFrame为基础，逐步连接其他DataFrame
integrated_df = df_list[0].copy()
for df in df_list[1:]:
integrated_df = integrated_df.merge(df, on=id_col, how="left")
# 检查连接后是否有新增缺失值（理论上不应有，因ID已统一）
new_missing = integrated_df.isnull().sum().sum() - integrated_df[id_col].isnull().sum()
if new_missing > 0:
print(f"连接{df.columns[1].split('_')[0]}后新增{new_missing}个缺失值，已填充为0")
integrated_df = integrated_df.fillna(0)
# 确保无重复行
integrated_df = integrated_df.drop_duplicates(subset=[id_col])
print(f"\n整合后多模态矩阵：患者数={len(integrated_df)}，特征数={len(integrated_df.columns)-3}（不含ID、结局类型、结局时间）")
return integrated_df
# 执行整合（顺序：metainfo→mutation→mrna→cna→outcomes）
df_integrated = integrate_multimodal_data([
df_metainfo_processed,
df_mutation_processed,
df_mrna_processed,
df_cna_processed,
df_outcomes_processed
])
# 查看整合结果
print("\n整合后数据示例（前5列+最后3列）：")
print(df_integrated.iloc[:, list(range(5)) + list(range(-3, 0))].head())
# 数据划分与输出（贴合文档验证策略）
def split_and_save_data(df_integrated, id_col="patient_id", save_path="./processed_data/"):
"""
数据划分：五折交叉验证划分训练/验证集，保存处理后数据（CSV格式）
参考文档🔶1 - 26的五折交叉验证逻辑
"""
import os
os.makedirs(save_path, exist_ok=True)
# 1. 保存完整整合数据
df_integrated.to_csv(os.path.join(save_path, "integrated_patients_data.csv"), index=False)
print(f"完整数据已保存至：{os.path.join(save_path, 'integrated_patients_data.csv')}")
# 2. 五折交叉验证划分
kf = KFold(n_splits=5, shuffle=True, random_state=42)
patients = df_integrated[id_col].unique()
fold = 1
for train_idx, val_idx in kf.split(patients):
# 划分训练/验证ID
train_ids = patients[train_idx]
val_ids = patients[val_idx]
# 生成训练/验证集
df_train = df_integrated[df_integrated[id_col].isin(train_ids)]
df_val = df_integrated[df_integrated[id_col].isin(val_ids)]
# 保存
df_train.to_csv(os.path.join(save_path, f"train_fold{fold}.csv"), index=False)
df_val.to_csv(os.path.join(save_path, f"val_fold{fold}.csv"), index=False)
print(f"Fold{fold}：训练集{len(df_train)}例，验证集{len(df_val)}例，已保存")
fold += 1
# 3. 保存特征说明（用于后续模型解释，参考文档FSR机制🔶1 - 61）
feature_info = pd.DataFrame({
"feature_name": [col for col in df_integrated.columns if col not in [id_col, "outcome_type", "outcome_days"]],
"feature_type": ["clinical" if col in df_metainfo_processed.columns else
"mutation" if col in df_mutation_processed.columns else
"mrna" if col in df_mrna_processed.columns else
"cna" if col in df_cna_processed.columns else
"outcome" if col in df_outcomes_processed.columns else ""],
"processing_method": ["binary" if col in df_mutation_processed.columns or col in CLINICAL_CAT_FEATURES else
"normalized" if col in df_metainfo_processed.columns else
"z_score" if col in df_mrna_processed.columns else
"ternary" if col in df_cna_processed.columns else
"outcome_encoding" if col in df_outcomes_processed.columns else ""]
})
feature_info.to_csv(os.path.join(save_path, "feature_metadata.csv"), index=False)
print(f"特征说明已保存至：{os.path.join(save_path, 'feature_metadata.csv')}")
return df_train, df_val
# 执行划分与保存
df_train, df_val = split_and_save_data(df_integrated)
# 关键结果可视化（验证处理效果）
# 1. 结局分布可视化（参考文档🔶1 - 30）
plt.figure(figsize=(12, 4))
# 结局类型分布
plt.subplot(1, 2, 1)
outcome_counts = df_integrated["outcome_type"].value_counts().sort_index()
outcome_labels = [OUTCOME_DEF[i] for i in outcome_counts.index]
plt.bar(outcome_labels, outcome_counts.values, color=["lightblue", "lightgreen", "salmon"])
plt.title("患者结局类型分布（参考文档🔶1 - 30）")
plt.ylabel("患者数")
# 恢复时间分布（仅恢复患者）
plt.subplot(1, 2, 2)
recover_days = df_integrated[df_integrated["outcome_type"] == 1]["outcome_days"]
plt.hist(recover_days, bins=10, color="lightgreen", edgecolor="black")
plt.title("恢复患者时间分布（≤31天，参考文档🔶1 - 60）")
plt.xlabel("恢复时间（天）")
plt.ylabel("患者数")
plt.tight_layout()
plt.savefig("./processed_data/outcome_distribution.png", dpi=300, bbox_inches="tight")
plt.close()
print("\n结局分布图表已保存至：./processed_data/outcome_distribution.png")
# 2. 关键生物标志物相关性（参考文档🔶1 - 94的Pearson分析）
# 由于当前没有生物标志物数据，此部分代码先注释掉
# markers = ["albumin", "hemoglobin", "total_protein", "ldh"]
# markers = [m for m in markers if m in df_integrated.columns]
# recover_days = df_integrated[df_integrated["outcome_type"] == 1]["outcome_days"]
# corr_results = []
# for marker in markers:
# marker_vals = df_integrated[df_integrated["outcome_type"] == 1][marker]
# corr, p_val = pearsonr(marker_vals, recover_days)
# corr_results.append({"marker": marker, "pearson_corr": corr.round(3), "p_value": p_val.round(4)})
# df_corr = pd.DataFrame(corr_results)
# print("\n关键生物标志物与恢复时间的Pearson相关性（参考文档🔶1 - 94）：")
# print(df_corr.to_string(index=False))
最新发布
扫一扫
2231
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
