如何用Python从5万+特征中选2个特征可视化区分Sarcopenia样本

我有一个CSV格式的数据库synthetic_feature_file，包含5万余个已处理的非原始特征，共43个样本。我希望从中找到两个特征，能将Sarcopenia样本与非Sarcopenia样本区分开，并用这两个特征作为坐标轴绘制散点图，使两类样本形成无重叠的聚类。目前通过随机选取特征反复运行代码查看结果，效率极低，请问该如何修改现有代码？

现有代码：

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# Read synthesized feature data
syn_data = pd.read_csv(synthetic_feature_file)

# labeled sample(people who suffer from Sarcopenia)
sample_indices = [1, 6, 7, 11, 14, 15, 27]

# Randomly pick two features as x and y axes
x_feature = np.random.choice(syn_data.columns[0:10000])
y_feature = np.random.choice(syn_data.columns[10001:20000])

# Clean feature names and remove illegal characters
x_feature = x_feature.strip().replace('\t', '')
y_feature = y_feature.strip().replace('\t', '')

plt.figure(figsize=(8, 6))

# Other samples(people who did not suffer from Sarcopenia)
other_samples = syn_data.drop(sample_indices)
plt.scatter(other_samples[x_feature], other_samples[y_feature], color='blue', label='Other Samples')

# Red sample
red_samples = syn_data.iloc[sample_indices]
plt.scatter(red_samples[x_feature], red_samples[y_feature], color='red', label='Sample Indices')

plt.xlabel(x_feature)
plt.ylabel(y_feature)
plt.title("Visualization")
plt.legend()
plt.show()

核心思路是先量化单个特征的分类区分能力，筛选出对两类样本区分度高的特征，再在这些高区分度特征中两两组合验证，自动找出能让两类样本无重叠的特征对，最后绘图，彻底替代随机尝试的低效方式。

步骤说明 #

1. 计算特征区分度：用Cohen's d指标（两类样本均值差除以合并标准差）衡量单个特征的区分能力，值越大说明特征越能分开两类样本
2. 筛选高区分度特征：设定阈值（比如d>1）缩小候选范围，避免遍历5万特征的海量组合
3. 验证特征对无重叠：遍历筛选后的特征对，检查两类样本在二维空间中是否完全无重叠，找到符合条件的组合后自动绘图

修改后的代码 #

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# 读取数据
syn_data = pd.read_csv("synthetic_feature_file")
# 标记Sarcopenia样本索引
sarc_indices = [1, 6, 7, 11, 14, 15, 27]
# 拆分两类样本
sarc_samples = syn_data.iloc[sarc_indices].reset_index(drop=True)
non_sarc_samples = syn_data.drop(sarc_indices).reset_index(drop=True)

# 计算单个特征的Cohen's d（区分度指标）
def calculate_cohens_d(feature):
    sarc_vals = sarc_samples[feature].values
    non_sarc_vals = non_sarc_samples[feature].values
    mean_diff = np.mean(sarc_vals) - np.mean(non_sarc_vals)
    pooled_std = np.sqrt(((len(sarc_vals)-1)*np.var(sarc_vals) + (len(non_sarc_vals)-1)*np.var(non_sarc_vals)) / (len(sarc_vals)+len(non_sarc_vals)-2))
    if pooled_std == 0:
        return 0  # 标准差为0说明该特征无区分度
    return abs(mean_diff / pooled_std)

# 计算所有特征的区分度，筛选高区分度特征（这里取d>1，可根据需求调整）
feature_scores = {col: calculate_cohens_d(col) for col in syn_data.columns}
high_d_features = [col for col, score in feature_scores.items() if score > 1]
print(f"筛选出高区分度特征数量：{len(high_d_features)}")

# 检查特征对是否能让两类样本无重叠
def check_no_overlap(x_col, y_col):
    # 获取两类样本的x、y值范围
    sarc_x_min, sarc_x_max = sarc_samples[x_col].min(), sarc_samples[x_col].max()
    sarc_y_min, sarc_y_max = sarc_samples[y_col].min(), sarc_samples[y_col].max()
    non_sarc_x_min, non_sarc_x_max = non_sarc_samples[x_col].min(), non_sarc_samples[x_col].max()
    non_sarc_y_min, non_sarc_y_max = non_sarc_samples[y_col].min(), non_sarc_samples[y_col].max()
    
    # 判断是否在x轴或y轴上完全分离，或二维矩形无交集
    if (sarc_x_max < non_sarc_x_min or sarc_x_min > non_sarc_x_max) or (sarc_y_max < non_sarc_y_min or sarc_y_min > non_sarc_y_max):
        return True
    return False

# 遍历寻找符合条件的特征对
valid_pairs = []
for i in range(len(high_d_features)):
    for j in range(i+1, len(high_d_features)):
        x_col = high_d_features[i]
        y_col = high_d_features[j]
        if check_no_overlap(x_col, y_col):
            valid_pairs.append((x_col, y_col))
            # 如果只需要一对符合条件的特征，找到后可直接break退出循环
            # break
    # if valid_pairs:
    #     break

print(f"找到符合条件的特征对数量：{len(valid_pairs)}")

# 绘制符合条件的特征对散点图（最多绘制前5个，可调整）
if valid_pairs:
    for idx, (x_feature, y_feature) in enumerate(valid_pairs[:5]):
        plt.figure(figsize=(8, 6))
        plt.scatter(non_sarc_samples[x_feature], non_sarc_samples[y_feature], color='blue', label='非Sarcopenia样本')
        plt.scatter(sarc_samples[x_feature], sarc_samples[y_feature], color='red', label='Sarcopenia样本')
        plt.xlabel(x_feature)
        plt.ylabel(y_feature)
        plt.title(f"无重叠聚类可视化 - 特征对{idx+1}")
        plt.legend()
        plt.show()
else:
    print("未找到能使两类样本无重叠的特征对，建议降低区分度阈值或调整无重叠判断条件")

关键优化点 #

• 效率提升：先筛选高区分度特征，把5万特征的组合量压缩到数千甚至数百次，避免无意义的随机尝试
• 可定制性：可以根据需求调整区分度阈值（比如把d>1改成d>0.8）或无重叠判断逻辑（比如允许少量重叠）
• 自动化：无需手动干预，代码会自动找到符合条件的特征对并绘图

内容的提问来源于stack exchange，提问作者董珈妤