机器学习实战:基于3大分类模型的中风病人预测

公众号:尤而小屋
作者:Peter
编辑:Peter

大家好,我是Peter~

今天给大家分享的是kaggle上一个关于中风疾病案例的数据集建模,文章的主要内容参考导图:

原数据地址:www.kaggle.com/datasets/fe…

导入库

数据基本信息

先把数据导进来,查看数据的基本信息

下面我们查看数据基本信息

In [3]:

df.shape

Out[3]:

(5110, 12)

In [4]:

df.dtypes

Out[4]:

id                     int64
gender                object
age                  float64
hypertension           int64
heart_disease          int64
ever_married          object  # 字符型
work_type             object
Residence_type        object
avg_glucose_level    float64
bmi                  float64
smoking_status        object
stroke                 int64
dtype: object

In [5]:

df.describe()  # 描述统计信息

Out[5]:

字段分布

gender统计

In [6]:

plt.figure(1, figsize=(12,5))

sns.countplot(y="gender", data=df)
plt.show()

age分布

In [7]:

px.violin(y=df["age"])

fig = px.histogram(df,
                   x="age",
                   color_discrete_sequence=['firebrick'])

fig.show()

ever_married

In [9]:

plt.figure(1, figsize=(12,5))

sns.countplot(y="ever_married", data=df)

plt.show()

本数据集中的结婚人士大约是未结婚的两倍。

work-type

查看不同工作状态的人员数量

In [10]:

plt.figure(1, figsize=(12,8))

sns.countplot(y="work_type", data=df)

plt.show()

Residence_type

In [11]:

plt.figure(1, figsize=(12,8))

sns.countplot(y="Residence_type", data=df)

plt.show()

avg_glucose_level

血糖水平的分布

fig = px.histogram(df,
                   x="avg_glucose_level",
                   color_discrete_sequence=['firebrick'])

fig.show()

可以看到大部分人的血糖还是在100以下,说明是正常的

bmi

bmi指标的分布情况

bmi指标的均值大约在28左右,呈现一定的正态分布

smoking_status

抽烟情况的统计

plt.figure(1, figsize=(12,8))

sns.countplot(y="smoking_status", data=df)

plt.show()

可以看到抽烟或者曾经抽烟的人相对来说是少一些的

缺失值情况

缺失值统计

df.isnull().sum()
id                     0
gender                 0
age                    0
hypertension           0
heart_disease          0
ever_married           0
work_type              0
Residence_type         0
avg_glucose_level      0
bmi                  201
smoking_status         0
stroke                 0
dtype: int64
201 / len(df)  # 缺失比例
0.03933463796477495

缺失值可视化

缺失值处理

使用决策树回归来预测缺失值的BMI值:通过年龄、性别和现有的bmi值来进行预测填充

dt_bmi = Pipeline(steps=[("scale",StandardScaler()), # 数据标准化
                         ("lr",DecisionTreeRegressor(random_state=42))
                        ])

取出3个指标来进行预测填充:

X = df[["age","gender","bmi"]].copy()

dic = {"Male":0, "Female":1, "Other":-1}

X["gender"] = X["gender"].map(dic).astype(np.uint8)
X.head()

取出非缺失值的部分进行训练:

# 缺失值部分
missing = X[X.bmi.isna()]

# 非缺失值部分
X = X[~X.bmi.isna()]
y = X.pop("bmi")
# 模型训练

dt_bmi.fit(X,y)
Pipeline(steps=[('scale', StandardScaler()),
                ('lr', DecisionTreeRegressor(random_state=42))])

In [23]:

# 模型预测

y_pred = dt_bmi.predict(missing[["age","gender"]])
y_pred[:5]

Out[23]:

array([29.87948718, 30.55609756, 27.24722222, 30.84186047, 33.14666667])

将预测的值转成Series,并且注意索引号:

predict_bmi = pd.Series(y_pred, index=missing.index)
predict_bmi
1       29.879487
8       30.556098
13      27.247222
19      30.841860
27      33.146667
          ...    
5039    32.716000
5048    28.313636
5093    31.459322
5099    28.313636
5105    28.476923
Length: 201, dtype: float64

填充到原来的df数据中:

df.loc[missing.index, "bmi"] = predict_bmi

进行上面的预测和填充之后,我们再次查看缺失值情况,发现已经没有任何缺失值:

df.isnull().sum()
id                   0
gender               0
age                  0
hypertension         0
heart_disease        0
ever_married         0
work_type            0
Residence_type       0
avg_glucose_level    0
bmi                  0
smoking_status       0
stroke               0
dtype: int64

数据EDA

variables = [variable for variable in df.columns if variable not in ['id','stroke']]

# 除去id号和是否中风外的全部字段
variables
['gender',
 'age',
 'hypertension',
 'heart_disease',
 'ever_married',
 'work_type',
 'Residence_type',
 'avg_glucose_level',
 'bmi',
 'smoking_status']

连续型变量

几点结论:

  • 年龄age:整体分布比较均衡,不同年龄段的人数差异小
  • 血糖水平:主要集中在100以下
  • bmi指标:呈现一定的正态分布

中风和未中风

上面我们查看了连续型变量的分布情况;可以看到bmi呈现明显的左偏态的分布。下面我们对比中风和未中风的情况:

从3个密度图中能够观察到:从上面的密度图中可以看出来:对于是否中风,年龄age是一个最主要的因素

对比不同年龄段的血糖和BMI指数

px.scatter(df,x="age",
           y="avg_glucose_level",
           color="stroke",
           trendline='ols'
          )

年龄和血糖、bmi关系

px.scatter(df,x="age",
           y="bmi",
           color="stroke",
           trendline='ols'
          )

年龄和患病几率

从散点分布图中看到:年龄可能真的是一个比较重要的因素,和BMI以及平均的血糖水平有着一定的关系。

可能随着年龄的增长,风险在增加。果真如此吗?

上面的图形说明了两点:

  1. 年龄越大,中风的几率的确越来越高
  2. 中风的几率是非常低的(y轴的值很低),这是由于中风和未中风的样本不均衡造成的

原数据5000个样本中只有249个中风样本,比例接近1:20

样本不均衡

整体属性分布

首先我们剔除gender中为Other的情况

In [34]:

str_only = df[df['stroke'] == 1]   # 中风
no_str_only = df[df['stroke'] == 0]  # 未中风

In [35]:

len(str_only) 

Out[35]:

249

In [36]:

# 剔除other
no_str_only = no_str_only[(no_str_only['gender'] != 'Other')]

比较在不同的属性下中风和未中风的情况:

建模

模型baseline

In [38]:

len(str_only)

Out[38]:

249

In [39]:

249 / len(df)  

Out[39]:

0.0487279843444227

说明总共有249个人是中风的。本数据的总人数是len(df),根据下面的表达式能够得到本次模型的baseline。

也就说,对于阳性中风患者的召回率,一个好的目标是4.8%。

字段编码

对4个字符型的字段进行编码工作:

In [40]:

df['gender'] = df['gender'].replace({'Male':0,
                                     'Female':1,
                                     'Other':-1}
                                   ).astype(np.uint8)

df['Residence_type'] = df['Residence_type'].map({'Rural':0,
                                                 'Urban':1}
                                               ).astype(np.uint8)

df['work_type'] = df['work_type'].map({'Private':0,
                                       'Self-employed':1,
                                       'Govt_job':2,
                                       'children':-1,
                                       'Never_worked':-2}
                                     ).astype(np.uint8)

df['ever_married'] = df['ever_married'].map({'No':0,'Yes':1}).astype(np.uint8)

df.head()

抽烟状态的独热码转换:

In [41]:

df["smoking_status"].value_counts()

Out[41]:

never smoked       1892
Unknown            1544
formerly smoked     885
smokes              789
Name: smoking_status, dtype: int64

In [42]:

df = df.join(pd.get_dummies(df["smoking_status"]))
df.drop("smoking_status",axis=1,inplace=True)

数据分割

In [43]:

# 选取特征
X  = df.drop("stroke",axis=1)
# 目标变量
y = df['stroke']
from sklearn.model_selection import train_test_split

# 3-7比例
X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.3, random_state=42)

上采样

前文中提到,本案例中风和未中风的数据比例接近1:20,在这里我们采样基于SMOTE的上采样方法

In [44]:

oversample = SMOTE()
X_train_smote, y_train_smote = oversample.fit_resample(X_train, y_train.ravel())

In [45]:

len(y_train_smote)

Out[45]:

2914

In [46]:

len(X_train_smote)

Out[46]:

2914

建模

采用3种不同的分类模型来建立模型:Random Forest, SVM, Logisitc Regression

In [47]:

rf_pipeline = Pipeline(steps = [('scale',StandardScaler()), # 标准化
                                ('RF',RandomForestClassifier(random_state=42))]  # 模型
                      )
svm_pipeline = Pipeline(steps = [('scale',StandardScaler()),
                                 ('SVM',SVC(random_state=42))])
logreg_pipeline = Pipeline(steps = [('scale',StandardScaler()),
                                    ('LR',LogisticRegression(random_state=42))])

10折交叉验证

3种模型得分对比

In [49]:

print('随机森林:', rf_cv.mean())
print('支持向量机:',svm_cv.mean())
print('逻辑回归:', logreg_cv.mean())
随机森林: 0.9628909366701726
支持向量机: 0.9363667907023254
逻辑回归: 0.8859930523017683

很明显:随机森林表现的最好!

模型训练fit

In [50]:

rf_pipeline.fit(X_train_smote,y_train_smote)

svm_pipeline.fit(X_train_smote,y_train_smote)

logreg_pipeline.fit(X_train_smote,y_train_smote)

Out[50]:

Pipeline(steps=[('scale', StandardScaler()),
                ('LR', LogisticRegression(random_state=42))])

In [51]:

# 3种模型预测

rf_pred =rf_pipeline.predict(X_test)
svm_pred = svm_pipeline.predict(X_test)
logreg_pred = logreg_pipeline.predict(X_test)

评价指标

In [52]:

# 1、混淆矩阵

rf_cm  = confusion_matrix(y_test, rf_pred )
svm_cm = confusion_matrix(y_test, svm_pred)
logreg_cm  = confusion_matrix(y_test, logreg_pred)

In [53]:

print(rf_cm)
print("----")
print(svm_cm)
print("----")
print(logreg_cm)
[[3338   66]
 [ 164    9]]
----
[[3196  208]
 [ 148   25]]
----
[[3138  266]
 [ 116   57]]

print('RF mean :',rf_f1)
print('SVM mean :',svm_f1)
print('LR mean :',logreg_f1)
RF mean : 0.07258064516129033
SVM mean : 0.1231527093596059
LR mean : 0.22983870967741934

随机森林模型的分类报告:

from sklearn.metrics import plot_confusion_matrix, classification_report

print(classification_report(y_test,rf_pred))

print('Accuracy Score: ',accuracy_score(y_test,rf_pred))
              precision    recall  f1-score   support

           0       0.95      0.98      0.97      3404
           1       0.12      0.05      0.07       173

    accuracy                           0.94      3577
   macro avg       0.54      0.52      0.52      3577
weighted avg       0.91      0.94      0.92      3577

Accuracy Score:  0.9357003075202683

随机森林模型调参

基于网格搜索的参数调优:

from sklearn.model_selection import GridSearchCV

n_estimators =[64,100,128,200]
max_features = [2,3,5,7]
bootstrap = [True,False]

param_grid = {'n_estimators':n_estimators,
             'max_features':max_features,
             'bootstrap':bootstrap}

rfc = RandomForestClassifier()
grid = GridSearchCV(rfc,param_grid)

grid.fit(X_train,y_train)
GridSearchCV(estimator=RandomForestClassifier(),
             param_grid={'bootstrap': [True, False],
                         'max_features': [2, 3, 5, 7],
                         'n_estimators': [64, 100, 128, 200]})
grid.best_params_  # 找到最优的参数
{'bootstrap': False, 'max_features': 3, 'n_estimators': 200}
# 再次建立随机森林模型

rfc = RandomForestClassifier(
    max_features=3,
    n_estimators=200,
    bootstrap=False)

rfc.fit(X_train_smote,y_train_smote)

rfc_tuned_pred = rfc.predict(X_test)
# 新的分类报告得分

print(classification_report(y_test,rfc_tuned_pred))

print('Accuracy Score: ',accuracy_score(y_test,rfc_tuned_pred))
print('F1 Score: ',f1_score(y_test,rfc_tuned_pred))
              precision    recall  f1-score   support

           0       0.95      0.98      0.97      3404
           1       0.05      0.02      0.03       173

    accuracy                           0.94      3577
   macro avg       0.50      0.50      0.50      3577
weighted avg       0.91      0.94      0.92      3577

Accuracy Score:  0.9362594352809617
F1 Score:  0.025641025641025644

逻辑回归模型调参

penalty = ['l1','l2']
C = [0.001, 0.01, 0.1, 1, 10, 100] 

log_param_grid = {'penalty': penalty, 
                  'C': C}

logreg = LogisticRegression()
grid = GridSearchCV(logreg,log_param_grid)
grid.fit(X_train_smote,y_train_smote)
GridSearchCV(estimator=LogisticRegression(),
             param_grid={'C': [0.001, 0.01, 0.1, 1, 10, 100],
                         'penalty': ['l1', 'l2']})
grid.best_params_
{'C': 1, 'penalty': 'l2'}
logreg_pipeline = Pipeline(steps = [('scale',StandardScaler()),
                                    ('LR',LogisticRegression(C=1,penalty='l2',random_state=42))])

logreg_pipeline.fit(X_train_smote,y_train_smote)

Out[65]:

Pipeline(steps=[('scale', StandardScaler()),
                ('LR', LogisticRegression(C=1, random_state=42))])

In [66]:

logreg_new_pred   = logreg_pipeline.predict(X_test) # 新预测

In [67]:

print(classification_report(y_test,logreg_new_pred))

print('Accuracy Score: ',accuracy_score(y_test,logreg_new_pred))
print('F1 Score: ',f1_score(y_test,logreg_new_pred))
              precision    recall  f1-score   support

           0       0.96      0.92      0.94      3404
           1       0.18      0.33      0.23       173

    accuracy                           0.89      3577
   macro avg       0.57      0.63      0.59      3577
weighted avg       0.93      0.89      0.91      3577

Accuracy Score:  0.8932065977075762
F1 Score:  0.22983870967741934

支持向量机调参

In [68]:

svm_param_grid = {
            'C': [0.1, 1, 10, 100, 1000],  
            'gamma': [1, 0.1, 0.01, 0.001, 0.0001], 
            'kernel': ['rbf']} 

svm = SVC(random_state=42)

grid = GridSearchCV(svm, svm_param_grid)

In [69]:

grid.fit(X_train_smote,y_train_smote)

Out[69]:

GridSearchCV(estimator=SVC(random_state=42),
             param_grid={'C': [0.1, 1, 10, 100, 1000],
                         'gamma': [1, 0.1, 0.01, 0.001, 0.0001],
                         'kernel': ['rbf']})

In [70]:

grid.best_params_

Out[70]:

{'C': 100, 'gamma': 0.0001, 'kernel': 'rbf'}

In [71]:

svm_pipeline = Pipeline(steps = [('scale',StandardScaler()),('SVM',SVC(C=100,gamma=0.0001,kernel='rbf',random_state=42))])

svm_pipeline.fit(X_train_smote,y_train_smote)

svm_tuned_pred   = svm_pipeline.predict(X_test)

In [72]:

print(classification_report(y_test,svm_tuned_pred))

print('Accuracy Score: ',accuracy_score(y_test,svm_tuned_pred))
print('F1 Score: ',f1_score(y_test,svm_tuned_pred))
              precision    recall  f1-score   support

           0       0.96      0.93      0.94      3404
           1       0.16      0.27      0.20       173

    accuracy                           0.90      3577
   macro avg       0.56      0.60      0.57      3577
weighted avg       0.92      0.90      0.91      3577

Accuracy Score:  0.8951635448700028
F1 Score:  0.19700214132762314

结论

  1. 在交叉验证的过程中,随机森林表现的最好。
  2. 3种模型的对比:随机森林的精度最好,但是F1-score缺失最低的
  3. 模型可能存在的特点:更能预测哪些人将会中风,而不是哪些人不会中风

猜你喜欢

转载自juejin.im/post/7114303398379257870