[信号处理] 基于 CNN 的心电图 (ECG) 信号分类的典型方法实现（tensorflow）

关于

本实验使用1维卷积神经网络实现心电信号的5分类。由于数据类别不均衡，这里使用典型的上采样方法，实现数据类别的均衡化处理。

工具

 

方法实现

数据加载

''' Read the CSV file datasets: NORMAL_LABEL=0 , ABNORMAL_LABEL=1,2,3,4,5 '''
ptbdb_abnormal=pd.read_csv(os.path.join('/input/heartbeat', 'ptbdb_abnormal.csv'),header=None)
ptbdb_normal=pd.read_csv(os.path.join('/input/heartbeat', 'ptbdb_normal.csv'),header=None)
ptbdb_train=pd.concat([ptbdb_abnormal,ptbdb_normal],ignore_index=True)
mitbih_train=pd.read_csv(os.path.join('/input/heartbeat', 'mitbih_train.csv'),header=None)
mitbih_test=pd.read_csv(os.path.join('/input/heartbeat', 'mitbih_test.csv'),header=None)


'''VISUALIZE THE [MITBIH_TRAIN] DATASET: You will observe that majority of the obsvns are of Label=0'''
fig,ax=plt.subplots(figsize=(8,8))
mitbih_train[187].value_counts().plot(ax=ax,kind='bar')
plt.show()


full_train=pd.concat([mitbih_train,ptbdb_train],ignore_index=True)
full_train.shape      #102106*188

'''BREAK DOWN THE FULL TRAIN DATASET INTO X & Y'''
full_train_y=full_train[187]
full_train_x=full_train.drop(columns=[187])

'''BREAK DOWN THE TEST DATASET INTO X & Y'''
mitbih_test_y=mitbih_test[187]
mitbih_test_x=mitbih_test.drop(columns=[187])

 信号可视化

plots= [['Normal Beat','Supraventricular Ectopic Beat'], ['Ventricular Ectopic Beat', 'Fusion Ectopic Beat'], ['Unknown']]
colors= [['green', 'orange'], ['red', 'blue'], ['grey']]
fig, axs = plt.subplots(3, 2, constrained_layout=True, figsize=(14,14))
fig.delaxes(axs[2,1])

for i in range(0,5,2):
    j=i//2
    axs[j][0].plot(mitbih_train[mitbih_train[187]==i%5].sample(1, random_state=100).iloc[0,:186], colors[j][0])
    axs[j][0].set_title('{}. {}'.format(i%5,plots[j][0]))
    
    if i%5!=4:
        axs[j][1].plot(mitbih_train[mitbih_train[187]==(i%5)+1].sample(1, random_state=100).iloc[0,:186], colors[j][1])
        axs[j][1].set_title('{}. {}'.format(i%5+1,plots[j][1]))

 使用上采样方法处理数据不均衡

'''RESAMPLING THE CLASSES TO HAVE EQUAL DATA DISTRIBUTION LED TO WORSEN PERFORMANCE (POSSIBLE UNDERFITTING REASON)'''
df0=mitbih_train[mitbih_train[187]==0].sample(n=20000,random_state=10)
df1=mitbih_train[mitbih_train[187]==1]
df2=mitbih_train[mitbih_train[187]==2]
df3=mitbih_train[mitbih_train[187]==3]
df4=mitbih_train[mitbih_train[187]==4]

df1_upsampled=resample(df1,replace=True,n_samples=20000,random_state=100)
df2_upsampled=resample(df2,replace=True,n_samples=20000,random_state=101)
df3_upsampled=resample(df3,replace=True,n_samples=20000,random_state=102)
df4_upsampled=resample(df4,replace=True,n_samples=20000,random_state=103)
train_df=pd.concat([df1_upsampled,df2_upsampled,df3_upsampled,df4_upsampled,df0])

print(train_df[187].value_counts())
plt.figure(figsize=(10,10))
plt.pie(train_df[187].value_counts(), labels=['N','Q','V','S','F'],
        colors=['orange','yellow','lightblue','lightgreen','grey'], autopct='%.2f%%')
plt.gcf().gca().add_artist(plt.Circle( (0,0), 0.7, color='white' ))
plt.title('Balanced classes after upsampling')
plt.show()

模型搭建

model=tf.keras.Sequential()
model.add(tf.keras.layers.Convolution1D(filters=50,kernel_size=20,activation='relu',kernel_initializer='glorot_uniform',input_shape=(187,1)))
#a1_0=> 187-20+1= 168,50
model.add(tf.keras.layers.MaxPool1D(pool_size=10,data_format='channels_first'))
#a1_1=> 50//10= 168,5
model.add(tf.keras.layers.Convolution1D(filters=20,kernel_size=15,activation='relu',kernel_initializer='glorot_uniform'))
#a2_0=> 168-15+1= 154,20
model.add(tf.keras.layers.MaxPool1D(pool_size=15,data_format='channels_first'))
#a2_1=> 20//15= 154,1
model.add(tf.keras.layers.Convolution1D(filters=10,kernel_size=10,activation='relu',kernel_initializer='glorot_uniform'))
#a3_0=> 154-10+1=145,10
model.add(tf.keras.layers.MaxPool1D(pool_size=10,data_format='channels_first'))
#a3_1=> 10//10=145,1
model.add(tf.keras.layers.Flatten())
#a4=> 145
model.add(tf.keras.layers.Dense(units=512,activation='relu',kernel_initializer='glorot_uniform'))
#a4=> 512
model.add(tf.keras.layers.Dense(units=128,activation='relu',kernel_initializer='glorot_uniform'))
#a5=> 128
model.add(tf.keras.layers.Dense(units=5,activation='softmax'))

model.compile(optimizer='Adam',loss='sparse_categorical_crossentropy',metrics=['acc'])
model.summary()

 模型训练

mitbih_test_x2=np.asarray(mitbih_test_x)
mitbih_test_x2=mitbih_test_x2.reshape(-1,187,1)


''' DATASET AFTER UPSAMPLING, WITH EVEN DISTRIBUTION ACROSS CLASSES '''
train_df_X=np.asarray(train_df.iloc[:,:187]).reshape(-1,187,1)
train_df_Y=train_df.iloc[:,187]
print(train_df_X.shape)


hist=model.fit(train_df_X,train_df_Y,batch_size=64,epochs=20)

 

代码获取

相关项目和问题，后台私信交流沟通。