早期败血症预测比赛记录

比赛官网

https://physionet.org/challenge/2019/

代码

原始数据整合

首先，手头数据的长相是这样式儿的

一个psv文件中存储着一个患者随时间变化的数据，一行代表一天的数据

• 整合并去掉缺失比例大于0.48的列

def get_data():
    path = "trainingA"
    filenames = os.listdir(path)
    result = pd.DataFrame()
    for filename in filenames:
        temp = pd.read_csv(path + "/" + filename, sep='|')
        result = pd.concat([result, temp])
    nan_columns = result.isna().mean()>0.48
return result.drop(nan_columns, inplace=True, axis=1)

填充缺失值

values = get_data()
# 先前向填充，再均值填充
# values = values.fillna(method='pad')
# values = values.fillna(values.mean())
values.interpolate(method='linear', inplace=True, limit_direction='both')

更新label值

由于比赛要求对发病做出提前六天的预警，因此对于每个人，如果某一天发病即label变成了1，我们把他前六天的观测结果也改成1。

 更改label
def change_label(value):
    label = value['SepsisLabel']
    if 1 in label.tolist():
        index = np.argmax(label)
        label[max(0, index-6):index] = 1

change_label(values)

新增累积列

考虑到时间上的累积效应，对每一个变量计算前七天的累积加权值，并获得一列新的变量，其实就是attention机制，只不过是固定的alpha值。

# 获取相乘的系数
def get_coefficient(num):
    coe = np.arange(1, num+1)
    return coe/sum(coe)

# 添加累积列
def compute_accu(value):
    length, _ = value.shape
    columns = value.columns
    seven = get_coefficient(7)
    for i in range(len(columns)-1):
        newname = "Accu_"+columns[i]
        temp = np.ones((length, 1))
        for j in range(len(temp)):
            if j == 0:
                temp[j] = value[columns[i]][j]
            elif j < 7:
                cor = get_coefficient(j)
                temp[j] = sum(value[columns[i]][0:j]*cor)
            else:
                temp[j] = sum(value[columns[i]][j-7:j] * seven)
        value[newname] = temp

xgboost预测法

Y = value['SepsisLabel']
X = value.drop(['SepsisLabel'], axis = 1)
test_size = 0.2
X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=test_size)
model = XGBClassifier()
eval_set = [(X_test, y_test)]
model.fit(X_train, y_train, early_stopping_rounds=10, eval_metric="logloss", eval_set=eval_set, verbose=True)
plot_importance(model)
pyplot.show()
y_pred = model.predict(X_test)
predictions = [round(value) for value in y_pred]
accuracy = accuracy_score(y_test, predictions)
print("Accuracy: %.2f%%" % (accuracy * 100.0))

逻辑回归预测法

• 对数据进行归一化

# 归一化
def normallization(data):
    columns = data.columns.tolist()
    for c in columns:
        d = data[c]
        MAX = d.max()
        MIN = d.min()
        if (MAX == 1 and MIN == 0) :
            continue
        if (MAX != MIN) :
            data[c] = ((d - MIN) / (MAX - MIN))

• 评估模型准确率

normallization(value)
Y = value['SepsisLabel']
X = value.drop(['SepsisLabel'], axis = 1)
test_size = 0.2
X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=test_size)
lr = LogisticRegression(C=1e5)
lr.fit(X_train, y_train)
y_pred = lr.predict(X_test)
predictions = [round(value) for value in y_pred]
accuracy = accuracy_score(y_test, predictions)
print("Accuracy: %.2f%%" % (accuracy * 100.0))

预测函数

由于比赛要求提交可执行文件，因此重新写了一个预测函数

• 尝试获取逻辑回归的参数并重新计算

weight = lr.coef_
weight = weight.transpose()
intercept = lr.intercept_[0]

sigmod = lambda x: float(1.0/(1.0+np.exp(-x)))
# 预测函数
def predict(F, weight, intercept):
    numSamples, _ = np.shape(F)
    result = []
    scores = []
    for i in range(numSamples):
        # 分数这样子计算
        # print(F[i, :] * weight + intercept)
        scores.append(sigmod(F[i, :] * weight + intercept))
        result.append(int(F[i, :] * weight >= -intercept))
    return scores, result

predictions = predict(np.mat(X_test), weight, intercept)
accuracy1 = accuracy_score(y_test, predictions)
print("Accuracy: %.2f%%" % (accuracy * 100.0))

• 将参数融入到代码当中去

# 读取函数
def evaluate(filename):
    filename = filename+'.psv'
    nan_columns = ['Temp', 'EtCO2', 'BaseExcess', 'HCO3', 'FiO2', 'pH', 'PaCO2', 'SaO2', 'AST', 'BUN', 'Alkalinephos', 'Calcium', 'Chloride', 'Creatinine', 'Bilirubin_direct', 'Glucose', 'Lactate', 'Magnesium', 'Phosphate', 'Potassium', 'Bilirubin_total', 'TroponinI', 'Hct', 'Hgb', 'PTT', 'WBC', 'Fibrinogen', 'Platelets']
    value = pd.read_csv(filename, sep='|')
    value.drop(nan_columns, inplace=True, axis=1)
    M =  {'HR': 0.24972140809778023, 'O2Sat': 0.9651197941386043, 'SBP': 0.3822739233740868, 'MAP': 0.21045135092077905, 'DBP': 0.1469540894675809, 'Resp': 0.2613221129135702, 'Age': 0.6334712914523257, 'Gender': 0.5777212530766943, 'Unit1': 0.5017673613352216, 'Unit2': 0.49823263866477846, 'HospAdmTime': 0.9796459436158841, 'ICULOS': 0.07820453171621204, 'SepsisLabel': 0.032193770049923125, 'Accu_HR': 0.3961986226205406, 'Accu_O2Sat': 0.9660205798556376, 'Accu_SBP': 0.3878985082015177, 'Accu_MAP': 0.26354211253459175, 'Accu_DBP': 0.1470331983530533, 'Accu_Resp': 0.2595029445367898, 'Accu_Age': 0.6334712914523257, 'Accu_Gender': 0.5777212530766943, 'Accu_Unit1': 0.5017673613352216, 'Accu_Unit2': 0.49823263866477846, 'Accu_HospAdmTime': 0.9796459436158841, 'Accu_ICULOS': 0.07065019821885823}
    value.interpolate(method='linear', inplace=True, limit_direction='both')
    compute_accu(value)
    # print(value.mean())
    if 'SepsisLabel' in value.columns.tolist():
        value.drop(['SepsisLabel'], axis=1, inplace=True)
    value.fillna(M, inplace=True)
    weight = np.array([[3.10101382],
           [0.9402448],
           [-0.22685776],
           [-0.80695164],
           [-1.28859308],
           [0.61754885],
           [0.01568997],
           [0.08981491],
           [-1.48981012],
           [-1.65665569],
           [-0.61166079],
           [6.84160107],
           [0.33597397],
           [1.07752734],
           [0.38562541],
           [-1.62192069],
           [1.55039952],
           [3.02623835],
           [0.01568997],
           [0.08981491],
           [-1.48981012],
           [-1.65665569],
           [-0.61166079],
           [-1.71220229]])
    intercept = -3.146465793867411
    scores, lables = predict(np.mat(value), weight, intercept)
    with open(filename+'.out', 'w') as f:
         for (s, l) in zip(scores, lables):
             f.write('%g|%d\n' % (s, l))

收获

了解了一些医学常识，提升了自己的数据分析和处理能力，通过在群里和大家交流收获了不少有趣的想法，对于机器学习类比赛有了更深入的了解。取得了进入复赛的好成绩，但由于不满足后续参加会议的要求不得不终止比赛。