deephit模型如何画校准曲线

Question

deephit模型如何画校准曲线 代码怎么实现，以下是目前的代码

num_durations = 130        #（最大生存时间/12）
labtrans = DeepHitSingle.label_transform(num_durations)
get_target = lambda df: (df['survival'].values, df['OS'].values)
y_train = labtrans.fit_transform(*get_target(df_train))
y_val = labtrans.transform(*get_target(df_val))
train = (x_train, y_train)
val = (x_val, y_val)
# 获取验证集的真实生存时间和事件状态  
durations_val, events_val = get_target(df_val)

durations_test, events_test = get_target(df_test)
in_features = x_train.shape[1]
num_nodes = [32, 64, 32]   #可变加层，调整节点数
out_features = labtrans.out_features
batch_norm = True
dropout = 0.5  #调整过拟合情况

net = tt.practical.MLPVanilla(in_features, num_nodes, out_features, batch_norm, dropout)
model = DeepHitSingle(net, tt.optim.AdamW, alpha=0.0001, sigma=0.1, duration_index=labtrans.cuts)     
batch_size = 256
lr_finder = model.lr_finder(x_train, y_train, batch_size, tolerance=3)
_ = lr_finder.plot()
lr_finder.get_best_lr()
model.optimizer.set_lr(lr_finder.get_best_lr())
epochs = 400
callbacks = [tt.callbacks.EarlyStopping()]
log = model.fit(x_train, y_train, batch_size, epochs, callbacks, val_data=val)
_ = log.plot()
surv = model.predict_surv_df(x_test)

Answer 1

引自免费微信小程序：皆我百晓生

在给定的代码基础上，我们可以添加一个函数来计算并绘制DeepHit模型的校准曲线。首先，我们需要导入所需的库，并定义一个函数来计算Brier分数，然后根据这个分数来绘制校准曲线。下面是完整的代码示例：

import numpy as np
import pandas as pd
from sklearn.metrics import brier_score_loss
import matplotlib.pyplot as plt
from collections import defaultdict
from typing import List, Tuple

def calculate_brier_scores(model, durations, events, n_intervals=10):
    """
    计算Brier分数
    :param model: 已训练好的DeepHit模型
    :param durations: 生存时间数组
    :param events: 事件发生标志数组
    :param n_intervals: 时间间隔数量
    :return: Brier分数列表
    """
    brier_scores = []
    
    # 将生存时间分为n个等间距的时间段
    interval_width = durations.max() / n_intervals
    intervals = np.arange(0, durations.max(), interval_width)
    interval_boundaries = [(intervals[i], intervals[i+1]) for i in range(len(intervals)-1)]

    for start, end in interval_boundaries:
        # 提取每个时间段内的样本
        mask = (durations >= start) & (durations < end)
        durations_in_interval = durations[mask]
        events_in_interval = events[mask]

        # 预测生存概率
        surv_probabilities = model.predict_surv_df(pd.DataFrame(x_test)[x_train.columns]).loc[:, 'survival_probability']

        # 计算Brier分数
        brier_score = brier_score_loss(events_in_interval, surv_probabilities[mask])
        brier_scores.append(brier_score)

    return brier_scores

def plot_calibration_curve(brier_scores, interval_boundaries):
    """
    绘制校准曲线
    :param brier_scores: Brier分数列表
    :param interval_boundaries: 时间段边界
    """
    plt.figure(figsize=(8, 6))
    plt.plot(interval_boundaries, brier_scores, label='Brier Score', marker='o')
    plt.xlabel('Time (months)')
    plt.ylabel('Brier Score')
    plt.title('Calibration Curve for DeepHit Model')
    plt.legend()
    plt.grid(True)
    plt.show()

# 调用上述函数计算Brier分数和绘制校准曲线
brier_scores = calculate_brier_scores(model, durations_val, events_val)
plot_calibration_curve(brier_scores, interval_boundaries)

这段代码首先定义了两个函数：calculate_brier_scores 和 plot_calibration_curve。前者用于计算Brier分数，后者则用来绘制校准曲线。接着，在主程序中调用这两个函数，传入已经训练好的模型、验证集中真实的生存时间和事件状态，以及要划分的时间间隔数量。最后，会显示一个表示模型预测准确性的校准曲线。

注意，你需要确保数据的格式与模型输入匹配，并且在运行这段代码之前，确保已经正确地完成了模型训练和验证。此外，你可以根据实际情况调整n_intervals参数来改变时间间隔的数量。