{
"cells": [
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "azrczYYVvEQb"
},
"source": [
"# Survival Analysis for Deep Learning\n",
"\n",
"Most machine learning algorithms have been developed to perform classification or regression. However, in clinical research we often want to estimate the time to and event, such as death or recurrence of cancer, which leads to a special type of learning task that is distinct from classification and regression. This task is termed *survival analysis*, but is also referred to as time-to-event analysis or reliability analysis.\n",
"Many machine learning algorithms have been adopted to perform survival analysis:\n",
"[Support Vector Machines](https://scholar.google.com/scholar?oi=bibs&cluster=18092275419152143443),\n",
"[Random Forest](https://scholar.google.com/scholar?cluster=16319510831191377024),\n",
"or [Boosting](https://scholar.google.com/scholar?cluster=14069073471114367075).\n",
"It has only been recently that survival analysis entered the era of deep learning, which is the focus of this post.\n",
"\n",
"You will learn how to train a convolutional neural network to predict time to a (generated) event from MNIST images, using a loss function specific to survival analysis. The [first part](#Primer-on-Survival-Analysis), will cover some basic terms and quantities used in survival analysis (feel free to skip this part if you are already familiar). In the [second part](#Generating-Synthetic-Survival-Data-from-MNIST), we will generate synthetic survival data from MNIST images and visualize it. In the [third part](#Cox's-Proportional-Hazards-Model), we will briefly revisit the most popular survival model of them all and learn how it can be used as a loss function for training a neural network.\n",
"[Finally](#Creating-a-Convolutional-Neural-Network-for-Survival-Analysis-on-MNIST), we put all the pieces together and train a convolutional neural network on MNIST and predict survival functions on the test data.\n",
"\n",
"\n",
"## Requirements:\n",
"\n",
"Please make sure you have the following packages installed. All are available via [PyPI](https://pypi.org) or [Anaconda](https://www.anaconda.com/distribution/).\n",
"\n",
"- [numpy](https://www.numpy.org/)\n",
"- [matplotlib](https://matplotlib.org/)\n",
"- [pandas](https://pandas.pydata.org/)\n",
"- [scikit-survival](https://github.com/sebp/scikit-survival/)\n",
"- [tensorflow](https://www.tensorflow.org/) >= 2.0.0\n",
"\n",
"You can also run this notebook in [Google Colaboratory](https://colab.research.google.com/github/sebp/survival-cnn-estimator/blob/master/tutorial_tf2.ipynb) and install scikit-survival using the command below."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 1000
},
"colab_type": "code",
"executionInfo": {
"elapsed": 12742,
"status": "ok",
"timestamp": 1589637598598,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "kb7TWFXivEQc",
"jupyter": {
"outputs_hidden": true
},
"outputId": "17fa15cf-e4dd-441e-e667-31e39722f5e8"
},
"outputs": [],
"source": [
"!pip uninstall --yes --quiet osqp\n",
"!pip install scikit-survival"
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 34
},
"colab_type": "code",
"executionInfo": {
"elapsed": 2942,
"status": "ok",
"timestamp": 1589637614533,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "ThvRyUyVvEQf",
"jupyter": {
"outputs_hidden": true
},
"outputId": "872381c5-d695-4bed-a4f8-bcc6d933aa34"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Using Tensorflow: 2.2.0\n"
]
}
],
"source": [
"from typing import Any, Dict, Iterable, Sequence, Tuple, Optional, Union\n",
"import numpy as np\n",
"import matplotlib.pyplot as plt\n",
"import pandas as pd\n",
"from pathlib import Path\n",
"from sksurv.nonparametric import kaplan_meier_estimator\n",
"from sksurv.metrics import concordance_index_censored\n",
"import tensorflow as tf\n",
"from tensorflow.keras.datasets import mnist\n",
"\n",
"print(\"Using Tensorflow:\", tf.__version__)"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"from distutils.version import LooseVersion\n",
"\n",
"assert LooseVersion(tf.__version__) >= LooseVersion(\"2.0.0\"), \\\n",
" \"This notebook requires TensorFlow 2.0 or above.\""
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "cdL25jjJvEQm"
},
"source": [
"## Primer on Survival Analysis\n",
"\n",
"The objective in survival analysis is to establish a connection between covariates and the time of an event. The name *survival analysis* originates from clinical research, where predicting the time to death, i.e., survival, is often the main objective. Survival analysis is a type of regression problem (one wants to predict a continuous value), but with a twist. It differs from traditional regression by the fact that parts of the training data can only be partially observed – they are *censored*.\n",
"\n",
"As an example, consider a clinical study that has been carried out over a 1 year period as in the figure below."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "7UZNLT1GvEQm"
},
"source": [
""
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "QqEgCNOxvEQn"
},
"source": [
"Patient A was lost to follow-up after three months with no recorded event, patient B experienced an event four and a half months after enrollment, patient C withdrew from the study two months after enrollment, and patient E did not experience any event before the study ended. Consequently, the *exact time* of an event could only be recorded for patients B and D; their records are *uncensored*. For the remaining patients it is unknown whether they did or did not experience an event after termination of the study. The only valid information that is available for patients A, C, and E is that they were event-free up to their last follow-up. Therefore, their records are *censored*.\n",
"\n",
"Formally, each patient record consists of the time $t>0$ when an event occurred or the time $c>0$ of censoring. Since censoring and experiencing and event are mutually exclusive, it is common to define an event indicator $\\delta \\in \\{0;1\\}$ and the observable survival time $y>0$. The observable time $y$ of a right censored time of event is defined as\n",
"\n",
"$$\n",
"y = \\min(t, c) = \n",
"\\begin{cases} \n",
"t & \\text{if } \\delta = 1 , \\\\\n",
"c & \\text{if } \\delta = 0 .\n",
"\\end{cases}\n",
"$$\n",
"\n",
"Consequently, survival analysis demands for models that take partially observed, i.e., censored, event times into account.\n",
"\n",
"\n",
"## Basic Quantities\n",
"\n",
"Typically, the survival time is modelled as a continuous non-negative random variable $T$, from which basic quantities for time-to-event analysis can be derived, most importantly, the *survival function* and the *hazard function*.\n",
"\n",
"- The **survival function** $S(t)$ returns the probability of survival beyond time $t$ and is defined as $S(t) = P(T > t)$. It is non-increasing with $S(0) = 1$, and $S(\\infty) = 0$.\n",
"- The **hazard function** $h(t)$ denotes an approximate probability (it is not bounded from above) that an event occurs in the small time interval $[t; t + \\Delta[$, under the condition that an individual would remain event-free up to time $t$:\n",
"$$\n",
"h(t) = \\lim_{\\Delta t \\rightarrow 0} \\frac{P(t \\leq T < t + \\Delta t \\mid T \\geq t)}{\\Delta t} \\geq 0\n",
"$$\n",
"Alternative names for the hazard function are conditional failure rate, conditional mortality rate, or instantaneous failure rate. In contrast to the survival function, which describes the absence of an event, the hazard function provides information about the occurrence of an event."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "Pb3xh-qGvEQn"
},
"source": [
"## Generating Synthetic Survival Data from MNIST\n",
"\n",
"To start off, we are using images from the MNIST dataset and will synthetically generate\n",
"survival times based on the digit each image represents.\n",
"We associate a survival time (or risk score) with each class of the ten digits in MNIST. First, we randomly assign each class label to one of four overall risk groups, such that some digits will correspond to better and others to worse survival. Next, we generate risk scores that indicate how big the risk of experiencing an event is, relative to each other."
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "5Wukwbc5vEQo",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"(x_train, y_train), (x_test, y_test) = mnist.load_data()\n",
"\n",
"x_train = x_train.astype(np.float32) / 255.\n",
"x_test = x_test.astype(np.float32) / 255.\n",
"\n",
"y_train = y_train.astype(np.int32)\n",
"y_test = y_test.astype(np.int32)\n",
"\n",
"y = np.concatenate((y_train, y_test))"
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 390
},
"colab_type": "code",
"executionInfo": {
"elapsed": 1008,
"status": "ok",
"timestamp": 1589637619418,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "XdgLp1FkvEQq",
"jupyter": {
"outputs_hidden": true
},
"outputId": "463ac931-4868-443e-a32f-de3a8de7a283"
},
"outputs": [
{
"data": {
"text/html": [
"
\n",
"\n",
"
\n",
" \n",
"
\n",
"
\n",
"
risk_score
\n",
"
risk_group
\n",
"
\n",
"
\n",
"
class_label
\n",
"
\n",
"
\n",
"
\n",
" \n",
" \n",
"
\n",
"
0
\n",
"
3.071
\n",
"
3
\n",
"
\n",
"
\n",
"
1
\n",
"
2.555
\n",
"
2
\n",
"
\n",
"
\n",
"
2
\n",
"
0.058
\n",
"
0
\n",
"
\n",
"
\n",
"
3
\n",
"
1.790
\n",
"
1
\n",
"
\n",
"
\n",
"
4
\n",
"
2.515
\n",
"
2
\n",
"
\n",
"
\n",
"
5
\n",
"
3.031
\n",
"
3
\n",
"
\n",
"
\n",
"
6
\n",
"
1.750
\n",
"
1
\n",
"
\n",
"
\n",
"
7
\n",
"
2.475
\n",
"
2
\n",
"
\n",
"
\n",
"
8
\n",
"
0.018
\n",
"
0
\n",
"
\n",
"
\n",
"
9
\n",
"
2.435
\n",
"
2
\n",
"
\n",
" \n",
"
\n",
"
"
],
"text/plain": [
" risk_score risk_group\n",
"class_label \n",
"0 3.071 3\n",
"1 2.555 2\n",
"2 0.058 0\n",
"3 1.790 1\n",
"4 2.515 2\n",
"5 3.031 3\n",
"6 1.750 1\n",
"7 2.475 2\n",
"8 0.018 0\n",
"9 2.435 2"
]
},
"execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"def make_risk_score_for_groups(y: np.ndarray,\n",
" n_groups: int = 4,\n",
" seed: int = 89) -> Tuple[pd.DataFrame, np.ndarray]:\n",
" rnd = np.random.RandomState(seed)\n",
"\n",
" # assign class labels `y` to one of `n_groups` risk groups\n",
" classes = np.unique(y)\n",
" group_assignment = {}\n",
" group_members = {}\n",
" groups = rnd.randint(n_groups, size=classes.shape)\n",
" for label, group in zip(classes, groups):\n",
" group_assignment[label] = group\n",
" group_members.setdefault(group, []).append(label)\n",
"\n",
" # assign risk score to each class label in `y`\n",
" risk_per_class = {}\n",
" for label in classes:\n",
" group_idx = group_assignment[label]\n",
" group = group_members[group_idx]\n",
" label_idx = group.index(label)\n",
" group_size = len(group)\n",
"\n",
" # allow risk scores in each group to vary slightly\n",
" risk_score = np.sqrt(group_idx + 1e-4) * 1.75\n",
" risk_score -= (label_idx - (group_size // 2)) / 25.\n",
" risk_per_class[label] = risk_score\n",
"\n",
" assignment = pd.concat((\n",
" pd.Series(risk_per_class, name=\"risk_score\"),\n",
" pd.Series(group_assignment, name=\"risk_group\")\n",
" ), axis=1).rename_axis(\"class_label\")\n",
" \n",
" risk_scores = np.array([risk_per_class[yy] for yy in y])\n",
" return assignment, risk_scores\n",
"\n",
"\n",
"risk_score_assignment, risk_scores = make_risk_score_for_groups(y)\n",
"\n",
"risk_score_assignment.round(3)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "EqQR-m5HvEQt"
},
"source": [
"We can see that class labels 2 and 8 belong to risk group 0, which has the lowest risk (close to zero). Risk group 1 corresponds to a risk score of about 1.7, risk group 2 of about 2.5, and risk group 3 is the group with the highest risk score of about 3.\n",
"\n",
"To generate survival times from risk scores, we are going to follow the protocol of\n",
"[Bender et al](https://scholar.google.com/scholar?cluster=11575471310627475868). We choose the exponential distribution for the survival time. Its probability density function is $f(t\\,|\\,\\lambda) = \\lambda \\exp(-\\lambda t)$, where $\\lambda > 0$ is a scale parameter that is the inverse of the expectation: $E(T) = \\frac{1}{\\lambda}$. The exponential distribution results in a relatively simple time-to-event model with no memory, because the hazard rate is constant: $h(t) = \\lambda$. For more complex cases, refer to the paper by [Bender et al](https://scholar.google.com/scholar?cluster=11575471310627475868).\n",
"\n",
"Here, we choose $\\lambda$ such that the mean survival time is 365 days. Finally, we randomly censor survival times drawing times of censoring from a uniform distribution such that we approximately obtain the desired amount of 45% censoring. The generated survival data comprises an observed time and a boolean event indicator for each MNIST image."
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 51
},
"colab_type": "code",
"executionInfo": {
"elapsed": 1071,
"status": "ok",
"timestamp": 1589637621537,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "ceJ4YhKEvEQu",
"jupyter": {
"outputs_hidden": true
},
"outputId": "8bb9bdca-7466-423f-d46f-269aff1cf0c6"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"46.19% samples are right censored in training data.\n",
"46.33% samples are right censored in test data.\n"
]
}
],
"source": [
"class SurvivalTimeGenerator:\n",
"\n",
" def __init__(self,\n",
" num_samples: int,\n",
" mean_survival_time: float,\n",
" prob_censored: float) -> None:\n",
" self.num_samples = num_samples\n",
" self.mean_survival_time = mean_survival_time\n",
" self.prob_censored = prob_censored\n",
"\n",
" def gen_censored_time(self,\n",
" risk_score: np.ndarray,\n",
" seed: int = 89) -> Tuple[np.ndarray,np.ndarray]:\n",
" rnd = np.random.RandomState(seed)\n",
" # generate survival time\n",
" baseline_hazard = 1. / self.mean_survival_time\n",
" scale = baseline_hazard * np.exp(risk_score)\n",
" u = rnd.uniform(low=0, high=1, size=risk_score.shape[0])\n",
" t = -np.log(u) / scale\n",
"\n",
" # generate time of censoring\n",
" qt = np.quantile(t, 1.0 - self.prob_censored)\n",
" c = rnd.uniform(low=t.min(), high=qt)\n",
"\n",
" # apply censoring\n",
" observed_event = t <= c\n",
" observed_time = np.where(observed_event, t, c)\n",
" return observed_time, observed_event\n",
"\n",
"\n",
"surv_gen = SurvivalTimeGenerator(\n",
" num_samples=y.shape[0],\n",
" mean_survival_time=365.,\n",
" prob_censored=.45\n",
")\n",
"time, event = surv_gen.gen_censored_time(risk_scores)\n",
"time_train = time[:y_train.shape[0]]\n",
"event_train = event[:y_train.shape[0]]\n",
"time_test = time[y_train.shape[0]:]\n",
"event_test = event[y_train.shape[0]:]\n",
"\n",
"print(\"%.2f%% samples are right censored in training data.\" % (np.sum(~event_train) * 100. / len(event_train)))\n",
"print(\"%.2f%% samples are right censored in test data.\" % (np.sum(~event_test) * 100. / len(event_test)))"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "37l5fJZvvEQx"
},
"source": [
"We can use the generated censored data and estimate the survival function $S(t)$ to see what the risk scores actually mean in terms of survival. We stratify the training data by class label, and estimate the corresponding survival function using the non-parametric [Kaplan-Meier estimator](https://en.wikipedia.org/wiki/Kaplan%E2%80%93Meier_estimator)."
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 330
},
"colab_type": "code",
"executionInfo": {
"elapsed": 1834,
"status": "ok",
"timestamp": 1589637624905,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "S1QJulD8vEQx",
"jupyter": {
"outputs_hidden": true
},
"outputId": "6e7b5ad1-0118-4e54-eade-d947fd7f5012"
},
"outputs": [
{
"data": {
"text/plain": [
""
]
},
"execution_count": 6,
"metadata": {},
"output_type": "execute_result"
},
{
"data": {
"image/png": "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\n",
"text/plain": [
""
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"styles = ('-', '--', '-.', ':')\n",
"\n",
"plt.figure(figsize=(6, 4.5))\n",
"for row in risk_score_assignment.itertuples():\n",
" mask = y_train == row.Index\n",
" coord_x, coord_y = kaplan_meier_estimator(event_train[mask], time_train[mask])\n",
" ls = styles[row.risk_group]\n",
" plt.step(coord_x, coord_y, where=\"post\", label=f\"Class {row.Index}\", linestyle=ls)\n",
"plt.ylim(0, 1)\n",
"plt.ylabel(\"Probability of survival $P(T > t)$\")\n",
"plt.xlabel(\"Time $t$\")\n",
"plt.grid()\n",
"plt.legend()"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "fps2XQo0vEQ0"
},
"source": [
"Classes 0 and 5 (dotted lines) correspond to risk group 3, which has the highest risk score. The corresponding survival functions drop most quickly, which is exactly what we wanted. On the other end of the spectrum are classes 2 and 8 (solid lines) belonging to risk group 0 with the lowest risk.\n",
"\n",
"## Evaluating Predictions\n",
"\n",
"One important aspect for survival analysis is that both the training data and the test data are subject to censoring, because we are unable to observe the exact time of an event no matter how the data was split. Therefore, performance measures need to account for censoring. The most widely used performance measure is Harrell's concordance index. Given a set of (predicted) risk scores and observed times, it checks whether the ordering by risk scores is concordant with the ordering by actual survival time. While Harrell's concordance index is widely used, it has its flaws, in particular when data is highly censored. Please refer to my [previous post on evaluating survival models](https://k-d-w.org/blog/111/evaluating-survival-models) for more details.\n",
"\n",
"We can take the risk score from which we generated survival times to check how good a model would perform if we knew the actual risk score."
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 34
},
"colab_type": "code",
"executionInfo": {
"elapsed": 2072,
"status": "ok",
"timestamp": 1589637626269,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "KlHkhxZdvEQ0",
"jupyter": {
"outputs_hidden": true
},
"outputId": "493e3329-96bf-4a23-f0c5-e2b7214fb276"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Concordance index on test data with actual risk scores: 0.705\n"
]
}
],
"source": [
"cindex = concordance_index_censored(event_test, time_test, risk_scores[y_train.shape[0]:])\n",
"\n",
"print(f\"Concordance index on test data with actual risk scores: {cindex[0]:.3f}\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "nHRvXxekvEQ3"
},
"source": [
"Surprisingly, we do not obtain a perfect result of 1.0. The reason for this is that generated survival times are randomly distributed based on risk scores and not deterministic functions of the risk score. Therefore, any model we will train on this data should not be able to exceed this performance value."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "TBqC0fjPvEQ3"
},
"source": [
"## Cox's Proportional Hazards Model\n",
"\n",
"By far the most widely used model to learn from censored survival data, is\n",
"[Cox's proportional hazards model](https://scholar.google.com/scholar?cluster=17981786408695305487) model.\n",
"It models the hazard function $h(t_i)$\n",
"of the $i$-th subject, conditional on the feature vector $\\mathbf{x}_i \\in \\mathbb{R}^p$,\n",
"as the product of an unspecified baseline hazard function $h_0$ (more on that later) and an\n",
"exponential function of the linear model $\\mathbf{x}_i^\\top \\mathbf{\\beta}$:\n",
"$$\n",
"h(t | x_{i1}, \\ldots, x_{ip}) = h_0(t) \\exp \\left( \\sum_{j=1}^p x_{ij} \\beta_j \\right)\n",
"\\Leftrightarrow\n",
"\\log \\frac{h(t | \\mathbf{x}_i)}{h_0 (t)} = \\mathbf{x}_i^\\top \\mathbf{\\beta} ,\n",
"$$\n",
"where $\\mathbf{\\beta} \\in \\mathbb{R}^p$ are the coefficients associated with each of the\n",
"$p$ features, and no intercept term is included in the model.\n",
"The key is that the hazard function is split into two parts: the baseline hazard function $h_0$ only depends on the time $t$, whereas the exponential is independent of time and only depends on the covariates $\\mathbf{x}_i$.\n",
"\n",
"Cox's proportional hazards model is fitted by maximizing the partial likelihood function, which is based on the probability that the $i$-th individual experiences\n",
"an event at time $t_i$, given that there is one event at time point $t_i$.\n",
"As we will see, by specifying the hazard function as above, the baseline hazard function $h_0$\n",
"can be eliminated and does not need be defined for finding the coefficients $\\mathbf{\\beta}$.\n",
"Let $\\mathcal{R}_i = \\{ j\\,|\\,y_j \\geq y_i \\}$\n",
"be the risk set, i.e., the set of subjects who remained event-free shortly before time point $y_i$,\n",
"and $I(\\cdot)$ the indicator function, then we have\n",
"\n",
"$$\n",
"\\begin{split}\n",
"&P(\\text{subject experiences event at $y_i$} \\mid \\text{one event at $y_i$}) \\\\\n",
"=& \\frac{P(\\text{subject experiences event at $y_i$} \\mid \\text{event-free up to $y_i$})}\n",
"{P (\\text{one event at $y_i$} \\mid \\text{event-free up to $y_i$})} \\\\\n",
"=& \\frac{h(y_i | \\mathbf{x}_i)}{ \\sum_{j=1}^n I(y_j \\geq y_i) h(y_j | \\mathbf{x}_j) } \\\\\n",
"=& \\frac{h_0(y_i) \\exp(\\mathbf{x}_i^\\top \\mathbf{\\beta})}\n",
"{ \\sum_{j=1}^n I(y_j \\geq y_i) h_0(y_j) \\exp(\\mathbf{x}_j^\\top \\mathbf{\\beta}) } \\\\\n",
"=& \\frac{\\exp( \\mathbf{x}_i^\\top \\beta)}{\\sum_{j \\in \\mathcal{R}_i} \\exp( \\mathbf{x}_j^\\top \\beta)} .\n",
"\\end{split}\n",
"$$\n",
"\n",
"By multiplying the conditional probability from above for all patients who experienced an event, and taking the logarithm, we obtain the *partial likelihood function*:\n",
"\n",
"$$\n",
"\\widehat{\\mathbf{\\beta}} = \\arg\\max_{\\mathbf{\\beta}}~\n",
"\\log\\,PL(\\mathbf{\\beta}) = \\sum_{i=1}^n \\delta_i \\left[ \\mathbf{x}_i^\\top \\mathbf{\\beta}\n",
"- \\log \\left( \\sum_{j \\in \\mathcal{R}_i} \\exp( \\mathbf{x}_j^\\top \\mathbf{\\beta}) \\right) \\right] .\n",
"$$\n",
"\n",
"\n",
"## Non-linear Survival Analysis with Neural Networks\n",
"\n",
"Cox's proportional hazards model as described above is a linear model, i.e., the predicted risk score is a linear combination of features. However, the model can easily be extended to the non-linear case by just replacing the linear predictor with the output of a neural network with parameters $\\mathbf{\\Theta}$."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "SK1NKDwYvEQ4"
},
"source": [
""
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "LOUPxXZkvEQ4"
},
"source": [
"This has been realized early on and was originally proposed in the work of [Faraggi and Simon](https://scholar.google.com/scholar?cluster=8523249692591517459) back in 1995. Farragi and Simon explore multilayer perceptrons, but the same loss can be used in combination with more advanced architectures such as convolutional neural networks or recurrent neural networks.\n",
"Therefore, it is natural to also use the same loss function in the era of deep learning.\n",
"However, this transition is not so easy as it may seem and comes with some caveats, both for training and for evaluation.\n",
"\n",
"### Computing the Loss Function\n",
"\n",
"When implementing the Cox PH loss function, the problematic part is the inner sum over the risk set:\n",
"$\\sum_{j \\in \\mathcal{R}_i} \\exp( \\mathbf{x}_j^\\top \\mathbf{\\beta})$. Note that the risk set is defined as $\\mathcal{R}_i = \\{ j\\,|\\,y_j \\geq y_i \\}$, which implies an ordering according to observed times $y_i$, which may lead to quadratic complexity if implemented naively. Ideally, we want to sort the data once in descending order by survival time and then incrementally update the inner sum, which leads to a linear complexity to compute the loss (ignoring the time for sorting).\n",
"\n",
"Another problem is that the risk set for the subject with the smallest uncensored survival time is over the whole dataset. This is usually impractical, because we may not be able to keep the whole dataset in GPU memory. If we use mini-batches instead, as it's the norm, (i) we cannot compute the exact loss, because we may not have access to all samples in the risk set, and (ii) we need to sort each mini-batch by observed time, instead of sorting the whole data once.\n",
"\n",
"For practical purposes, computing the Cox PH loss over a mini-batch is usually fine, as long as the batch contains several uncensored samples, because otherwise the outer sum in the partial likelihood function would be over an empty set.\n",
"Here, we implement the sum over the risk set by multiplying the exponential of the predictions (as a row vector) by a squared boolean matrix that contains each sample's risk set as its rows. The sum over the risk set for each sample is then equivalent to a row-wise summation."
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "Y-HgnZlEvEQ5",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"def _make_riskset(time: np.ndarray) -> np.ndarray:\n",
" \"\"\"Compute mask that represents each sample's risk set.\n",
"\n",
" Parameters\n",
" ----------\n",
" time : np.ndarray, shape=(n_samples,)\n",
" Observed event time sorted in descending order.\n",
"\n",
" Returns\n",
" -------\n",
" risk_set : np.ndarray, shape=(n_samples, n_samples)\n",
" Boolean matrix where the `i`-th row denotes the\n",
" risk set of the `i`-th instance, i.e. the indices `j`\n",
" for which the observer time `y_j >= y_i`.\n",
" \"\"\"\n",
" assert time.ndim == 1, \"expected 1D array\"\n",
"\n",
" # sort in descending order\n",
" o = np.argsort(-time, kind=\"mergesort\")\n",
" n_samples = len(time)\n",
" risk_set = np.zeros((n_samples, n_samples), dtype=np.bool_)\n",
" for i_org, i_sort in enumerate(o):\n",
" ti = time[i_sort]\n",
" k = i_org\n",
" while k < n_samples and ti == time[o[k]]:\n",
" k += 1\n",
" risk_set[i_sort, o[:k]] = True\n",
" return risk_set\n",
"\n",
"\n",
"class InputFunction:\n",
" \"\"\"Callable input function that computes the risk set for each batch.\n",
" \n",
" Parameters\n",
" ----------\n",
" images : np.ndarray, shape=(n_samples, height, width)\n",
" Image data.\n",
" time : np.ndarray, shape=(n_samples,)\n",
" Observed time.\n",
" event : np.ndarray, shape=(n_samples,)\n",
" Event indicator.\n",
" batch_size : int, optional, default=64\n",
" Number of samples per batch.\n",
" drop_last : int, optional, default=False\n",
" Whether to drop the last incomplete batch.\n",
" shuffle : bool, optional, default=False\n",
" Whether to shuffle data.\n",
" seed : int, optional, default=89\n",
" Random number seed.\n",
" \"\"\"\n",
"\n",
" def __init__(self,\n",
" images: np.ndarray,\n",
" time: np.ndarray,\n",
" event: np.ndarray,\n",
" batch_size: int = 64,\n",
" drop_last: bool = False,\n",
" shuffle: bool = False,\n",
" seed: int = 89) -> None:\n",
" if images.ndim == 3:\n",
" images = images[..., np.newaxis]\n",
" self.images = images\n",
" self.time = time\n",
" self.event = event\n",
" self.batch_size = batch_size\n",
" self.drop_last = drop_last\n",
" self.shuffle = shuffle\n",
" self.seed = seed\n",
"\n",
" def size(self) -> int:\n",
" \"\"\"Total number of samples.\"\"\"\n",
" return self.images.shape[0]\n",
"\n",
" def steps_per_epoch(self) -> int:\n",
" \"\"\"Number of batches for one epoch.\"\"\"\n",
" return int(np.floor(self.size() / self.batch_size))\n",
"\n",
" def _get_data_batch(self, index: np.ndarray) -> Tuple[np.ndarray, Dict[str, np.ndarray]]:\n",
" \"\"\"Compute risk set for samples in batch.\"\"\"\n",
" time = self.time[index]\n",
" event = self.event[index]\n",
" images = self.images[index]\n",
"\n",
" labels = {\n",
" \"label_event\": event.astype(np.int32),\n",
" \"label_time\": time.astype(np.float32),\n",
" \"label_riskset\": _make_riskset(time)\n",
" }\n",
" return images, labels\n",
"\n",
" def _iter_data(self) -> Iterable[Tuple[np.ndarray, Dict[str, np.ndarray]]]:\n",
" \"\"\"Generator that yields one batch at a time.\"\"\"\n",
" index = np.arange(self.size())\n",
" rnd = np.random.RandomState(self.seed)\n",
"\n",
" if self.shuffle:\n",
" rnd.shuffle(index)\n",
" for b in range(self.steps_per_epoch()):\n",
" start = b * self.batch_size\n",
" idx = index[start:(start + self.batch_size)]\n",
" yield self._get_data_batch(idx)\n",
"\n",
" if not self.drop_last:\n",
" start = self.steps_per_epoch() * self.batch_size\n",
" idx = index[start:]\n",
" yield self._get_data_batch(idx)\n",
"\n",
" def _get_shapes(self) -> Tuple[tf.TensorShape, Dict[str, tf.TensorShape]]:\n",
" \"\"\"Return shapes of data returned by `self._iter_data`.\"\"\"\n",
" batch_size = self.batch_size if self.drop_last else None\n",
" h, w, c = self.images.shape[1:]\n",
" images = tf.TensorShape([batch_size, h, w, c])\n",
"\n",
" labels = {k: tf.TensorShape((batch_size,))\n",
" for k in (\"label_event\", \"label_time\")}\n",
" labels[\"label_riskset\"] = tf.TensorShape((batch_size, batch_size))\n",
" return images, labels\n",
"\n",
" def _get_dtypes(self) -> Tuple[tf.DType, Dict[str, tf.DType]]:\n",
" \"\"\"Return dtypes of data returned by `self._iter_data`.\"\"\"\n",
" labels = {\"label_event\": tf.int32,\n",
" \"label_time\": tf.float32,\n",
" \"label_riskset\": tf.bool}\n",
" return tf.float32, labels\n",
"\n",
" def _make_dataset(self) -> tf.data.Dataset:\n",
" \"\"\"Create dataset from generator.\"\"\"\n",
" ds = tf.data.Dataset.from_generator(\n",
" self._iter_data,\n",
" self._get_dtypes(),\n",
" self._get_shapes()\n",
" )\n",
" return ds\n",
"\n",
" def __call__(self) -> tf.data.Dataset:\n",
" return self._make_dataset()\n",
"\n",
"\n",
"def safe_normalize(x: tf.Tensor) -> tf.Tensor:\n",
" \"\"\"Normalize risk scores to avoid exp underflowing.\n",
"\n",
" Note that only risk scores relative to each other matter.\n",
" If minimum risk score is negative, we shift scores so minimum\n",
" is at zero.\n",
" \"\"\"\n",
" x_min = tf.reduce_min(x, axis=0)\n",
" c = tf.zeros_like(x_min)\n",
" norm = tf.where(x_min < 0, -x_min, c)\n",
" return x + norm\n",
"\n",
"\n",
"def logsumexp_masked(risk_scores: tf.Tensor,\n",
" mask: tf.Tensor,\n",
" axis: int = 0,\n",
" keepdims: Optional[bool] = None) -> tf.Tensor:\n",
" \"\"\"Compute logsumexp across `axis` for entries where `mask` is true.\"\"\"\n",
" risk_scores.shape.assert_same_rank(mask.shape)\n",
"\n",
" with tf.name_scope(\"logsumexp_masked\"):\n",
" mask_f = tf.cast(mask, risk_scores.dtype)\n",
" risk_scores_masked = tf.math.multiply(risk_scores, mask_f)\n",
" # for numerical stability, substract the maximum value\n",
" # before taking the exponential\n",
" amax = tf.reduce_max(risk_scores_masked, axis=axis, keepdims=True)\n",
" risk_scores_shift = risk_scores_masked - amax\n",
"\n",
" exp_masked = tf.math.multiply(tf.exp(risk_scores_shift), mask_f)\n",
" exp_sum = tf.reduce_sum(exp_masked, axis=axis, keepdims=True)\n",
" output = amax + tf.math.log(exp_sum)\n",
" if not keepdims:\n",
" output = tf.squeeze(output, axis=axis)\n",
" return output\n",
"\n",
"\n",
"class CoxPHLoss(tf.keras.losses.Loss):\n",
" \"\"\"Negative partial log-likelihood of Cox's proportional hazards model.\"\"\"\n",
"\n",
" def __init__(self, **kwargs):\n",
" super().__init__(**kwargs) \n",
"\n",
" def call(self,\n",
" y_true: Sequence[tf.Tensor],\n",
" y_pred: tf.Tensor) -> tf.Tensor:\n",
" \"\"\"Compute loss.\n",
"\n",
" Parameters\n",
" ----------\n",
" y_true : list|tuple of tf.Tensor\n",
" The first element holds a binary vector where 1\n",
" indicates an event 0 censoring.\n",
" The second element holds the riskset, a\n",
" boolean matrix where the `i`-th row denotes the\n",
" risk set of the `i`-th instance, i.e. the indices `j`\n",
" for which the observer time `y_j >= y_i`.\n",
" Both must be rank 2 tensors.\n",
" y_pred : tf.Tensor\n",
" The predicted outputs. Must be a rank 2 tensor.\n",
"\n",
" Returns\n",
" -------\n",
" loss : tf.Tensor\n",
" Loss for each instance in the batch.\n",
" \"\"\"\n",
" event, riskset = y_true\n",
" predictions = y_pred\n",
"\n",
" pred_shape = predictions.shape\n",
" if pred_shape.ndims != 2:\n",
" raise ValueError(\"Rank mismatch: Rank of predictions (received %s) should \"\n",
" \"be 2.\" % pred_shape.ndims)\n",
"\n",
" if pred_shape[1] is None:\n",
" raise ValueError(\"Last dimension of predictions must be known.\")\n",
"\n",
" if pred_shape[1] != 1:\n",
" raise ValueError(\"Dimension mismatch: Last dimension of predictions \"\n",
" \"(received %s) must be 1.\" % pred_shape[1])\n",
"\n",
" if event.shape.ndims != pred_shape.ndims:\n",
" raise ValueError(\"Rank mismatch: Rank of predictions (received %s) should \"\n",
" \"equal rank of event (received %s)\" % (\n",
" pred_shape.ndims, event.shape.ndims))\n",
"\n",
" if riskset.shape.ndims != 2:\n",
" raise ValueError(\"Rank mismatch: Rank of riskset (received %s) should \"\n",
" \"be 2.\" % riskset.shape.ndims)\n",
"\n",
" event = tf.cast(event, predictions.dtype)\n",
" predictions = safe_normalize(predictions)\n",
"\n",
" with tf.name_scope(\"assertions\"):\n",
" assertions = (\n",
" tf.debugging.assert_less_equal(event, 1.),\n",
" tf.debugging.assert_greater_equal(event, 0.),\n",
" tf.debugging.assert_type(riskset, tf.bool)\n",
" )\n",
"\n",
" # move batch dimension to the end so predictions get broadcast\n",
" # row-wise when multiplying by riskset\n",
" pred_t = tf.transpose(predictions)\n",
" # compute log of sum over risk set for each row\n",
" rr = logsumexp_masked(pred_t, riskset, axis=1, keepdims=True)\n",
" assert rr.shape.as_list() == predictions.shape.as_list()\n",
"\n",
" losses = tf.math.multiply(event, rr - predictions)\n",
"\n",
" return losses"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "to_jpSRxvEQ9"
},
"source": [
"To monitor the training process, we would like to compute the concordance index with respect to a separate validation set. Similar to the Cox PH loss, the concordance index needs access to predicted risk scores and ground truth of *all* samples in the validation data. While we had to opt for computing the Cox PH loss over a mini-batch, I would not recommend this for the validation data. For small batch sizes and/or high amount of censoring, the estimated concordance index would be quite volatile, which makes it very hard to interpret. In addition, the validation data is usually considerably smaller than the training data, therefore we can collect predictions for the whole validation data and compute the concordance index accurately."
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 231
},
"colab_type": "code",
"executionInfo": {
"elapsed": 1232,
"status": "error",
"timestamp": 1589637629266,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "BUBru4MSvEQ-",
"jupyter": {
"outputs_hidden": true
},
"outputId": "c3ec2a25-6712-4395-c0b4-aebcdcfac4b4"
},
"outputs": [],
"source": [
"class CindexMetric:\n",
" \"\"\"Computes concordance index across one epoch.\"\"\"\n",
"\n",
" def reset_states(self) -> None:\n",
" \"\"\"Clear the buffer of collected values.\"\"\"\n",
" self._data = {\n",
" \"label_time\": [],\n",
" \"label_event\": [],\n",
" \"prediction\": []\n",
" }\n",
"\n",
" def update_state(self, y_true: Dict[str, tf.Tensor], y_pred: tf.Tensor) -> None:\n",
" \"\"\"Collect observed time, event indicator and predictions for a batch.\n",
"\n",
" Parameters\n",
" ----------\n",
" y_true : dict\n",
" Must have two items:\n",
" `label_time`, a tensor containing observed time for one batch,\n",
" and `label_event`, a tensor containing event indicator for one batch.\n",
" y_pred : tf.Tensor\n",
" Tensor containing predicted risk score for one batch.\n",
" \"\"\"\n",
" self._data[\"label_time\"].append(y_true[\"label_time\"].numpy())\n",
" self._data[\"label_event\"].append(y_true[\"label_event\"].numpy())\n",
" self._data[\"prediction\"].append(tf.squeeze(y_pred).numpy())\n",
"\n",
" def result(self) -> Dict[str, float]:\n",
" \"\"\"Computes the concordance index across collected values.\n",
"\n",
" Returns\n",
" ----------\n",
" metrics : dict\n",
" Computed metrics.\n",
" \"\"\"\n",
" data = {}\n",
" for k, v in self._data.items():\n",
" data[k] = np.concatenate(v)\n",
"\n",
" results = concordance_index_censored(\n",
" data[\"label_event\"] == 1,\n",
" data[\"label_time\"],\n",
" data[\"prediction\"])\n",
"\n",
" result_data = {}\n",
" names = (\"cindex\", \"concordant\", \"discordant\", \"tied_risk\")\n",
" for k, v in zip(names, results):\n",
" result_data[k] = v\n",
"\n",
" return result_data"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "Q53UFMrZvERA"
},
"source": [
"## Creating a Convolutional Neural Network for Survival Analysis on MNIST\n",
"\n",
"Finally, after many considerations, we can create a convolutional neural network (CNN) to learn a high-level representation from MNIST digits such that we can estimate each image's survival function. The CNN follows the LeNet architecture where the last linear has one output unit that corresponds to the predicted risk score. The predicted risk score, together with the binary event indicator and risk set, are the input to the Cox PH loss."
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "ZuSGDRpCvERA",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"import tensorflow.compat.v2.summary as summary\n",
"from tensorflow.python.ops import summary_ops_v2\n",
"\n",
"\n",
"class TrainAndEvaluateModel:\n",
"\n",
" def __init__(self, model, model_dir, train_dataset, eval_dataset,\n",
" learning_rate, num_epochs):\n",
" self.num_epochs = num_epochs\n",
" self.model_dir = model_dir\n",
"\n",
" self.model = model\n",
"\n",
" self.train_ds = train_dataset\n",
" self.val_ds = eval_dataset\n",
"\n",
" self.optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)\n",
" self.loss_fn = CoxPHLoss()\n",
"\n",
" self.train_loss_metric = tf.keras.metrics.Mean(name=\"train_loss\")\n",
" self.val_loss_metric = tf.keras.metrics.Mean(name=\"val_loss\")\n",
" self.val_cindex_metric = CindexMetric()\n",
"\n",
" @tf.function\n",
" def train_one_step(self, x, y_event, y_riskset):\n",
" y_event = tf.expand_dims(y_event, axis=1)\n",
" with tf.GradientTape() as tape:\n",
" logits = self.model(x, training=True)\n",
"\n",
" train_loss = self.loss_fn(y_true=[y_event, y_riskset], y_pred=logits)\n",
"\n",
" with tf.name_scope(\"gradients\"):\n",
" grads = tape.gradient(train_loss, self.model.trainable_weights)\n",
" self.optimizer.apply_gradients(zip(grads, self.model.trainable_weights))\n",
" return train_loss, logits\n",
"\n",
" def train_and_evaluate(self):\n",
" ckpt = tf.train.Checkpoint(\n",
" step=tf.Variable(0, dtype=tf.int64),\n",
" optimizer=self.optimizer,\n",
" model=self.model)\n",
" ckpt_manager = tf.train.CheckpointManager(\n",
" ckpt, str(self.model_dir), max_to_keep=2)\n",
"\n",
" if ckpt_manager.latest_checkpoint:\n",
" ckpt.restore(ckpt_manager.latest_checkpoint)\n",
" print(f\"Latest checkpoint restored from {ckpt_manager.latest_checkpoint}.\")\n",
"\n",
" train_summary_writer = summary.create_file_writer(\n",
" str(self.model_dir / \"train\"))\n",
" val_summary_writer = summary.create_file_writer(\n",
" str(self.model_dir / \"valid\"))\n",
"\n",
" for epoch in range(self.num_epochs):\n",
" with train_summary_writer.as_default():\n",
" self.train_one_epoch(ckpt.step)\n",
"\n",
" # Run a validation loop at the end of each epoch.\n",
" with val_summary_writer.as_default():\n",
" self.evaluate(ckpt.step)\n",
"\n",
" save_path = ckpt_manager.save()\n",
" print(f\"Saved checkpoint for step {ckpt.step.numpy()}: {save_path}\")\n",
"\n",
" def train_one_epoch(self, step_counter):\n",
" for x, y in self.train_ds:\n",
" train_loss, logits = self.train_one_step(\n",
" x, y[\"label_event\"], y[\"label_riskset\"])\n",
"\n",
" step = int(step_counter)\n",
" if step == 0:\n",
" # see https://stackoverflow.com/questions/58843269/display-graph-using-tensorflow-v2-0-in-tensorboard\n",
" func = self.train_one_step.get_concrete_function(\n",
" x, y[\"label_event\"], y[\"label_riskset\"])\n",
" summary_ops_v2.graph(func.graph, step=0)\n",
"\n",
" # Update training metric.\n",
" self.train_loss_metric.update_state(train_loss)\n",
"\n",
" # Log every 200 batches.\n",
" if step % 200 == 0:\n",
" # Display metrics\n",
" mean_loss = self.train_loss_metric.result()\n",
" print(f\"step {step}: mean loss = {mean_loss:.4f}\")\n",
" # save summaries\n",
" summary.scalar(\"loss\", mean_loss, step=step_counter)\n",
" # Reset training metrics\n",
" self.train_loss_metric.reset_states()\n",
"\n",
" step_counter.assign_add(1)\n",
"\n",
" @tf.function\n",
" def evaluate_one_step(self, x, y_event, y_riskset):\n",
" y_event = tf.expand_dims(y_event, axis=1)\n",
" val_logits = self.model(x, training=False)\n",
" val_loss = self.loss_fn(y_true=[y_event, y_riskset], y_pred=val_logits)\n",
" return val_loss, val_logits\n",
"\n",
" def evaluate(self, step_counter):\n",
" self.val_cindex_metric.reset_states()\n",
" \n",
" for x_val, y_val in self.val_ds:\n",
" val_loss, val_logits = self.evaluate_one_step(\n",
" x_val, y_val[\"label_event\"], y_val[\"label_riskset\"])\n",
"\n",
" # Update val metrics\n",
" self.val_loss_metric.update_state(val_loss)\n",
" self.val_cindex_metric.update_state(y_val, val_logits)\n",
"\n",
" val_loss = self.val_loss_metric.result()\n",
" summary.scalar(\"loss\",\n",
" val_loss,\n",
" step=step_counter)\n",
" self.val_loss_metric.reset_states()\n",
" \n",
" val_cindex = self.val_cindex_metric.result()\n",
" for key, value in val_cindex.items():\n",
" summary.scalar(key, value, step=step_counter)\n",
"\n",
" print(f\"Validation: loss = {val_loss:.4f}, cindex = {val_cindex['cindex']:.4f}\")"
]
},
{
"cell_type": "code",
"execution_count": 11,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "hG2EK2s5vERD",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"model = tf.keras.Sequential([\n",
" tf.keras.layers.Conv2D(6, kernel_size=(5, 5), activation='relu', name='conv_1'),\n",
" tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),\n",
" tf.keras.layers.Conv2D(16, (5, 5), activation='relu', name='conv_2'),\n",
" tf.keras.layers.MaxPooling2D(pool_size=(2, 2)),\n",
" tf.keras.layers.Flatten(),\n",
" tf.keras.layers.Dense(120, activation='relu', name='dense_1'),\n",
" tf.keras.layers.Dense(84, activation='relu', name='dense_2'),\n",
" tf.keras.layers.Dense(1, activation='linear', name='dense_3')\n",
"])"
]
},
{
"cell_type": "code",
"execution_count": 12,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "YQMI1JuLvERI",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"train_fn = InputFunction(x_train, time_train, event_train,\n",
" drop_last=True,\n",
" shuffle=True)\n",
"\n",
"eval_fn = InputFunction(x_test, time_test, event_test)\n",
"\n",
"trainer = TrainAndEvaluateModel(\n",
" model=model,\n",
" model_dir=Path(\"ckpts-mnist-cnn\"),\n",
" train_dataset=train_fn(),\n",
" eval_dataset=eval_fn(),\n",
" learning_rate=0.0001,\n",
" num_epochs=15,\n",
")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "M5Cwhd8gvERO"
},
"source": [
"To obverse training, we can start TensorBoard."
]
},
{
"cell_type": "code",
"execution_count": 13,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "LiVyl83uvERP",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"# Load the TensorBoard notebook extension.\n",
"%load_ext tensorboard"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 821
},
"colab_type": "code",
"executionInfo": {
"elapsed": 4600,
"status": "ok",
"timestamp": 1589637897978,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "yB9hqXnSvERR",
"jupyter": {
"outputs_hidden": true
},
"outputId": "23ae239d-76b7-4a25-9ccd-ca73930d3ff8"
},
"outputs": [],
"source": [
"%tensorboard --logdir ckpts-mnist-cnn"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "Wi1cPZKkvERT"
},
"source": [
"Let the training begin…"
]
},
{
"cell_type": "code",
"execution_count": 14,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/",
"height": 1000
},
"colab_type": "code",
"executionInfo": {
"elapsed": 92306,
"status": "ok",
"timestamp": 1589637998781,
"user": {
"displayName": "Sebastian Pölsterl",
"photoUrl": "https://lh3.googleusercontent.com/a-/AOh14GibzrfdaHThaPgjaoGC9Dfb7YXvuTd-tFLbzoO2Gb6WEwyKUsHIqQpwFQAnUAKIewfdDQm7LzvGMH1MzU0PGgU9JwdQ2_9F-5kiQH_DlB1ZaFKpkST5Oha3_n4379GpkI6TgsLF0WZU_7qikJ61kKM2ytdtJeEz5VVwoz3XdhEPaqbu57hGpX4JZ2aGKRbmVu9JQOU9u8Ym0_w4HOaywrK2s5F1H700i1y89hljff2afH6WLPCP2XSIW2-eK7Mkk1rWCYHvdKt2Q1F2cjNOVoPO3C_LDkAfl1U33HWfwTKRKrlf_fsw5BrBVeV65FDP2xxtFj47t1uNTni3fq9DSzMb30dX4v0k0zjKVI_PtxFOmm0VAhr1NYrNh5PgBfbgxjcCooOJbNg21wsosLvYazfQdbLZfeCNq79hK6ljJblvcDUdu9l8oV5WftCmYipe-pWi5_hd3RSeiJoHg1bRQctViY6KvOx8taENqNS6P3IY1zYVTlNYgews5dtAVR11ei3ofgB5vcBa-bfqgal4ZlJNcsCSwNzUaKMiQ3twG19ESCSnbgJTbLEb6hHeCyhGKoyRwFjCgvEixoU04BnxGH5SEh_qiXf4euMiEaALYK7SrH35KWoZTkW9wXShGv3CmgCdqyOloiG3QsusKVmB9PPCuLjw0A9ixzd3ktRotErkEH2N1_EAdQqti9CK9A3yirLJSyk7Vs6Uem3Jv1Jr21mHsFocw53FciKfwUXm-LydQGUQ9TvgiZepRPHJCypj3l-6Dg=s64",
"userId": "18353690321324822306"
},
"user_tz": -120
},
"id": "SgS0SawuvERK",
"jupyter": {
"outputs_hidden": true
},
"outputId": "dc051681-3d59-4749-fa8f-4ed44393fddb"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"step 0: mean loss = 2.3711\n",
"step 200: mean loss = 2.0204\n",
"step 400: mean loss = 1.9728\n",
"step 600: mean loss = 1.9706\n",
"step 800: mean loss = 1.9514\n",
"Validation: loss = 1.9304, cindex = 0.6784\n",
"step 1000: mean loss = 1.9386\n",
"step 1200: mean loss = 1.9526\n",
"step 1400: mean loss = 1.9553\n",
"step 1600: mean loss = 1.9484\n",
"step 1800: mean loss = 1.9391\n",
"Validation: loss = 1.9233, cindex = 0.6831\n",
"step 2000: mean loss = 1.9381\n",
"step 2200: mean loss = 1.9446\n",
"step 2400: mean loss = 1.9407\n",
"step 2600: mean loss = 1.9428\n",
"step 2800: mean loss = 1.9339\n",
"Validation: loss = 1.9204, cindex = 0.6851\n",
"step 3000: mean loss = 1.9415\n",
"step 3200: mean loss = 1.9331\n",
"step 3400: mean loss = 1.9472\n",
"step 3600: mean loss = 1.9308\n",
"Validation: loss = 1.9186, cindex = 0.6862\n",
"step 3800: mean loss = 1.9243\n",
"step 4000: mean loss = 1.9305\n",
"step 4200: mean loss = 1.9552\n",
"step 4400: mean loss = 1.9306\n",
"step 4600: mean loss = 1.9286\n",
"Validation: loss = 1.9167, cindex = 0.6877\n",
"step 4800: mean loss = 1.9209\n",
"step 5000: mean loss = 1.9323\n",
"step 5200: mean loss = 1.9410\n",
"step 5400: mean loss = 1.9314\n",
"step 5600: mean loss = 1.9245\n",
"Validation: loss = 1.9152, cindex = 0.6886\n",
"step 5800: mean loss = 1.9296\n",
"step 6000: mean loss = 1.9273\n",
"step 6200: mean loss = 1.9388\n",
"step 6400: mean loss = 1.9274\n",
"Validation: loss = 1.9149, cindex = 0.6888\n",
"step 6600: mean loss = 1.9178\n",
"step 6800: mean loss = 1.9243\n",
"step 7000: mean loss = 1.9421\n",
"step 7200: mean loss = 1.9323\n",
"step 7400: mean loss = 1.9226\n",
"Validation: loss = 1.9137, cindex = 0.6898\n",
"step 7600: mean loss = 1.9103\n",
"step 7800: mean loss = 1.9368\n",
"step 8000: mean loss = 1.9336\n",
"step 8200: mean loss = 1.9334\n",
"step 8400: mean loss = 1.9064\n",
"Validation: loss = 1.9138, cindex = 0.6894\n",
"step 8600: mean loss = 1.9265\n",
"step 8800: mean loss = 1.9184\n",
"step 9000: mean loss = 1.9380\n",
"step 9200: mean loss = 1.9223\n",
"Validation: loss = 1.9133, cindex = 0.6900\n",
"step 9400: mean loss = 1.9153\n",
"step 9600: mean loss = 1.9147\n",
"step 9800: mean loss = 1.9377\n",
"step 10000: mean loss = 1.9291\n",
"step 10200: mean loss = 1.9240\n",
"Validation: loss = 1.9133, cindex = 0.6900\n",
"step 10400: mean loss = 1.9003\n",
"step 10600: mean loss = 1.9353\n",
"step 10800: mean loss = 1.9215\n",
"step 11000: mean loss = 1.9303\n",
"step 11200: mean loss = 1.9078\n",
"Validation: loss = 1.9129, cindex = 0.6904\n",
"step 11400: mean loss = 1.9216\n",
"step 11600: mean loss = 1.9189\n",
"step 11800: mean loss = 1.9335\n",
"step 12000: mean loss = 1.9158\n",
"Validation: loss = 1.9127, cindex = 0.6905\n",
"step 12200: mean loss = 1.9108\n",
"step 12400: mean loss = 1.9192\n",
"step 12600: mean loss = 1.9245\n",
"step 12800: mean loss = 1.9312\n",
"step 13000: mean loss = 1.9165\n",
"Validation: loss = 1.9127, cindex = 0.6906\n",
"step 13200: mean loss = 1.9004\n",
"step 13400: mean loss = 1.9269\n",
"step 13600: mean loss = 1.9200\n",
"step 13800: mean loss = 1.9296\n",
"step 14000: mean loss = 1.9080\n",
"Validation: loss = 1.9125, cindex = 0.6908\n",
"Saved checkpoint for step 14055: ckpts-mnist-cnn/ckpt-1\n"
]
}
],
"source": [
"trainer.train_and_evaluate()"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "HMdg3qsKvERV"
},
"source": [
"We can make a couple of observations:\n",
"\n",
"1. The final concordance index on the validation data is close to the optimal value we computed above using the actual underlying risk scores.\n",
"2. The loss during training is quite volatile, which stems from the small batch size (64) and the varying number of uncensored samples that contribute to the loss in each batch. Increasing the batch size should yield smoother loss curves."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "33q_ISFevERW"
},
"source": [
"### Predicting Survival Functions\n",
"\n",
"For inference, things are much easier, we just pass a batch of images and record the predicted risk score. To estimate individual survival functions, we need to estimate the baseline hazard function $h_0$, which can be done analogous to the linear Cox PH model by using [Breslow's estimator](https://www.jstor.org/stable/1402659)."
]
},
{
"cell_type": "code",
"execution_count": 15,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "cI8ESLZcvERW",
"jupyter": {
"outputs_hidden": true
}
},
"outputs": [],
"source": [
"from sklearn.model_selection import train_test_split\n",
"from sksurv.linear_model.coxph import BreslowEstimator\n",
"\n",
"\n",
"class Predictor:\n",
"\n",
" def __init__(self, model, model_dir):\n",
" self.model = model\n",
" self.model_dir = model_dir\n",
"\n",
" def predict(self, dataset):\n",
" ckpt = tf.train.Checkpoint(\n",
" step=tf.Variable(0, dtype=tf.int64),\n",
" optimizer=tf.keras.optimizers.Adam(),\n",
" model=self.model)\n",
" ckpt_manager = tf.train.CheckpointManager(\n",
" ckpt, str(self.model_dir), max_to_keep=2)\n",
"\n",
" if ckpt_manager.latest_checkpoint:\n",
" ckpt.restore(ckpt_manager.latest_checkpoint).expect_partial()\n",
" print(f\"Latest checkpoint restored from {ckpt_manager.latest_checkpoint}.\")\n",
"\n",
" risk_scores = []\n",
" for batch in dataset:\n",
" pred = self.model(batch, training=False)\n",
" risk_scores.append(pred.numpy())\n",
"\n",
" return np.row_stack(risk_scores)"
]
},
{
"cell_type": "code",
"execution_count": 16,
"metadata": {
"scrolled": false
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Latest checkpoint restored from ckpts-mnist-cnn/ckpt-1.\n"
]
}
],
"source": [
"train_pred_fn = tf.data.Dataset.from_tensor_slices(x_train[..., np.newaxis]).batch(64)\n",
"\n",
"predictor = Predictor(model, trainer.model_dir)\n",
"train_predictions = predictor.predict(train_pred_fn)\n",
"\n",
"breslow = BreslowEstimator().fit(train_predictions, event_train, time_train)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "MkhKNlgnvERZ"
},
"source": [
"Once fitted, we can use Breslow's estimator to obtain estimated survival functions for images in the test data. We randomly draw three sample images for each digit and plot their predicted survival function."
]
},
{
"cell_type": "code",
"execution_count": 17,
"metadata": {
"colab": {},
"colab_type": "code",
"id": "HSdfNEJMvERa",
"jupyter": {
"outputs_hidden": true
},
"outputId": "16bf44f0-427b-4dd1-8a91-c22883e5e5c7"
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Latest checkpoint restored from ckpts-mnist-cnn/ckpt-1.\n"
]
},
{
"data": {
"image/png": "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\n",
"text/plain": [
""
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"sample = train_test_split(x_test, y_test, event_test, time_test,\n",
" test_size=30, stratify=y_test, random_state=89)\n",
"x_sample, y_sample, event_sample, time_sample = sample[1::2]\n",
"\n",
"sample_pred_ds = tf.data.Dataset.from_tensor_slices(\n",
" x_sample[..., np.newaxis]).batch(64)\n",
"sample_predictions = predictor.predict(sample_pred_ds)\n",
"\n",
"sample_surv_fn = breslow.get_survival_function(sample_predictions)\n",
"\n",
"plt.figure(figsize=(6, 4.5))\n",
"for surv_fn, class_label in zip(sample_surv_fn, y_sample):\n",
" risk_group = risk_score_assignment.loc[class_label, \"risk_group\"]\n",
" plt.step(surv_fn.x, surv_fn.y, where=\"post\",\n",
" color=f\"C{class_label}\", linestyle=styles[risk_group])\n",
"\n",
"plt.ylim(0, 1)\n",
"plt.ylabel(\"Probability of survival $P(T > t)$\")\n",
"plt.xlabel(\"Time $t$\")\n",
"plt.grid()"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "1Wxzu2NGvERc"
},
"source": [
"Solid lines correspond to images that belong to risk group 0 (with lowest risk), which the model was able to learn. Samples from the group with the highest risk are shown as dotted lines. Their predicted survival functions have the steepest descent, confirming that the model correctly identified different risk groups from images."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text",
"id": "dKl8PQotvERc"
},
"source": [
"## Conclusion\n",
"\n",
"We successfully built, trained, and evaluated a convolutional neural network for survival analysis on MNIST. While MNIST is obviously not a clinical dataset, the exact same approach can be used for clinical data. For instance, [Mobadersany et al.](https://www.pnas.org/content/115/13/E2970) used the same approach to predict overall survival of patients diagnosed with brain tumors from microscopic images, and [Zhu et al.](https://scholar.google.com/scholar?cluster=3381426605939025516) applied CNNs to predict survival of lung cancer patients from pathological images."
]
}
],
"metadata": {
"accelerator": "GPU",
"colab": {
"collapsed_sections": [],
"name": "tutorial_tf2.ipynb",
"provenance": []
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.7"
}
},
"nbformat": 4,
"nbformat_minor": 1
}