US 2022/0092430 A1 · TIME SERIES DEEP SURVIVAL ANALYSIS SYSTEM IN COMBINATION WITH ACTIVE LEARNING

Abstract

Provided is a time series deep survival analysis system combined with active learning. The system includes: a data collection module, an active learning module, and a time series deep survival analysis module; the data collection module is used for obtaining survival data of objects to be analyzed; combined with an active learning method, the active learning module selects a part of right censored data to label a survival time; and the time series deep survival analysis module constructs a time series deep survival analysis neural network model, and takes uncensored data and right censored data as model inputs, so as to obtain survival time prediction results of the objects to be analyzed. The present application can make full use of the right censored data in the survival data and time series features.

A time series deep survival analysis system combined with active learning, comprising:

a data collection module, an active learning module, and a time series deep survival analysis module;
wherein,
the data collection module is configured to obtain survival data of objects to be analyzed;
combined with an active learning method, the active learning module selects a part of right censored data to label a survival time, which is specifically as follows:
(1) putting all the survival data in a preparatory training set pool, in which time series features, survival times and whether censoring of all the objects to be analyzed are stored;
(2) using a cox risk proportional regression model to perform cox regression analysis on a preparatory training set, so as to calculate a concordance index (Concordance Index, C-index)C0;
(3) putting all the right censored data in a censored data pool, in which the time series features and censoring times of all the objects to be analyzed are stored;
(4) combined with the active learning method, according to a novel sampling strategy, sorting the data in the censored data pool, wherein the specific steps are as follows:
(4.1) performing, for each data instance Ii in the censored data pool, cox regression analysis on the data instance and all the uncensored data, so as to calculate a concordance index Ci, and meanwhile calculating a change in the concordance index ΔCi=Ci−C0;
and
(4.2) sorting the ΔCi of the data instances in the censored data pool in an order from greatest to least;
and
(5) selecting the most important batch of right censored data ranked first, labeling a survival time of selected right censored data, updating the labeled data into the preparatory training set pool, and recording whether censoring as NOT;
the time series deep survival analysis module constructs a time series deep survival analysis neural network model, and takes the uncensored data and the right censored data as model inputs, so as to obtain survival time prediction results of the objects to be analyzed;
the time series deep survival analysis neural network model comprises a time series data input layer, a bidirectional LSTM recurrent neural network layer, an attention mechanism layer, a fully connected layer and a softmax output layer, which are sequentially connected;
the survival time or the censoring time of each object to be analyzed is divided into k intervals at an interval of m days, so as to obtain an interval tag of each object to be analyzed, and a survival analysis problem is converted into a multi-category tag classification prediction problem;
the survival data, the survival time tag and whether censoring in the preparatory training set are input into the time series data input layer of the network model, and by means of the softmax output layer, each input data instance will output k values:
h0, h1, h2 . . . hk-1, representing the probabilities that an ending event occurs in the 0, 1 . . . (k−1) intervals, respectively;
a loss function Lossall of the time series deep survival analysis neural network model consists of two parts, which are denoted as Loss1 and Loss2, respectively:
Loss
1
=
-
log
⁢
⁢
Π
(
X
,
T
)
∈
D
uncensored
❘
(
X
,
T
)
∈
D
censored
⁢
S
⁡
(
T
-
1
❘
X
)
=
-
⁢
Σ
(
X
,
T
)
∈
D
uncensored
❘
(
X
,
T
)
∈
D
censored
⁢
log
⁡
[
Π
l
:
l
≤
(
T
-
1
)
⁡
(
1
-
h
t
)
]
=
-
Σ
(
X
,
T
)
∈
D
uncensored
❘
(
X
,
T
)
∈
D
censored
⁢
⁢
Σ
l
:
l
≤
(
T
-
1
)
⁢
log
⁡
(
1
-
h
t
)
Loss
2
=
-
log
⁢
⁢
Π
(
X
,
T
)
∈
D
uncensored
⁢
Pr
⁡
(
z
=
T
❘
X
)
=
-
log
⁢
⁢
Π
(
X
,
T
)
∈
D
uncensored
⁡
[
h
T
⁢
Π
l
:
l
≤
T
⁡
(
1
-
h
t
)
]
=
-
Σ
(
X
,
T
)
∈
D
uncensored
[
log
⁢
⁢
h
T
+
Σ
l
:
l
≤
T
⁢
log
⁡
(
1
-
h
t
)
⁢
⁢
⁢
Loss
all
=
Loss
1
+
Loss
2
where, X represents a sum of the time series features of the data, comprising the time series features of s step lengths:
X0, X1, X2 . . . Xs-2, Xs-1;
T represents an interval tag to which the survival time or the censoring time of the data is divided;
Duncensored and Dcensored represent a set of uncensored data and a set of censored data, respectively;
ha represents the probability that the ending event occurs in an ath interval;
z represents the interval tag to which the survival time or the censoring time predicted by the model is divided, and the calculation formula of z is z=argmax([h0, h1, h2 . . . hk-1]);
S(t|X) represents the probability that the ending event still does not occur at the end of the time t where the feature of the data is X, and Pr(z=T|X) represents the probability of z=T where the feature of the data is X.

245 more elements

Time series deep survival analysis system in combination with active learning

Life of this patent

Time series deep survival analysis system in combination with active learning

Life of this patent

Abstract

Claims5