Medical event prediction using a personalized dual-channel combiner network

1 . A computer implemented method for predicting an occurrence of a medical event for a patient using a trained neural network, comprising:

preprocessing received historical patient data for a plurality of patients to generate a plurality of normalized training samples;

pretraining a personalized deep convolutional neural network (P-DCCN) by updating of model parameters using a multi-layer perceptron (MLP) prediction layer that combines outputs of a static channel that generates compact static representations of the normalized training samples and a temporal channel of the P-DCCN that combines compact temporal representations of the normalized training samples based on attention weights of the compact temporal representations for each time step of the normalized training samples;

storing the pretrained model in a remote server for utilization for personalization by a local machine during a preparation time period for a medical treatment;

generating a normalized finetuning set as output from the P-DCCN by processing input personal data for the patient from the local machine;

iteratively finetuning the model parameters of the P-DCCN by performing a plurality of pretraining iterations for the P-DCCN sent to the local machine using the generated normalized finetuning set to increase prediction accuracy of the P-DCCN with data limited to personalized patient data obtained from the local machine;

generating a personalized prediction score for future medical events for the patient using the P-DCCN; and

controlling an operation of a medical treatment device responsive to the personalized prediction score for future medical events.

2 . The method of claim 1 , wherein the P-DCCN is finetuned by optimizing using a regression loss function as follows:

l

=

1

N

⁢

∑

i

=

1

N

⁢



y

ˆ

i

-

y

i



2

2

+

λ

⁢



θ



2

2

where y i is a true indicator of an incidence of an event for an i-th sample in the training samples, ŷ i is a predicted score for the i-th sample, N is a total number of the training samples, θ represents the model parameters, λ is a hyperparameter which controls a regularization on the model parameters to avoid overfitting during the training.

3 . The method of claim 1 , wherein a static channel is utilized for processing static and first temporal features, and a temporal channel is utilized for processing second temporal features having higher frequency compared to the first temporal features.

4 . The method of claim 3 , wherein the static channel includes a multilayer perceptron (MLP) configured to encode information in static features x s to a compact representation of the static features h s by:

h

s

=

f

MLP

(

x

s

)

where f MLP (·) represents multiple layers of a fully connected network with the form W s x s +b s , with W s and b s being model parameters to be trained.

5 . The method of claim 3 , wherein the temporal channel includes a plurality of long short term memory (LSTM) layers for processing the temporal features, represented by a sequence of vectors x 1 , . . . , x T , to output a sequence of compact representations h 1 , . . . , h T by:

h

1

,

…

,

h

T

=

f

LSTM

(

x

1

,

…

,

x

T

)

where f LSTM (·) includes multiple layers of LSTM units, which include trainable model parameters.

6 . The method of claim 1 , further comprising computing a probability of an incidence of a medical event by concatenating static and temporal representations h s and h d , received from a static channel and a temporal channel, respectively, and computing the probability using a multilayer perceptron (MLP) by:

y

^

=

f

MLP

(

[

h

s

,

h

d

]

)

where ŷ is a score which indicates the probability of the incidence of the medical event.

7 . The method of claim 1 , further comprising generating, during the preparation time period, a recommendation for decision making by a medical professional, the recommendation identifying one or more treatment parameter adjustments based on the personalized prediction score.

8 . A system for predicting an occurrence of a medical event for a patient using a trained neural network, comprising:

a processor operatively coupled to a computer-readable storage medium, the processor being configured for:

preprocessing received historical patient data for a plurality of patients to generate a plurality of normalized training samples;

pretraining a personalized deep convolutional neural network (P-DCCN) by updating of model parameters using a multi-layer perceptron (MLP) prediction layer that combines outputs of a static channel that generates compact static representations of the normalized training samples and a temporal channel of the P-DCCN that combines compact temporal representations of the normalized training samples based on attention weights of the compact temporal representations for each time step of the normalized training samples;

storing the pretrained model in a remote server for utilization for personalization by a local machine during a preparation time period for a medical treatment;

generating a normalized finetuning set as output from the P-DCCN by processing input personal data for the patient from the local machine;

iteratively finetuning the model parameters of the P-DCCN by performing a plurality of pretraining iterations for the P-DCCN sent to the local machine using the generated normalized finetuning set to increase prediction accuracy of the P-DCCN with data limited to personalized patient data obtained from the local machine;

generating a personalized prediction score for future medical events for the patient using the P-DCCN; and

controlling an operation of a medical treatment device responsive to the personalized prediction score for future medical events.

9 . The system of claim 8 , wherein the P-DCCN is finetuned by optimizing using a regression loss function as follows:

l

=

1

N

⁢

∑

i

=

1

N

⁢



y

ˆ

i

-

y

i



2

2

+

λ

⁢



θ



2

2

where y i is a true indicator of an incidence of an event for an i-th sample in the training samples, ŷ i is a predicted score for the i-th sample, N is a total number of the training samples, θ represents the model parameters, λ is a hyperparameter which controls a regularization on the model parameters to avoid overfitting during the training.

10 . The system of claim 8 , wherein the preprocessing received historical patient data for a plurality of patients further comprises segmenting recording data and generating time series samples.

11 . The system of claim 8 , wherein a static channel is utilized for processing static and first temporal features, and a temporal channel is utilized for processing second temporal features having higher frequency compared to the first temporal features.

12 . The system of claim 11 , wherein the static channel includes a multilayer perceptron (MLP) configured to encode information in static features x s to a compact representation of the static features h s by:

h

s

=

f

MLP

(

x

s

)

where f MLP (·) represents multiple layers of a fully connected network with the form W s x s +b s , with W s and b s being model parameters to be trained.

13 . The system of claim 11 , wherein the temporal channel includes a plurality of long short term memory (LSTM) layers for processing the temporal features, represented by a sequence of vectors x 1 , . . . , x T , to output a sequence of compact representations h 1 , . . . , h T by:

h

1

,

…

,

h

T

=

f

LSTM

(

x

1

,

…

,

x

T

)

where f LSTM (·) includes multiple layers of LSTM units, which include trainable model parameters.

14 . The system of claim 8 , wherein the processor is further configured for computing a probability of an incidence of a medical event by concatenating static and temporal representations h s and h d , received from a static channel and a temporal channel, respectively, and computing the probability using a multilayer perceptron (MLP) by:

y

^

=

f

MLP

(

[

h

s

,

h

d

]

)

where ŷ is a score which indicates the probability of the incidence of the medical event.

15 . A non-transitory computer-readable storage medium comprising a computer-readable program for predicting an occurrence of a medical event for a patient using a trained neural network, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:

preprocessing received historical patient data for a plurality of patients to generate a plurality of normalized training samples;

pretraining a personalized deep convolutional neural network (P-DCCN) by updating of model parameters using a multi-layer perceptron (MLP) prediction layer that combines outputs of a static channel that generates compact static representations of the normalized training samples and a temporal channel of the P-DCCN that combines compact temporal representations of the normalized training samples based on attention weights of the compact temporal representations for each time step of the normalized training samples;

storing the pretrained model in a remote server for utilization for personalization by a local machine during a preparation time period for a medical treatment;

generating a normalized finetuning set as output from the P-DCCN by processing input personal data for the patient from the local machine;

iteratively finetuning the model parameters of the P-DCCN by performing a plurality of pretraining iterations for the P-DCCN sent to the local machine using the generated normalized finetuning set to increase prediction accuracy of the P-DCCN with data limited to personalized patient data obtained from the local machine;

generating a personalized prediction score for future medical events for the patient using the P-DCCN; and

controlling an operation of a medical treatment device responsive to the personalized prediction score for future medical events.

16 . The computer-readable storage medium of claim 15 , wherein the P-DCCN is finetuned by optimizing using a regression loss function as follows:

l

=

1

N

⁢

∑

i

=

1

N

⁢



y

ˆ

i

-

y

i



2

2

+

λ

⁢



θ



2

2

where y i is a true indicator of an incidence of an event for an i-th sample in the training samples, ŷ i is a predicted score for the i-th sample, N is a total number of the training samples, θ represents the model parameters, λ is a hyperparameter which controls a regularization on the model parameters to avoid overfitting during the training.

17 . The computer-readable storage medium of claim 15 , wherein a static channel is utilized for processing static and first temporal features, and a temporal channel is utilized for processing second temporal features having higher frequency compared to the first temporal features.

18 . The computer-readable storage medium of claim 17 , wherein the static channel includes a multilayer perceptron (MLP) configured to encode information in static features x s to a compact representation of the static features h s by:

h

s

=

f

MLP

(

x

s

)

where f MLP (·) represents multiple layers of a fully connected network with the form W s x s +b s , with W s and b s being model parameters to be trained.

19 . The computer-readable storage medium of claim 17 , wherein the temporal channel includes a plurality of long short term memory (LSTM) layers for processing the temporal features, represented by a sequence of vectors x 1 , . . . , x T , to output a sequence of compact representations h 1 , . . . , h T by:

h

1

,

…

,

h

T

=

f

LSTM

(

x

1

,

…

,

x

T

)

where f LSTM (·) includes multiple layers of LSTM units, which include trainable model parameters.

20 . The computer-readable storage medium of claim 19 , further comprising computing a probability of an incidence of a medical event by concatenating static and temporal representations h s and h d , received from a static channel and a temporal channel, respectively, and computing the probability using a multilayer perceptron (MLP) by:

y

^

=

f

MLP

(

[

h

s

,

h

d

]

)

where ŷ is a score which indicates the probability of the incidence of the medical event.