Method and apparatus for automatic event prediction
A method and apparatus for predicting hospital bed exit events from video camera systems is disclosed. The system processes video data with a deep convolutional neural network consisting of five main layers: a 1×1 3D convolutional layer used for generating feature maps from raw video data, a context-aware pooling layer used for rectifying data from different camera angles, two fully connected layers used for applying pre-trained deep features, and an output layer used to provide a likelihood of a bed exit event.
1. An apparatus for predicting that a patient is about to exit from a patient support apparatus, the apparatus comprising
a camera positioned with the patient support apparatus in the field of view of the camera,
a controller receiving signals representative of images from the camera, the controller operable to:
capture time sequenced video images from the camera,
input the time sequenced video images to a convolution layer and convolving the time sequenced video images with a convolution kernel to produce a defined number of feature maps,
input the feature maps into a context-aware pooling layer to extract relevant features of interest from the feature maps and generate feature vectors, wherein the context-aware pooling layer applies a rectification factor to the feature of interest to correct for the differences between the developed kernel and the appropriate kernel to be applied to account for a difference in the relationship of the position and orientation of the bed in the developed kernel and the actual position and orientation of the patient support apparatus,
input a feature vector to a first fully connected layer such that each element of the feature vector is connected to a plurality of artificial neurons in the first fully connected layer and each combination outputs a first connected layer value,
input the values derived by each combination of first fully connected layer into a second fully connected layer such that each value is connected to a plurality of artificial neurons in the second fully connected layer such that each combination outputs a second connected layer value,
input the second connected layer values into an output layer which provides a non-exit probability which defines the likelihood that a patient exit event will not occur in a predetermined time and an exit probability which defines the likelihood that a patient exit event will occur in the predetermined time,
utilize the non-exit probability and exit probability to determine the likelihood of a patient exit event to generate a signal when the determine likelihood of a patient exit event exceeds a threshold value, and
if the signal is generated based on the determined likelihood of a patient exit event exceeds a threshold value, generating a notification of the impending event.
2. The apparatus of claim 1 , wherein the convolution layer applies a rectifier when the feature maps are generated.
3. The apparatus of claim 2 , wherein the rectifier introduces non-saturating linearity to the features maps.
4. The apparatus of claim 3 , wherein the rectifier is an absolute value function.
5. The apparatus of claim 1 , wherein the first fully connected layer includes 50 artificial neurons.
6. The apparatus of claim 5 , wherein the second fully connected layer comprises 10 artificial neurons.
7. The apparatus of claim 6 , wherein first fully connected layer and second fully connected layer apply a transfer function.
8. The apparatus of claim 7 , wherein the transfer function is the tansig(x) function.
9. The apparatus of claim 1 , wherein first fully connected layer and second fully connected layer apply a transfer function.
10. The apparatus of claim 9 , wherein the transfer function is the tansig(x) function.
11. The apparatus of claim 10 , wherein the first fully connected layer and the second fully connected layer are developed by training via stochastic gradient descent to produce a set of deep features.
12. The apparatus of claim 1 , wherein the first fully connected layer and the second fully connected layer are developed by training via stochastic gradient descent to produce a set of deep features.
13. A method of predicting that a patient is about to exit from a patient support apparatus that is in the field of view of a camera, the method comprising
receiving signals representative of images from the camera,
capturing time sequenced video images from the camera,
inputting the time sequenced video images to a convolution layer and convolving the time sequenced video images with a convolution kernel to produce a defined number of feature maps,
inputting the feature maps into a context-aware pooling layer to extract relevant features of interest from the feature maps and generate feature vectors, wherein the context-aware pooling layer applies a rectification factor to the feature vectors to correct for the differences between the developed kernel and the appropriate kernel to be applied to account for a difference in the relationship of the position and orientation of the bed in the developed kernel and the actual position and orientation of the patient support apparatus,
inputting a feature vector to a first fully connected layer such that each element of the feature vector is connected to a plurality of artificial neurons in the first fully connected layer and each combination outputs a first connected layer value,
inputting the values derived by each combination of first fully connected layer into a second fully connected layer such that each value is connected to a plurality of artificial neurons in the second fully connected layer such that each combination outputs a second connected layer value,
inputting the second connected layer values into an output layer which provides a non-exit probability which defines the likelihood that a patient exit event will not occur in a predetermined time and an exit probability which defines the likelihood that a patient exit event will occur in the predetermined time,
utilizing the non-exit probability and exit probability to determine the likelihood of a patient exit event to generate a signal when the determined likelihood of a patient exit event exceeds a threshold value, and
if the signal is generated based on the determined likelihood of a patient exit event exceeds a threshold value, generating a notification of the impending event.
14. The method of claim 13 , wherein the convolution layer applies a rectifier function when the feature maps are generated.
15. The method of claim 14 , wherein the rectifier function introduces non-saturating linearity to the features maps.
16. The method of claim 15 , wherein the rectifier function is an absolute value function.
17. The method of claim 16 , wherein the first fully connected layer includes 50 artificial neurons.
18. The method of claim 17 , wherein the second fully connected layer comprises 10 artificial neurons.
19. The method of claim 13 , wherein first fully connected layer and second fully connected layer apply a transfer function.
20. The method of claim 19 , wherein the transfer function is the tansig(x) function.
21. The method of claim 20 , wherein the first fully connected layer and the second fully connected layer are developed by training via stochastic gradient descent to produce a set of deep features.
22. The method of claim 13 , wherein the first fully connected layer and the second fully connected layer are developed by training via stochastic gradient descent to produce a set of deep features.
23. The method of claim 13 , wherein the first fully connected layer includes 50 artificial neurons.
24. The method of claim 23 , wherein the second fully connected layer comprises 10 artificial neurons.