11 Healthcare IoT
543
• The average of 10 consequent pre-RR values (local-RR)
• Similar to local-RR but for the last 20 minutes instead of the last 10 values
(Global-RR)
– Neural network: As mentioned above, we utilize a shallow neural network
with a few layers and neuron as our machine learning on chip (IoT devices).
Our rich experiments show that the accuracy of this network for arrhythmia
detection is above 90%.
• Edge/fog intelligence: Utilizing a fog layer allows delay-sensitive applications
(e.g., arrhythmia detection) to make online real-time decisions. The fog enables
us to develop more complex machine learning algorithms offering better accuracy compared to shallow neural networks which we used in endpoint IoT
devices. In this case study, we design and implement a convolutional neural
network that have been proven to be very effective in several areas such as
image classification. To do so, the fog node converts the ECG signal to an
equivalent image and feeds it into a CNN. In this case, we use an AlexNet CNN.
Our experimental results show that CNN is able to significantly improve the
performance of arrhythmia detection. Indeed, AlexNet can detect the anomalies
in ECG signals with the accuracy of 97%. As a result, the patient can switch
between machine learning on chip (endpoint IoT device) and fog node depending
on the working conditions and requirements. Note that the wearable ECG
monitoring device (endpoint IoT device) can be connected to the fog node (e.g.,
smartphone) via Bluetooth. However, transferring raw ECG signals can drain the
battery.
• Big Data analytics in the cloud: Although we moved the decision making task
to endpoint IoT devices and fog nodes, we need to train the machine learning
algorithms (shallow neural network and CNN) in the cloud. The reason is that
training is a huge computational burden for such devices. In addition, in the
cloud we have access to more data (ECG signals) from several patients resulting
in better machine learning modes. The cloud enables us to continuously train
and improve the models, and periodically, we can dispatch the model parameters
(e.g., coefficient, weights, etc.) to edge.
11.8 Summary
As the Internet of Things (IoT) paradigm becomes more widespread, a host of novel
opportunities have arisen. Technologies such as miniature wearable biosensors,
along with advances in Big Data, especially with respect to efficient handling of
large, multiscale, multimodal, distributed, and heterogeneous data sets, have opened
the floodgates for e-Health and m-Health services that are more personalized and
precise than ever before. However, IoT hints at an even greater change in healthcare
paradigms; it promises greater accessibility and availability, personalization and
tailored content, and improved returns on investments in delivery. Even so, as IoT
543
• The average of 10 consequent pre-RR values (local-RR)
• Similar to local-RR but for the last 20 minutes instead of the last 10 values
(Global-RR)
– Neural network: As mentioned above, we utilize a shallow neural network
with a few layers and neuron as our machine learning on chip (IoT devices).
Our rich experiments show that the accuracy of this network for arrhythmia
detection is above 90%.
• Edge/fog intelligence: Utilizing a fog layer allows delay-sensitive applications
(e.g., arrhythmia detection) to make online real-time decisions. The fog enables
us to develop more complex machine learning algorithms offering better accuracy compared to shallow neural networks which we used in endpoint IoT
devices. In this case study, we design and implement a convolutional neural
network that have been proven to be very effective in several areas such as
image classification. To do so, the fog node converts the ECG signal to an
equivalent image and feeds it into a CNN. In this case, we use an AlexNet CNN.
Our experimental results show that CNN is able to significantly improve the
performance of arrhythmia detection. Indeed, AlexNet can detect the anomalies
in ECG signals with the accuracy of 97%. As a result, the patient can switch
between machine learning on chip (endpoint IoT device) and fog node depending
on the working conditions and requirements. Note that the wearable ECG
monitoring device (endpoint IoT device) can be connected to the fog node (e.g.,
smartphone) via Bluetooth. However, transferring raw ECG signals can drain the
battery.
• Big Data analytics in the cloud: Although we moved the decision making task
to endpoint IoT devices and fog nodes, we need to train the machine learning
algorithms (shallow neural network and CNN) in the cloud. The reason is that
training is a huge computational burden for such devices. In addition, in the
cloud we have access to more data (ECG signals) from several patients resulting
in better machine learning modes. The cloud enables us to continuously train
and improve the models, and periodically, we can dispatch the model parameters
(e.g., coefficient, weights, etc.) to edge.
11.8 Summary
As the Internet of Things (IoT) paradigm becomes more widespread, a host of novel
opportunities have arisen. Technologies such as miniature wearable biosensors,
along with advances in Big Data, especially with respect to efficient handling of
large, multiscale, multimodal, distributed, and heterogeneous data sets, have opened
the floodgates for e-Health and m-Health services that are more personalized and
precise than ever before. However, IoT hints at an even greater change in healthcare
paradigms; it promises greater accessibility and availability, personalization and
tailored content, and improved returns on investments in delivery. Even so, as IoT