Identify Patients and Health Care Workers Using Transfer Learning on a Pre-trained Convolution Neural Network
Aim
There has been extensive research in leveraging on Deep Learning to train Convolution Neural Networks (CNN) that can predict the presence of objects, humans and scenes in images. However, the results from these research cannot be applied directly to healthcare problems such as monitoring of falls, daily activities, hand hygiene compliance, and be of practical use. This project propose to train a CNN with the ability to foremost detect and predict human bodies as doctors, nurses or patients, as the groundwork for building future healthcare monitoring applications. The model shall be trained using transfer learning on Resnet18.
Conclusion and Future Work
The model achieved a 100% accuracy on the test set and highest overall accuracy of 93.75% on the validation set. The predictive performance of the classes for the validation set are as follows:
Doctor Acc:90.48%
Nurse Acc:91.67%%
Patient Acc:97.44%
​
