Real-time Location Systems (RTLS) in a clinical environment are an important way to make the workflows more efficient and increase the safety of the patients. Locating patients and equipment can save a lot of time, and by analyzing the flow of patient’s bottlenecks can be found. If there is a disease outbreak in the hospital the tracking history can be used to streamline the search for patient who could be at risk.
In this study, four different models are implemented and trained to predict the position of a patient using Bluetooth hardware. A combined CNN and ANN architecture, a LSTM and two variants of a TCN architecture are evaluated on four different datasets. The datasets were collected during the study and consist of data from two different environments, an office environment and a clinical environment. In each of these environments two different type of datasets were collected. In the first one, the data was collected from each zone and labeled thereafter and in the second type multiple trajectories including zone switches was collected. The trajectories were used for evaluation of robustness and responsiveness.
All models showed good results (0.99 accuracy) on the validation and test sets when the history length of the input data was longer. The best result regarding robustness and responsiveness was achieved by the CNN+ANN model when a history length of 10 was used, with a degree of robustness of 0.99 and a averaged delay of zone switches by 2.19 seconds. The TCN architecture best performance was achieved with a history length of 12 were it had a degree of robustness of 0.98 and an averaged delay of 2.54 seconds.
It was concluded that the goal to collect two datasets in two different environment was successfully made. The aim to verify the two of the models from the related work in these new settings, using different hardware, was also successfully made. However, it could not be concluded that the proposed TCN models performed better regarding the robustness and responsiveness.