Nurse Activity Recognition based on Temporal Frequency Features

Recognition of nurse care activities is a subset of recognizing human activities. Using the accelerometers embedded into smartphones, raw activity data is captured and then analyzed to recognize different activities. It is difficult to identify this kind of infrequent and unpredictable movement based on accelerometer data alone and traditional imbalanced learning does not bring out the best outcome. Due to the inconsistency of the accelerometer data provided, only the frequency analysis of care record data was the focus of our study. After the data were pre-processed in a basic manner and timestamped, we extracted 5 time based features and tried to predict the future activity based on temporal frequency. The care record dataset provided is strikingly similar to data from weather forecasts. So, using time-based features, our study focuses on roughly correlating weather prediction and nurse care activity prediction upto some extent. We tried several classifiers including Random Forest(RF), Extra Trees(EXT) and KNN to model the data. However, only RF brought out reasonably satisfactory results. Therefore, we based our model on RF which resulted in Precision, Recall, F1 score of 52.98%, 50.13% and 50.21% respectively during training. Test Result of the team Hippocrates: Accuracy: 85% F1-Score: 4.2%