Date of Award
Master of Science (MS)
Adam W Hoover
This thesis investigates the difference between raw and linear acceleration in wrist motion for detecting eating episodes. In previous work, our group developed a classifier that analyzed linear motion and achieved good accuracy. However, the classifier can be volatile in the sense that when retrained and tested on the same data, accuracy varies, especially when trained on small amounts of data such as for a single individual. We hypothesize that in part this may be due to the noise in linear acceleration which is significantly larger relative to normal human wrist motions as compared to the noise in raw acceleration. We therefore perform a set of experiments to determine if classifier accuracy and/or stability can be improved by analyzing raw acceleration instead of linear acceleration.
The dataset used for this work is the Clemson All-Day Eating (CAD) dataset. This was collected over a period of one year, in 2014. In the process of data collection, 351 participants were recruited and 354 days of wrist data was recorded. The recorded data contained 1,133 meals spread over 250 hours of eating. The total length of the recorded data was nearly 4,680 hours. In this work, the CAD dataset was reduced to 342 days and 1034 meals because for some recordings, raw acceleration data was not saved.
Previous work developing a classifier based on linear acceleration achieved a time-based weighted accuracy of 80%, a true positive rate of 89% on eating episodes, and a false positive per true positive rate of 1.7. However, these results were based upon a single run of train and test. Recently we discovered that the model accuracy varies somewhat between runs. We therefore perform a replication experiment on the linear classifier to confirm these results by rerunning the entire experiment 10 times. We report the average and standard deviation of all metrics across these runs. This helps establish a better baseline for comparison of our new classifier that analyzes raw acceleration.
We next analyze the same set of data, using the same neural network model and general approach as for the linear acceleration-based classifier, to compare its accuracy and stability. Evaluating all results, we found that the linear acceleration classifier achieved (average ± standard deviation across 10 runs) a TPR of 86% ±1.2% and a FP/TP of 1.7 ± 0.3. It also achieved a weighted accuracy of 79 % ± 0.5 %. Thus, we concluded that the results of original experiment were above the average results and could either be due to a freak training and testing run or due to contamination of the testing data. These results set up a new baseline with which we compare the raw acceleration model metrics. We found that the raw acceleration achieved a TPR of 84% ± 1.3% and a FP/TP of 1.7 ± 0.3. In the case of time metrics, the raw acceleration model achieved a weighted accuracy of 78% ± 0.4%. Thus, on average, we found that the linear acceleration performed slightly better than raw acceleration in episode detection. The time metrics for both raw and linear acceleration were more or less similar but we did see a higher standard deviation for the raw models.
Our results indicate that linear acceleration does provide greater accuracy than raw acceleration. Even though raw acceleration has a higher signal-to-noise ratio than linear acceleration, in terms of normal human wrist motions, our classifier model has relatively equal volatility when analyzing either signal. We conclude that the main source of model volatility is still unknown. Thus, we found that linear acceleration is, overall, a better predictor of eating as compared to raw acceleration. It should be noted that the difference in the accuracies is very minor and the volatility in the training process could account for some of the differences.
Gupta, Shaurya, "Raw Versus Linear Acceleration in the Recognition of Wrist Motions Related to Eating During Everyday Life" (2021). All Theses. 3632.