TDA-Aware Long Short-Term Memory Model Improves Accuracy of Human Activity Recognition
Rowan J Barker-Clarke, Nicholas C. Latina, Daniel Suh, Harshita Kumar, Jacob Gardinier Scott, Andrew Dhawan · 2025
Reliable recognition of human activity from time-series actigraphy data could significantly advance patient care by enabling automated evaluation and monitoring of patient functionality outside the clinic. It remains an open question whether topological features, which describe the shape properties of the sensor time series, can improve the predictive ability of such recognition models. Here, we devise a method to compute topological features from time series data, calculate topological persistence, and test whether this approach improves activity recognition. We evaluate the variance and correlations within this feature set and whether topological features that can represent the ‘shape’ within sensor time series have independent predictive value in activity recognition. We implement a deep learning architecture that combines time-series features extracted using topological data analysis with raw sensor values and passes inputs through long short-term memory units. Our model performance, when trained and evaluated on the WISDM 6-activity and WISDM 18-activity datasets, highlights that the addition of topological features to the CNN+LSTM architecture improves maximum accuracy in both settings. In the WISDM-18 dataset, we achieve an average F1 score of 74.5% across folds (68.1-82.5%) compared to an average score of 70.7% (62.5-77.3%) for the same model without TDA input. Our results are comparable to the highest published accuracies for other methods with an explicit subject-independent training-validation split. We find that constructed topological features improve the classification of activity from time-series data, and provide complementary information to traditional time series datasets. These results motivate further exploration of TDA in additional activity recognition datasets.