Neural Network-Based Granular Activity Recognition from Accelerometers: Assessing Generalizability Across Diverse Mobility Profiles.

The gated recurrent unit (GRU) architecture achieved the best cross-cohort generalization when trained solely on young cohorts' data and tested on old cohorts' data.

• GRU achieved a weighted F1-score of 0.95 ± 0.05 on young cohorts and 0.93 ± 0.05 on old cohorts.
• Neural networks were trained on young cohorts' data and tested on old cohorts' data to assess cross-cohort generalizability.
• Performance was evaluated using accuracy, recall, precision, F1-score, and confusion matrices.
• Multiple network architectures were compared in the study.

Combining datasets from young and old cohorts significantly improved classification performance on the old cohort.

• Combined dataset training resulted in a weighted F1-score of 0.97 ± 0.02 on the old cohort.
• This improvement was attributed to increased variability in the training data.
• This represents an improvement compared to the 0.93 ± 0.05 achieved when training only on young cohorts' data.

The thigh-mounted sensor consistently achieved higher classification performance than the lower back sensor across most activities.

• The thigh sensor outperformed the lower back sensor across activities except for lying.
• Both sensors sampled at 50 Hz.
• Sensor locations evaluated were the right thigh and lower back.
• Activities were annotated using video recordings from chest-mounted cameras synchronised with the accelerometers.

Classification performance across multiple sampling frequencies was comparable.

• The effect of sampling frequency on HAR performance was one of the investigated factors.
• The original sampling frequency of the accelerometers was 50 Hz.
• No substantial performance degradation was found across the different sampling frequencies tested.

Traditional HAR methods struggle to adapt to variable-length, real-world activity data and to generalise across cohorts.

• Cross-cohort generalization (e.g., from young to old cohorts) was identified as a key challenge in HAR.
• The study used daily-life activities annotated via video recordings synchronised with wearable accelerometers.
• Dataset composition and network architecture were identified as important factors influencing HAR performance.
• The study highlights the potential of neural networks for robust, real-world activity recognition across age-defined cohorts.