Sex adaptive deep recurrent neural networks for Parkinson’s disease detection using 5-second vertical ground reaction force signals

This study introduces an innovative sex-stratified methodology for the identification of Parkinson’s disease (PD) using vertical ground reaction force (VGRF) data obtained from foot sensors during ambulation. We devised and evaluated four distinct recurrent neural network designs. We identified a substantial deficiency in existing diagnostic methodologies that infrequently utilize sex-stratified techniques by creating distinct models for male and female individuals over the age of 50 years and architectural benchmarking for variants of recurrent neural networks in processing vertical ground reaction force signals. We trained Long Short-Term Memory (LSTM), bidirectional LSTM (bi-LSTM), Gated Recurrent Unit (GRU), and bidirectional GRU (bi-GRU) on brief 5 s intervals of vertical ground reaction force (VGRF) data. Our results indicate that the GRU models, which use fewer resources, perform very well, achieving high accuracy (98.5% for males and 99.46% for females), recall (98.91% for males and 100% for females), and F1 scores (98.78% for males and 99.48% for females), compared to the more complex LSTM models. The bidirectional models demonstrated similar performance but necessitated increased processing resources. The efficacy of these sex-specific models with abbreviated time frames underscores the possibility of more tailored, efficient, and accessible Parkinson’s disease screening instruments that could facilitate earlier intervention and treatment. Our discovery signifies a significant step in streamlining the diagnostic procedure for Parkinson’s disease patients non-invasively while also establishing a basis for the future development of real-time diagnostic in clinical devices.