Skip to main content
. 2026 May 1;16:20271. doi: 10.1038/s41598-026-50771-0

Table 13.

Ablation study results on test set (5-fold cross-validation).

Configuration Acc. (%) Inline graphic (%)
Full BATL 93.33 ± 0.94 –
Component ablations
 w/o biomechanical constraints 85.55 ± 1.83 Inline graphic
 w/o phase discovery 88.89 ± 1.47 Inline graphic
 w/o dual-stream fusion 84.44 ± 2.11 Inline graphic
Individual constraint ablations
 w/o Inline graphic 90.00 ± 1.56 Inline graphic
 w/o Inline graphic 91.11 ± 1.34 Inline graphic
 w/o Inline graphic 92.22 ± 1.02 Inline graphic
Architecture variants
 Replace BiLSTM with GRU 91.11 ± 1.45 Inline graphic
 Replace BiLSTM with transformer 92.22 ± 1.28 Inline graphic
 Single-layer LSTM 88.89 ± 1.67 Inline graphic
Training strategy ablations
 w/o data agmentation 86.67 ± 2.34 Inline graphic
 w/o dropout 85.55 ± 2.45 Inline graphic
 w/o weight decay 87.78 ± 1.98 Inline graphic
 w/o class confusion aug 91.11 ± 1.23 Inline graphic