Fig 1.
Scheme of the experimental structure and LLP classifier.
Top to bottom: The sentence “Franzy jagt …” is spelled three times. To spell a single character in one trial, 68 highlighting events occur, with 32 belonging to sequence 1 and 36 belonging to sequence 2. The resulting 68 ERP responses (epochs) are averaged for each sequence, and these averages are exploited to reconstruct the mean target and non-target ERP responses.
Fig 2.
Classification results of LLP applied on artificial data sets generated from an auditory (A) and a visual (B) ERP paradigm.
For each artificial data set, the target vs non-target accuracy result for a supervised shrinkage-LDA and different mixing matrices Π1–Π4 is shown. The first column in the matrices always denotes the target ratios while the second column denotes the non-target ratio. NAF = noise amplification factor.
Fig 3.
Grand average (N = 13) visual ERP response.
Top row: Average responses evoked by visual target (blue) and non-target (green) stimuli in the occipital channel O1 (thick) and the central channel Cz (thin) during the online experiment. Prior to averaging, a baseline correction was performed based on data within the interval [-200, 0] ms. The signed R2 values for channels O1 and Cz over time are provided by two horizontal colour bars. Their scale is identical to the scale of the plots in the bottom row of scalp plots. Middle rows: Scalp plots visualising the spatial distribution of mean target and non-target responses within four selected time intervals: [50 120], [120 200], [201 380] and [381 700] ms relative to stimulus onset. Bottom row: Scalp plots with signed R2 values indicate spatial areas with high class-discriminative information.
Table 1.
Overview of neurophysiological features and supervised classification performance.
Fig 4.
ERP responses for S11 of sequence 1 (A) and sequence 2 (B).
Top row: Average responses evoked by target (blue) and non-target (green) stimuli in the occipital channel O1 (thick) and the central channel Cz (thin). Prior to averaging, a baseline correction was performed based on data within the interval [-200, 0] ms. Bottom rows: Scalp plots visualising the spatial distribution of mean target and non-target responses within four selected time intervals: [50 120], [120 200], [201 380] and [381 700] ms.
Fig 5.
ERP responses for S6 of the reconstructed class-wise means using LLP (A) and original labelled data (B). C shows the LLP target estimations in [120 200] ms for different numbers of training points.
For details, see description of Fig 4.
Fig 6.
Spelling performance as seen online using the LLP (A) and after the post hoc re-analysis with LLP (B).
Top: Each row represents a single spelling of the test sentence “Franzy jagt …”, with yellow squares indicating incorrectly spelled characters and blue squares indicating correctly spelled characters. Bottom: The averaged spelling accuracy across sentences and subjects is shown for each character.
Fig 7.
Comparison of LLP to an unsupervised EM-based classification approach for each sentence (A) and after 5 characters (B).
A: Thin lines represent the binary target vs. non-target AUC performance of the two learning models with every line corresponding to the spelling of a single sentence. Each of the subjects (N = 13) spelled each sentence three times resulting in 39 lines. Dashed lines depict average performances. Please note that the supervised shrinkage-LDA was trained and tested in a 5-fold crossvalidation to avoid overfitting. This means that the supervised method only had 80% of the data compared to the unsupervised methods. B: Each dot represents the EM and LLP performance after 5 characters for the same subject and sentence.