Cross-Subject Priors for Maximum a Posteriori Estimation of Latent Structure Influence Models in Motor-Imagery EEG With Few Trials
Authors: Sagnik De, Anil Kumar Gupta
Organizations: Institute of Radio Physics & Electronics, University of Calcutta, Kolkata 700009, India · HPC Infrastructure and EcoSystem Group, Centre for Development of Advanced Computing, Pune 411008, India
Coupled hidden Markov models such as the latent structure influence model (LSIM) describe how the sensorimotor rhythms of the two hemispheres evolve and interact during motor imagery, but subject-specific maximum likelihood (ML) estimates are unstable when few trials are available. This paper derives a maximum a posteriori (MAP) estimator for the LSIM whose conjugate prior equals the expected statistics of λ pseudo-trials from a population model fitted to other subjects. The M-step remains in closed form, expectation maximization keeps its monotone ascent, and ML (λ=0) and the population model (λ→∞) are limiting cases. The prior strength was selected on 35 subjects of the PhysioNet EEG Motor Movement/Imagery Dataset and applied unchanged to 70 held-out subjects and to BCI Competition IV-2a. With λ=1 the held-out log-likelihood exceeded that of ML in 61 of 70 subjects (Holm-adjusted p=9.1×10−8) and in all 70 subjects with five training trials per class. A prior learned only from the PhysioNet data improved the held-out likelihood of all nine BCI IV-2a subjects at five trials per class and reduced the error of the joint transition matrix by 61%. Decoding accuracy did not improve (p=0.17). Cross-subject priors therefore stabilize few-trial dynamic models of motor-imagery EEG, but they do not improve decoding, which in these data relies mainly on structure shared across subjects.
Figures & tables
Analysis
Data
Subjects
Trials
Selection of λ
EEGMMIDB
35
14 to 20
Primary outcomes
EEGMMIDB
70
15 to 20
Reduced training sets
EEGMMIDB
70
5, 10
Split-half reliability
EEGMMIDB
70
9 to 12
External validation
BCI IV-2a
9
5 to 30
Exploratory analysis
EEGMMIDB
105
5, 14 to 20
TABLE I: Summary of the analyses. Trials are training trials per class and subject.
Fig. 1: (a) Mean change in held-out log-likelihood per frame relative to ML on the 35 tuning subjects as a function of the prior strength. The dotted line marks the selected λ=1 . Flat-MAP continues below the axis for λ≥2 . (b) Excess held-out loss relative to the true model in the simulation with ground truths fitted to the tuning subjects (mean and standard error over 35 synthetic subjects, logarithmic axes) for the selected prior strengths.
Trials
ℓˉ
Group-MAP minus ML
Group-MAP minus Flat-MAP
per class
ML
Flat-MAP
Group-MAP
Pooled
mean [95% CI]
better
p
mean [95% CI]
better
p
5
− 5.4837
− 5.4443
− 5.4021
− 5.4121
0.0816 [0.0689, 0.1019]
70/70
3.6×10−13
0.0422 [0.0366, 0.0487]
69/70
2.1×10−12
10
− 5.3847
− 5.3772
− 5.3657
− 5.4121
0.0190 [0.0133, 0.0237]
67/70
4.0×10−10
0.0115 [0.0091, 0.0141]
65/70
9.5×10−11
15 to 20
− 5.3523
− 5.3503
− 5.3472
− 5.4121
0.0051 [0.0026, 0.0072]
61/70
3.0×10−8
0.0031 [0.0018, 0.0043]
60/70
1.2×10−7
accuracy (%)
Group-MAP minus ML (points)
ML
Flat-MAP
Group-MAP
Pooled
LDA
mean [95% CI]
better/worse
p
TABLE II: Results for the 70 EEGMMIDB test subjects: held-out log-likelihood per frame ℓˉ (nats), paired differences with bootstrap 95% CIs, numbers of subjects improved, two-sided Wilcoxon p -values (unadjusted) and decoding accuracies.
Fig. 2: Per-subject differences in held-out log-likelihood per frame for the 70 test subjects with 5, 10 and all (15 to 20) training trials per class. The axis is symmetric logarithmic and linear within ±0.01 . Horizontal bars mark the median and quartiles, and the counts under each group give the subjects with a positive difference.
Summary
ML
Group-MAP
difference
desynchronized-state occupancy, C3 ′
0.56
0.58
[ − 0.06, 0.10]
desynchronized-state occupancy, C4 ′
0.33
0.45
[ − 0.04, 0.28]
IR→L
0.17
0.22
[ − 0.18, 0.27]
IL→R
0.31
0.33
[ − 0.21, 0.25]
state μ -power difference, C3 ′
0.60
0.57
[ − 0.20, 0.16]
state μ -power difference, C4 ′
0.68
0.58
[ − 0.28, 0.07]
TABLE III: Split-half reliability (Spearman ρ across the 70 test subjects) of subject-level summaries, and bootstrap 95% CI of the difference Group-MAP minus ML.
Trials
held-out ℓˉ
Group vs ML
∥P^−Pref∥F2
Group vs ML
accuracy (%)
per class
ML
Flat-MAP
Group-MAP
Group-MAP (IV)
better
p
ML
Group-MAP
better
p
ML
Group-MAP
LDA
5
− 5.2367
− 5.2142
− 5.1736
− 5.1741
9/9
0.004
0.0793
0.0308
9/9
0.004
58.2
58.9
60.7
10
− 5.1560
− 5.1525
− 5.1416
− 5.1417
7/9
0.02
0.0421
0.0241
9/9
0.004
60.5
61.4
65.3
20
− 5.1226
− 5.1217
− 5.1190
− 5.1190
8/9
0.10
0.0263
0.0198
9/9
0.004
61.9
62.3
67.3
30
− 5.1121
− 5.1115
− 5.1106
− 5.1103
6/9
0.25
0.0220
0.0172
6/9
0.36
63.9
62.4
68.5
TABLE IV: External validation on BCI IV-2a ( n=9 ). Group-MAP uses a prior learned only from EEGMMIDB, and Group-MAP (IV) uses a prior from the other eight BCI IV-2a subjects. The Pooled EEGMMIDB model, which uses no BCI IV-2a data, reached ℓˉ=−5.2660 , a transition error of 0.0469 and an accuracy of 64.2%. Values are means over subjects and, for fewer than 30 trials, over 20 training draws.
Fig. 3: External validation on BCI IV-2a as a function of the number of training trials per class (means over nine subjects). (a) Held-out log-likelihood per frame. (b) Squared error of the joint transition matrix relative to an independent reference fit (logarithmic axis). (c) Decoding accuracy. The dashed line is the pooled EEGMMIDB model, which uses no BCI IV-2a data.
Fig. 4: Exploratory analysis on all 105 EEGMMIDB subjects. (a) Held-out log-likelihood per frame and (b) decoding accuracy of Group-MAP as functions of the prior strength (ML isplotted at the left end). The dashed line marks the pooled models, which use no data of the target subject.