Three pain decoders (NPS, pPV, and pNsy in
A–C) and a voice decoder
(D) were used to test identification
for mental states. x-axes are the
normalized dot products between decoder and decodee, while
y-axes are the posterior
probability of being in pain (A–C) or
listening to voices (D). Distributions of
normalized dot products and posterior probabilities include
both the decodee (light grey & colors) and comparator
(dark grey) tasks. (A–C) Normalized dot
products of the pain condition span the entire distribution
of comparator normalized dot products, and as a result, pain
is not adequately isolated from the comparator conditions.
Quantitatively, this is evidenced by the strong
decodee-comparator overlap for (A) NPS (overlap
(95%CI) = 68% (59–82)), (B) pPV (79%
(73–90)), and (C) pNsy (73%
(66–84)). This is reflected in the Bayesian model,
which shows similar probabilities of being in pain for both
pain and pain-free conditions (each dot/line). To this end,
all three decoders perform similarly, and cannot
unequivocally identify pain, as indicated
by their sensitivity/specificity (threshold from
Youden’s J statistic, chosen in-sample) of (NPS,
A) 0.64/0.74, (pPV, B)
0.6/0.64, and (pNsy, C) 0.54/0.76.
(D) In contrast to pain, a contrast map
decoder for identifying when a participant is listening to
human voices separates more clearly the normalized dot
products of the decodee (red) from comparator (dark grey),
but still performs poorly (overlap = 54% (46–66)).
This separation is reflected in the Bayesian model, which
shows high probabilities when individuals are listening to
human voices and lower probabilities when they are not.
Using a threshold determined by Youden’s J statistic
(chosen in-sample), the voice decoder has a
sensitivity/specificity of 0.77/0.64. In (A),
(B), (C) the dataset used
were not used in the training of the decoders (NPS, pPV,
pNsy); tests are all out of sample. In (D), we
split the dataset into a training set (107 subjects) and a
testing set (106 subjects).