To summarize briefly, I followed the step of Beam et al. (2021, Nature Neuroscience) to find evidence for the structure of human cognition. Firstly, filtering all studies that included the testing paradigms that consisted with those used in our project from the database of Neurosynth, I conducted 28 meta-analysis separately on each group of fMRI studies. Thanks to NiMARE (a python research environment for neuroimaging meta-analysis), I could easily perform Multilevel Kernel Density Analysis (MKDA) on hundreds fMRI studies without manually coding.

Code

cog_label <- readxl::read_xlsx(box::file("Selected_cogconstruct_label.xlsx"))
label_occur <- readxl::read_xlsx(box::file("dtm_210114.xlsx"))

occur_combined <- label_occur %>%
  pivot_longer(cols = -c("study_id"), names_to = "term_name", values_to = "occurrence") %>%
  left_join(cog_label, by = "term_name") %>%
  filter(!is.na(meta_name)) %>%
  group_by(study_id, meta_name) %>%
  summarize(task_occurrence = sum(occurrence)) %>%
  filter(task_occurrence > 1)
total_occur <- occur_combined %>%
  group_by(meta_name) %>%
  summarize(total_occurrence = length(study_id)) %>%
  filter(total_occurrence > 16) %>%
  mutate(Paradigm_name = meta_name,
         Number_of_studies = total_occurrence) %>%
  select(Paradigm_name, Number_of_studies) %>%
  arrange(desc(Number_of_studies))

Here is the summarize table for the number of included fMRI studies in each meta-analysis.

Code

total_occur |>
  htmlTable::htmlTable()

	Paradigm_name	Number_of_studies
1	Stroop_task	1193
2	Go/no_go_task	752
3	N_back_task	572
4	Word_generation_task	359
5	Oddball_task	203
6	Iowa_gambling_task	201
7	Recognition_task	184
8	Wisconsin_card_game	179
9	Anti_saccade_task	177
10	Flanker_task	170
11	Switching_task	165
12	Calculation_task	151
13	Stop_signal_task	130
14	Mental_rotation_task	117
15	Selective_attention_task	117
16	Tower_of_London_task	103
17	Verbal_fluency_task	86
18	Visual_search_task	85
19	Auditory_discrimination_task	84
20	Continuous_performance_task	83
21	Simon_task	82
22	Choice_reaction_time_task	74
23	Reasoning_task	61
24	Delay_memory_task	51
25	Orientation_match_task	28
26	Trial_making_test	28
27	Forward_span_task	25
28	Complex_span_task	20
29	Simple_reaction_time_task	17

Code

total_matrix <- fs::dir_ls(box::file("data"), regexp = "xlsx") %>%
  map(readxl::read_xlsx, col_names = TRUE) %>%
  reduce(left_join, by = "voxel_num")

Correlation Matrix

Here is the representational similarity matrix of the 28 meta-analysis results, representing the similarity of the brain activities during these 28 different types of cognitive tasks.

Code

total_nonzero <- total_matrix[which(rowSums(total_matrix) > 0),] %>%
  select(-voxel_num)
cor_total <- round(cor(total_matrix),2)
cor_nonzero <- round(cor(total_nonzero),2)
melted_cor_nonzero <- melt(cor_nonzero) %>%
 mutate(percrank=rank(value)/length(value))

correlate(total_nonzero, quiet = TRUE) |>
  rearrange(method = "HC") |>
  stretch() |>
  mutate(across(c(x, y), as_factor)) |>
  ggplot(aes(x, y)) +
  geom_tile(aes(fill = r)) +
  scico::scale_fill_scico(palette = "bam", midpoint = 0, direction = -1) +
  coord_fixed() +
  theme_minimal(base_size = 18) +
  labs(x = "", y = "", fill = "Pearson's", color = "") +
  theme(axis.text.x = element_text(angle = 45, hjust = 1))

Code

# ggplot(data = melted_cor_nonzero, aes(Var2, Var1, fill = percrank))+
#  geom_tile(color = "black")+
#   scale_fill_viridis(discrete=FALSE)+
#   theme_minimal()+
#  theme(axis.text.x = element_text(angle = 45, vjust = 1,
#     size = 7, hjust = 1),
#     axis.text.y = element_text(size = 7))+
#  coord_fixed()

Multidimensional Scaling

Code

mds <- total_nonzero |>
  cor(use = "pairwise") |>
  smacof::sim2diss(to.dist = TRUE) |>
  smacof::mds(ndim = 2, type = "mspline")
# plot(mds, plot.type = "Shepard", main = "Shepard Diagram (Ratio Transformation)")
# par(family = "SimHei")
plot(mds)

Exploratory Factor Analysis

Traditional

Code

nfactors_test <- psych::nfactors(cor_nonzero)

Correlated latent factors

Code

fit <- psych::fa(cor_nonzero, 4)
psych::fa.diagram(fit)

Bifactor model

Code

fit_bifac <- psych::omega(total_nonzero, 3, plot = FALSE)
psych::omega.diagram(fit_bifac)