Tuesday, June 29, 2021

The rs-FC fMRI Law of Attraction (i.e., Resting-State Functional Connectivity of Speed Dating Choice)


Feeling starved for affection after 15 months of pandemic-mandated social distancing? Ready to look for a suitable romantic partner by attending an in-person speed dating event? Just recline inside this noisy tube for 10 minutes, think about anything you like, and our algorithm will Predict [the] Compatibility of a Female-Male Relationship!

This new study by Kajimura and colleagues garnered a lot of attention on Twitter, where it was publicized by @INM7_ISN (Simon Eickhoff) and @Neuro_Skeptic. The prevailing sentiment was not favorable (check the replies)...

 

 

Full disclosure: I was immediately biased against the claims made in this study...

This research emphasizes the utility of neural information to predict complex phenomena in a social environment that behavioral measures alone cannot predict.

...and have covered earlier attempts at linking speed dating choice to a proxy of neural activity. But I wanted to be fair and see what the authors did, since their results reflect an enormous amount of work.

Here I will argue that a 10 minute brain scan cannot predict who you will choose at a speed dating event. The resultant measures are even further away from identifying a compatible mate for you, since only 5% of speed dating interactions result in a relationship of any sort (6% for sexual relationships and 4% for romantic relationships, according to one study).

I was flabbergasted that anyone would think a “resting” state MRI scan (looking at “+” for 10 min) and its resulting pattern of correlated BOLD signal fluctuations would reflect a level of superficial desirability that can be detected by a potential mate at greater than chance level. Another disclosure: this is far from my field of expertise. So I searched the literature. Apparently, “patterns of functional brain activity during rest encode latent similarities (e.g., in terms of how people think and behave) that are associated with friendship” (Hyon et al., 2020). However, that study was conducted in a small town in South Korea (total population 860), allowing a detailed social network analysis. Plus, people knew each other well and experienced many of the same day to day events, which could shape their functional connectomes. Not exactly relevant for predicting strangers' speed dating choices, eh?

Another paper identified a “global personality network” based on data from 984 participants in the Human Connectome Project (Liu et al., 2019). The sample was large enough to support a training set of n=801 and a “hold-out” dataset (n=183) for validation purposes. The results supported the authors' “similar brain, similar personality” hypothesis. But in the dating world, how much do “similars” attract (compared to the popular saying, “opposites attract”)? Well why not construct (dis)similarity profiles between potential pairs by taking the absolute value of differences in functional connectivity (FC), and combine those with values of similarities in FC? Does that make sense?? And in order to arrive at this metric, there's a whole lot of machine learning (but with much smaller training sets)...

Identity Classification 

A separate sample of 44 individuals from the Human Connectome Project was used to construct the Similarity of Connectivity Pattern between pairs (Kajimura et al., 2021). These 44 participants had each been scanned twice, allowing 44 self-self pairs (Jessica at time 1 vs. Jessica at time 2), which were compared to 44 self-other pairs (Jessica at time 1 vs. Jennifer at time 2). Self-self “feature values” always show a positive correlation, and these were used to define “individual-specific information.”

26,680 feature values?

To start, 116 regions of interest (ROIs) were defined by Automated Anatomical Labeling (AAL). Pairwise comparisons of these for Self scan #1 vs. Self scan #2 (or vs. Other scan #2) resulted in a vector of 6,670 functional connectivities for each data point [(116 × 115)/2]. Then multiply this by four (!!) and you get 26,680 values fed into a machine learning classifier. Why four? Because the slow fluctuating BOLD signals were decomposed into four frequency bands for the classification procedure. Was this necessary? Does it add robustness, or merely more opportunities for false positive results?


Fig. 3 (Kajimura et al., 2021). Top 100 feature values, i.e. absolute values of differences between functional connectivity that contributed to identity classification for three frequency bands [the fourth was eliminated because the classifier could not distinguish between self-self and self-other pairs].

The machine learning algorithm was sparse logistic regression with elastic-net regularization (SLR-EN), which usually prevents overfitting, but I don't know if the algorithm can overcome 26,680 feature values with only 44 subjects. Maybe I'm misunderstanding (and others can correct me if I'm wrong), but the number of participants is rather low for SLR-EN given the number of input parameters? Then...
The classification accuracy was evaluated using a stratified k-fold cross-validation procedure. ... The ratio of the number of correctly classified labels was then obtained as the classification accuracy.


The regional results are below, showing a 7 x 7 brain network matrix with similarity in red (positive coefficients) and dissimilarity in blue (negative coefficients). We're still in the realm of correctly classifying self-self, so dissimilarites were considered artifacts of overfitting [but similarities were not?]. If the contribution from similar > dissimilar with binomial tests, this was considered an indicator of self. This was true of F1 (53 out of 67, p<.001) and F2 (52 out of 67, p<.001), but not F3, which was at chance (33 out of 67).


Fig. 4 (modified from Kajimura et al., 2021). Ratio of self-self classification connectivity in terms of brain networks. Red and blue matrices display the results of similarity- and dissimilarity-based contributions [at three frequency bands]. ... Vis, visual network; Som, somatosensory-motor network; Sal, salience network; Lim, limbic system; Con, executive control network; Def, default mode network; Cer, cerebellum.


Separate Statistical Analysis — a bevy of Pearsons 

Before we turn to speed dating, two more analyses are shown below for the identity classification study. The first involved a boatload of FDR-corrected Pearson’s correlations of the functional connectivity vectors for self–self pairs vs. self–other pairs (Fig 2A). The next shows the effectiveness of the machine learning (ML) algorithm in classifying these pairs (Fig 2B).

 

Fig. 2 (modified from Kajimura et al., 2021). Identity classification. (A) Similarities in overall functional connectivity profile was significantly higher for the self–self pair (dark-colored distribution) than the self–other pair (light-colored distribution) for all frequency bands. [I've included arrows to point out where they start to diverge] (B) Distribution of differences between ML classification accuracy.

As the authors predicted, self-self comparisons yielded more similar connectivites than self-other pairs. The ML algorithm identified self for three of the frequency bands (F1-F3) at greater than chance levels (12.4%. 14.8%, and 16.3% better than chance, respectively). However, the algorithm is still wrong a lot of the time. This is especially important for the matchmaking study...


Speed Dating

The authors provided a nice self-explanatory graphic presenting an overview of the Speed Dating study (click on image for a larger view). Data collection and analysis followed the flow of the Identity experiment.



Participants and Social Event

The participants were 42 heterosexual young adults (20-23 yrs), with 20 females and 22 males. Why these numbers were not perfectly matched, I do not know. The resting state fMRI scan was several days before the first speed dating session. [I'm assuming it was the first, because the Methods say there were three speed dating events. There was also a post-dating scan, which was described in another paper]. The three hour event was held in a large room where pairs of participants had 3 min long conversations with every member of the opposite sex. After each conversation, all the men moved to the next table. When all the speed dates were over, each person was asked to identify at least half of the opposite sex individuals they'd like to chat with again.

Well, there's a problem here — a requirement to select half the dates could result in less-than-optimal choices in some individuals. This requirement was necessary for sampling purposes, but it makes you wonder about the quality of the matches. Also, there was a strong possibility of unilateral matches — one individual thinks they've found their dream partner but that feeling was not reciprocated. When both members of a pair said "yes" they were considered compatible. Out of a total of 440 possible pairs, 158 were compatible and 282 incompatible.

The Compatible vs. Incompatible comparisons are the key findings of the study (Fig. 5, with A and B panels as above). Unlike the Identity comparison, compatible male-female pairs did not show more similar functional connnectivity patterns than incompatible pairs (Fig. 5A).

Well then...

“This indicates that the compatibility of female–male relationships is not necessarily represented by the similarity of functional connectivity patterns.”

Yes.

“Unlike identity classification, compatibility classification was supported by the considerable negative coefficients of the features” (shown in Fig. 6 of the paper). We shall not interpret this as opposites attract

 


 
Fig. 5 (Kajimura et al., 2021). Compatibility classification. (A) Similarity of overall functional connectivity profile. There was no significant difference between compatible (dark-colored distribution) and incompatible (light-colored distribution) pairs. (B) Distribution of differences between the classification accuracy with true labels of pairs and that with a randomized label for each frequency band. Vertical lines indicate chance levels.


Fig. 5B shows classification accuracy for compatible pairs, which was above chance for F1 and F2. Before investing in a commercial venture,  however, you should know that the benefit beyond guessing is only 5.47% and 4.95%, respectively. Thus, I disagree with the claim that...
...the current results indicate that resting-state functional connectivity has information about behavioral tendencies that two individuals actually exhibit during a dyadic interaction, which cannot be measured by self-report methods and thus may remain hidden unless we use neuroimaging methods concurrently.

To review the potential limitations of the study, we can't assess the quality of matches (meh vs. enthusiastic), we don't know what the participants were thinking about during their rsfMRI scan (see Gonzalez-Castillo et al., 2021), and we don't know their mental state during the scan. Although rs-FC fMRI is often considered a stable trait”, state factors and motion artifacts can affect the results on a given day (Geerligs et al., 2015). Indeed, ~35% of the time, the present paper was unsuccessful in classifying the same person run on two different days (and that's excluding one of four frequencies that was not above chance).

Is there something intrinsic encoded in BOLD signal fluctuations that can predict who we will find appealing (and a potential “match) after a three minute interaction?  Decisions at speed dating events are mostly based on physical attractiveness, so it seems very implausible to me.


Further Reading (the Speed Dating Collection)

The Neuroscience of Speed Dating Choice

The Electroencephalogram Cocktail Party

EEG Speed Dating

The Journal of Speed Dating Studies

Winner of Best Title

How I Meditated with Your Mother: Speed Dating at Temples and Shrines in Contemporary Japan

 

References

Geerligs L, Rubinov M, Henson RN. (2015). State and trait components of functional connectivity: individual differences vary with mental state. Journal of Neuroscience 35(41):13949-61.

Gonzalez-Castillo J, Kam JW, Hoy CW, Bandettini PA. (2021). How to Interpret Resting-State fMRI: Ask Your Participants. Journal of Neuroscience 41(6):1130-41.
 
Hyon R, Youm Y, Kim J, Chey J, Kwak S, Parkinson C. (2020). Similarity in functional brain connectivity at rest predicts interpersonal closeness in the social network of an entire village. Proceedings of the National Academy of Sciences 117(52):33149-60.

 
 
 

Subscribe to Post Comments [Atom]

eXTReMe Tracker