Emotion-laden events and objects are typically better remembered than neutral ones. This is usually explained by stronger functional coupling in the brain evoked by emotional content. However, most research on this issue has focused on functional connectivity evoked during or after learning. The effect of an individual’s functional connectivity at rest is unknown. Our pre-registered study addresses this issue by analysing a large database, the Cambridge Centre for Ageing and Neuroscience, which includes resting-state data and emotional memory scores from 303 participants aged 18–87 years. We applied regularised regression to select the relevant connections and replicated previous findings that whole-brain resting-state functional connectivity can predict age and intelligence in younger adults. However, whole-brain functional connectivity predicted neither an emotional enhancement effect (i.e., the degree to which emotionally positive or negative events are remembered better than neutral events) nor a positivity bias effect (i.e., the degree to which emotionally positive events are remembered better than negative events), failing to support our pre-registered hypotheses. These results imply a small or no association between individual differences in functional connectivity at rest and emotional memory, and support recent notions that resting-state functional connectivity is not always useful in predicting individual differences in behavioural measures.
Emotional events are typically remembered better and more vividly relative to neutral ones1,2,3. This emotional enhancement effect has been found in laboratory studies4,5,6,7,8 as well as autobiographical memory9. Previous research has suggested that individual differences in this emotional memory enhancement effect may have important consequences on wellbeing and psychopathology. For example, the enhancement effects of emotion on memory are considered to result in spontaneous and intrusive recollection of traumatic memories10. Likewise, one’s tendency to preferentially remember negative information is frequently present in psychopathological conditions, including depression and anxiety, and is associated with symptom severity11. This negative memory bias has also been associated with smaller hippocampal grey and white matter volume12, which is in turn associated with major depression13. In contrast, one’s tendency to preferentially remember positive over negative information is referred to as “positivity bias”, and is often associated with better emotional wellbeing in old age14,15. In the current study, we tested whether such individual differences in the emotional enhancement effects of memory can be predicted by resting-state functional connectivity in the brain. Functional connectivity (FC) refers to the strength of connections between brain areas that share functional properties. We distinguish task-induced FC in response to a stimulus, from resting-state FC, which reflects the connectivity of an individual at rest.
The brain mechanisms behind the emotion-induced enhancement effects in memory have been intensively studied in task-fMRI studies, where researchers examined blood oxygen level dependent (BOLD) signals obtained while participants encoded emotional vs. non-emotional information. Meta-analyses based on these studies reported that the enhancement of emotional memory is associated with increased activation in the amygdala, hippocampus, and regions in the ventral visual stream during the encoding of emotional items16,17. In addition to the activation level, increased task-induced FC across the amygdala, hippocampus and the prefrontal cortex (PFC) during encoding of emotional items is associated with enhanced memory for emotional compared with neutral items18,19,20,21,22. Previous studies also extended their focus to FC after learning (i.e., during consolidation)23,24. Stronger FC between the amygdala and visuosensory areas after learning was associated with the negative memory bias in memory, whereas stronger FC between the amygdala and anterior cingulate after learning was associated with the positivity bias in memory24.
In contrast, it has been less clear whether resting-state FC before learning predicts individual differences in emotional memory. Resting-state FC refers to the temporal correlation in activity between regions that are not actively engaged in any task, and is considered to reflect the brain’s functional and structural connectivity25. Individual differences in resting-state FC have been used to predict individual differences in brain activation during various tasks, including working memory, language tasks, emotion recognition, and interpreting social interactions26. Research on memory has further demonstrated an association between memory performance for neutral items and resting-state FC of the MTL27,28 and the default mode network (DMN) which has been implicated in age-related cognitive decline29,30.
In addition, recent advances in machine learning have allowed researchers to identify and study complex data models, that can be used to predict individual differences from a wide range of behavioural and cognitive measures31. Studies implementing such analyses found that resting-state FC predicts behavioural measures including attention span32,33, decision-making strategies34, intelligence35,36, motor skills learning37 and personality36, acting as a behavioural “fingerprint”35.
In contrast, few studies have investigated whether resting-state FC predicts individual differences in emotional memory. On the one hand, FC during rest resembles the FC observed during a task38,39 and previous findings support an association between emotional memory enhancement effects and FC during rest before22,40 or after encoding23,24. Therefore, it is reasonable to hypothesise that whole-brain resting-state FC is predictive of individual differences in emotional memory. On the other hand, recent evidence emphasised that robust cognitive tasks may not always yield reliable inter-individual measures41. Similar low reliability was also reported for the emotional enhancement effect in memory, despite robust and strong group-wise effects for better memory for emotional rather than neutral items42. Therefore, even though resting-state FC has a relatively high temporal reliability35,43, resting-state FC may not be able to reliably predict emotional memory enhancement effects.
We investigated whether resting-state FC predicts emotional memory using a large database—the Cambridge Centre for Ageing and Neuroscience (Cam-CAN)—that includes emotional memory scores, structural and functional MRI (fMRI) scans of 303 individuals of ages 18–87 years44,45. In the Cam-CAN project, participants completed an emotional memory task (in a different session from the MRI session), where they learned neutral objects superimposed onto emotionally positive, neutral, and negative backgrounds46. Consistent with the emotion induced enhancement effect observed in the literature, participants had a better memory for objects learned with positive or negative backgrounds than objects learned with neutral backgrounds (Table 1). Based on this task, we created two continuous measures of emotional memory: (a) better memory for positive and negative information than neutral information (the emotional enhancement effects) and (b) preferential memory for positive rather than negative information (the positivity bias). Our study also attempts to predict age and intelligence from resting-state FC; these latter analyses served as control checks to ensure that our method and data can replicate previous findings35,36,47.
We preregistered the above hypotheses and analysis pipelines, which are accessible at https://osf.io/untzm. Following an analysis pipeline previously used to predict individual differences in personality and intelligence from resting-state FC36, we used regularised linear regression to predict the emotional enhancement effect and the positivity bias in memory from whole-brain resting-state FC. The brain was parcellated into 268 nodes obtained from Shen et al.48. Seven nodes were excluded from the analysis due to missing data, therefore comprising a total number of predictors of 33,930 connections. Due to the expected collinearity and large number of predictors, we used common parameter regularization techniques to avoid over-fitting of the data models. Specifically, we used Elastic Net penalization, which combines ridge (L1) and lasso (L2) penalization schemes. Ridge regularization adds a Gaussian prior to the parameters of the model. Lasso penalization provides an upper bound to the parameter, while creating opportunities to reduce the number of predictors altogether. Additionally, we used leave-one-out cross-validation to train and test the models, and permutation testing to compute a p-value when R2 showed a positive relationship (permutation analyses were not run when R2 was negative because negative R2 means that the models performed poorly).
The analysis procedure for the main analyses (where we predicted the emotional enhancement effects, the positivity bias and intelligence from resting-state FC across all participants) was preregistered, and the scripts used are publicly available (https://osf.io/bm98y).
A composite score for intelligence was computed from the four subsets of Cattell through principal component analysis. The derived factor explained 67.8% of the total variance, and had loadings ranging from 0.81 to 0.84 with the four Cattell subsets. As reported in the original paper about the dataset46, participants showed better object memory for positive and negative backgrounds than neutral backgrounds (Table 1). We further computed a measure of the emotional enhancement effect variable by subtracting object memory performance in the neutral condition from the average memory performance in the positive and negative conditions. We also created another measure of positivity bias by subtracting object memory performance in the negative condition from memory performance in the positive condition.
From the original dataset, we used 303 participants (Table 2)—all the participants in the database who completed the resting-state fMRI, emotional memory task, and the intelligence test. The data included 261 nodes, as one or more of those seven nodes—located in the left and right temporal lobes—were missing for 45 participants. We computed the following exploratory correlations analyses as quality checks. Older individuals performed more poorly on the intelligence score than younger individuals, r(301) = − 0.63, p < 0.001. In contrast, age was not significantly correlated with the positivity bias, r(301) = 0.11, p = 0.06, nor with the emotional enhancement effect of memory, r(301) = 0.01, p = 0.89. There were no significant gender differences in intelligence, t(301) = 1.77, p = 0.08, the positivity bias, t(301) = − 1.40, p = 0.17, or in the emotional enhancement effect, t(301) = 0.27, p = 0.79.
Preregistered predictive modelling
We followed a strategy first described by Dubois et al.36. Before running prediction analyses, preprocessing and denoising pipelines were run on the resting-state images. The pipelines included (1) applying motion correction, (2) registration to the standard Montreal Neurological Institute (MNI) brain template, (3) detrending the white matter and cerebrospinal fluid through removing temporal drifts with third-degree Legendre polynomial regressors, (4) regressing out mean signals of the white matter and cerebrospinal fluid from the grey matter signal, (5) regressing out motion parameters from the whole brain, (6) removing high-frequency noise by applying a low-pass filter (1 TR which is 1970 ms in this study), (7) detrending the grey matter signal though removing temporal drifts with third-degree polynomial Legendre regressors, and (8) regressing out global signals from the whole brain signal.
Prediction analyses began with filtering, whereby only edges with correlations of p-value < 0.01 with the predicted variable were included in the subsequent analyses. We used Elastic Net models with a high ratio of ridge (0.9) and tuned the models’ alpha parameter through a grid search. Analyses were run to predict the emotional enhancement effect in memory, the positivity bias in memory, age, and intelligence from the connectivity matrix. The control variables were age, gender, handedness and intelligence (unless they are the predicted variable) which were regressed out from the predicted variables (see Supplementary Table 1 for analyses including motion as a control variable). The models were trained in leave-one-out cross validation. We ran one thousand permutations of the data, which allowed us to calculate one-tailed p-values for each model that returned positive R2. Results are shown in Fig. 1. The models predicting the emotional enhancement effect and the positivity bias performed poorly, demonstrating negative correlations between the predicted and observed values (Table 3; Fig. 1). The model predicting intelligence also performed poorly and did not achieve a significant correlation between predicted and observed values (Table 3; Fig. 1).
As described in the previous section, our pre-registered analyses failed to predict our two emotional memory measures from resting-state FC. We also failed to replicate previous findings showing that resting-state FC can predict intelligence. We therefore ran a series of unplanned exploratory analyses to identify when resting-state FC predicts behavioural measures. First, we ran an exploratory analysis to test if we can replicate previous findings that one’s chronological age is predicted by resting-state FC47,49. The model obtained good prediction, achieving strong correlation between predicted and observed values, r(301) = 0.44 (Table 3; Fig. 1), suggesting that resting-state FC is predictive of an individual’s age.
Analysis for each age group
Next, we performed an exploratory analysis after splitting the sample into three age groups: younger (aged 18–40 years; N = 85), middle-aged (aged 41–60 years; N = 98) and older adults (aged 61 years and over; N = 120) given that previous studies on intelligence and resting-state FC primarily focused on younger adults35,36, whereas our participants included those aged between 18 and 87 years. Note that past studies also showed the non-linear effects of age; suggesting that older adults may rely on a different set of regions (relative to younger adults) to perform the same task50,51. To test the possibility that the emotional enhancement effect, the positivity bias and intelligence are successfully predicted after splitting participants into separate age groups, the analyses were repeated separately for each age group, with the same methodology as the whole-sample analyses described above. The model successfully predicted intelligence in younger adults, but not for middle-aged or older adults (Table 4). However, the model still failed to predict the emotional memory enhancement effect and the positivity bias across all groups (Table 4).
Other emotional memory measures
Results presented so far concerned memory accuracy for neutral objects that were superimposed on negative, neutral or positive images (so called ‘object memory’). Yet the CamCAN study tested three types of memory: object, associative valence and background memory45. While the effects of valence on this object memory measure were significant, they were relatively small46; which may have resulted in our failure to predict the emotional memory enhancement effects using resting-state FC. To address this issue, we applied the same analysis method again to the two other types of memory in the Cam-CAN dataset: associative valence memory and background memory. The associative valence memory measure concerns whether each correctly-recalled neutral object was associated with a positive, negative or neutral background and showed stronger effects of valence compared with object memory46 (see Table 1). In contrast, the background memory concerns participants’ gist memory for contents of the positive, negative and neutral background images. This gist background memory also showed significant effects of valence, such that participants had a better background memory for the negative than the positive condition, which was better than the neutral condition (see Table 1).
As done in the object memory, we obtained the emotional enhancement effect and the positivity bias for both the associative and background memory measures and ran the same set of analyses. But the models derived from resting-state FC could not significantly predict either the emotional enhancement effect or the positivity bias even in these measures (Table 5). We also ran the same analysis after splitting participants into three age groups, but the models could not predict the emotional enhancement effect or positivity bias in any group.
To check that the results were not specific to the analysis method we used, we ran a series of analyses with other methods and parameters. First, we ran the same set of analyses while changing the lasso-to-ridge ratio from 0.01 to an automatic selection in threefold nested cross-validation among 6 ratios (0.1, 0.5, 0.7, 0.9, 0.99, 1), to check whether the quality of parameter regularization would impact the results52,53,54,55,56. The results showed similar patterns; resting-state FC successfully predicted age and intelligence in younger adults but none of the other variables (Table 6).
The same set of analyses were also run again but using tenfold cross-validation instead of leave-one-out cross-validation, as k-fold cross-validation may show higher robustness than leave-one-out cross-validation57. Once again, age was successfully predicted. Likewise, intelligence in younger adults was predicted by resting-state FC. Nevertheless, none of the rest of variables were predicted by resting-state FC (Table 6).
Thirdly, the analyses were run using a Random Forest Regressor instead of a linear regression. Leave-one-out cross-validation was used, allowing for a maximum depth among 5 values (5, 10, 20, 40, 50), similar to parameters used in previous studies58,59,60,61. The results were again similar, showing good prediction for age and intelligence in younger adults, but not the other variables (Table 6).
Fourthly, we also changed the edge filtering threshold. Across all the analyses described so far, we applied the edge filtering threshold of 0.01 to include only edges that correlated with the predicted behavioural measure with p value < 0.01. To test the effects of this filtering threshold, we ran exploratory analyses using filtering thresholds of 0.02, 0.03, 0.04 and 0.05 for intelligence in younger adults, a behavioural domain where we saw one of the strongest prediction results (Supplementary Table 2). This exploratory analysis showed that intelligence in younger adults was best predicted by a filtering threshold of p = 0.05, resulting in r = 0.37, R-square = 0.12, and nRMSD = 0.94. Based on this analysis, the main analyses (collapsed across age groups) were performed again for all prediction models using this new filtering threshold (p = 0.05). However, the results were also similar; the resting-state FC predicted age but not emotional memory measures nor intelligence (Table 6).
Including all edges
As described earlier, we excluded data from seven nodes given that 45 participants did not have data from one of these nodes. The excluded nodes are located in the left and right temporal lobes. To ensure that the results are not affected by our exclusion of these nodes, the main analyses were conducted again using the preregistered methods (edges filtering at p = 0.01, Elastic Net, L1 = 0.01, leave one-out cross validation) while including all the edges; this resulted in 258 participants without the 45 participants who had missing data in these edges (N = 13 younger adults; N = 15 middle-aged adults; N = 17 older adults). Once again, the analyses showed that only age was significantly predicted by resting-state FC (Table 7).
In this study, we examined whether resting-state FC predicts individual differences in the emotional enhancement effect in memory, the positivity bias in memory, intelligence and age. Neither the emotional memory enhancement effect, nor the positivity bias was significantly predicted by resting-state FC. In contrast to these measures of emotional memory, models derived from resting-state FC successfully predicted chronologic age, replicating previous findings47,49. These results suggest that the methods used in this study were able to predict behavioural phenotypes based on resting-state FC. Yet, contrary to our prediction (preregistered), intelligence was not predicted from resting-state FC when participants of all ages were included.
To check whether the failure to predict intelligence or emotional memory measures using resting-state FC was due to the wide age range of participants, we split participants into three groups: younger (18–40 years), middle-aged (41–60 years), and older adults (61–87 years). Previous studies suggest that resting-state FC patterns undergo a nonlinear trajectory with age, such as increasing FC within DMN during late adulthood before its rapid decline after age 7429. In addition, age-related compensatory recruitment of the prefrontal cortex can result in age-related shifts in brain regions responsible for tasks relevant to intelligence50. However, neither the emotional memory enhancement effect nor the positivity bias was predicted by the models in any age groups. The only exception was intelligence in younger adults; when including only younger adults as done in past studies35,36, resting-state FC successfully predicted individual differences in intelligence.
Importantly, even after splitting participants into three age groups, resting-state FC did not predict the two emotional memory measures in any age groups. These results suggest that the predictive power of resting-state FC is lower for emotional memory measures than for intelligence. The results could also suggest that the utilized emotional memory measures are not appropriate or reflective of a reliable effect. These results are in line with those from past studies on resting-state FC. For example, a recent study failed to replicate past findings in predicting habitual use of emotion regulation strategies from resting-state FC62. Another study showed that resting-state FC predicts working memory, but not executive control, language, or verbal memory performance in older adults63. Similarly, in Dubois et al.36, resting-state FC predicted intelligence in younger adults, but not personality traits of neuroticism, consciousness, extraversion, and agreeableness. There are several possible reasons behind the weaker predictive power of resting-state FC for our emotional memory measures.
The first possibility might be a low reliability of emotional memory measures. A recent study reported that the emotional memory enhancement effect had a very low test–retest reliability when the same participants were tested twice over a delay of 10 weeks42 presumably due to the correlation between emotional and neutral memory measures and low between-subject variability in these subtraction scores41. In the Cam-CAN data, there were strong correlations between emotional and neutral memory measures; positive and negative object memory scores were highly correlated, r(301) = 0.90, p < 0.011; and neutral object memory performance was also highly correlated with both positive, r(301) = 0.89, p < 0.001, and negative memory performance, r(301) = 0.88, p < 0.001. The associative valence memory performance also showed high correlations between positive and negative conditions, r(301) = 0.77, p < 0.001, between neutral and negative conditions, r(301) = 0.85, p < 0.001, and between positive and neutral conditions, r(301) = 0.84, p < 0.001; although the magnitudes of correlation were weaker for the background memory, it still showed moderate correlations between negative and positive r(301) = 0.41, p < 0.001, between negative and neutral r(301) = 0.35, p < 0.001, and between positive and neutral conditions r(301) = 0.39, p < 0.001. Such strong correlations could have resulted in low reliability for our dependent variables (i.e., the emotional memory enhancement effect and the positivity bias score) that were derived by subtracting one from another highly correlated variable64. Thus, our failure to predict emotional memory measures may have been driven by the limited reliability of the measures.
The second possibility concerns our dependent measures. In our main analysis, we used memory performance for neutral objects learned with emotional backgrounds as the key dependent variable. Thus, the dependent measures were not about emotional items themselves but more about the effects of emotion (induced by the background images) on memory for nearby neutral information (i.e., neutral objects presented with the background images). Previous research has repeatedly shown that while emotional items are preferentially remembered better than neutral items in many situations, the effects of emotional items on nearby neutral information are more complex7,65,66; such that emotion sometimes enhances memory for nearby neutral information but sometimes impairs memory for nearby neutral information66,67,68,69. These findings point to the likelihood that resting-state FC has low prediction power for individual differences in memory for neutral items nearby emotional items (arguably due to the complex nature of the effects) but may be able to predict individual differences in memory for emotional items themselves. To address this possibility, we performed exploratory analyses on two additional memory measures that are more about emotional background images (i.e., valence and the content of a background image associated with each neutral object). However, once again, resting-state FC failed to predict individual differences in these two measures. Thus, resting-state FC does not seem to reliability predict individual differences in the effects of emotion on memory, irrespective of whether memory concerns emotional items per se or nearby neutral information.
Nevertheless, it is important to note that in the Cam-CAN project, the two measures of the emotional background images were not independent from the object memory measure; participants were given a chance to answer the valence and the content of a background image associated with each neutral object only when they recognized the neutral object as studied (see “Methods”). Thus, it is possible that resting-state FC can predict individual differences in emotional memory in other tasks (e.g., a simple recognition test; a free recall without constraints of associated object memory). Future research needs to address this issue.
Third and relatedly, the current study used performance in the memory test performed 10 min after the encoding session. However, previous research has suggested that the effects of emotion on memory are due to long-term consolidation effects19; thus future studies with long-term memory measures obtained after consolidation may obtain a different result. Furthermore, in the Cam-CAN project, the resting-state BOLD signals were obtained on a different day from the emotional memory task. Thus, the design could be particularly vulnerable to the low reliability of our emotional memory measures42. It is therefore possible that resting-state BOLD signals have stronger predictive power for emotional memory measures when they are obtained on the same day.
Fourth, recent research points out that FC derived from 5 to 10 min of resting-state data have low reliability to detect reliable individual differences70,71,72,73,74. Given that the Cam-CAN project has a relatively short resting-state data, the lack of significant effects in the present study may have been due to the low reliability of FC analysed in the present study. Future research needs to address this issue with data from longer resting-state scans. Likewise, recent research suggests that FC derived from task-state fMRI scans can enhance predictions of individual differences74. Therefore, future research could also combine using task-state functional connectivity and resting-state in order to achieve stronger predictability75.
Fifth, we had a relatively large sample size; in fact, our total sample size (n = 303) is sufficient to detect a relatively small sized correlation (see “Methods”). However, our sample size was modest after splitting participants into three age groups57,76, which could have resulted in the failure to predict emotional memory measures by resting-state FC. On the one hand, even with this same sample size, we still found that intelligence in younger adults was predicted by resting-state FC as observed in previous studies36,49,77,78,79. Yet, our sample size may not have been large enough to address the heterogeneity within older adults80. In addition, participants only had one resting-state session, which may have resulted in higher noise and lower prediction power than combining two or more sessions36. Future research needs to use a larger sample combined with multiple resting-state sessions and address the effects of resting-state FC.
Finally, although the main analyses investigated the effect across ages, it is notable that the preprocessing methods and predictive models used may be most appropriate for younger adults as the methods were developed and used primarily for a younger adult sample36. Unplanned exploratory analyses showed that intelligence was successfully predicted from resting-state FC in younger adults but not in middle-aged or older adults. Yet, as described earlier, age was predicted by resting-state FC successfully for participants from different age groups in this study. Such results for age were consistent with other findings49, suggesting that our analysis and denoising method was appropriate. The alternative reason behind the failure of predictions for middle-aged and older adults concerns the effects of age on individual differences. Previous longitudinal studies have suggested heterogeneity within older adults in their cognitive performance, brain structure and its functioning80,81. Thus, the age-related increases in the heterogeneity may have made it difficult for us to predict cognitive measures in middle-aged or older adults relative to younger adults. In line with this idea, a recent large-scale study including 711 older adults also found no association between cognitive performance and resting-state FC63 (but see Ref.82). Future research needs to take into account the effects of age on heterogeneity within participants.
In summary, the present study used a machine learning approach (which allowed us to select the most informative connections across the whole brain rather than relying on a priori selected regions) in predicting individual differences in emotional memory measures. While models derived from resting-state FC predicted age (for all participants) and intelligence for younger adults, they did not reliably predict the emotional memory enhancement effect and the positivity bias in memory for any age group. The results suggest the neural basis of individual differences in the emotional memory enhancement effect and positivity bias may not be meaningful or large enough to be predicted from resting-state FC. The results are in line with recent findings on low-reliability of the emotional enhancement effects in memory42, suggesting that more research should be done on the viability of the emotional enhancement effect and positivity bias as stable traits. Our results also support the use of an existing pipeline36 to denoise and predict traits at least for adult participants. Future research would be able to use this pipeline to minimise bias in choosing methods based on the results obtained (p-hacking)76.
Data used in the preparation of this work were obtained from the CamCAN repository (available at http://www.mrc-cbu.cam.ac.uk/datasets/camcan/)44,45. A total of 306 participants, aged 18–87, have completed the structural MRI brain scans, resting-state fMRI scans, the emotional memory test and the intelligence test in the Cam-CAN dataset. Two participants were completely missing signal in significant portions of the cerebellum and the brain stem leading to errors in preprocessing. One participant had an incomplete resting-state fMRI scan lasting less than the database’s acquisition time of 8 min and 40 s. Therefore, the final sample size included 303 participants (N = 155 females; 18–87 years, Mage = 54.3, SD = 18.1) who had structural and functional resting-state brain scans, behavioural measures on emotional memory, and intelligence scores. The data analysed in this study was the Cam-CAN consortium which has gained ethical approval from the Cambridgeshire 2 (now East of England-Cambridge Central) Ethics Committee. We did not perform a formal power analysis; The power computation for prediction R2 is not established because (1) there is no single true data generation model corresponding to a specific R2 value and (2) the true data generation model also varies depending on the algorithm (e.g., random forest, elastic net). But the sensitivity analysis suggests that our total sample size (n = 303) is sufficient to detect a relatively small sized correlation (r) of 0.16 at 80% statistical power with alpha = 0.0583. After splitting the sample into three age groups, the sample sizes were sufficient to detect a correlation (r) of 0.30 in younger adults, (r) of 0.28 in middle-aged adults, and (r) of 0.25 in older adults, with 80% power and alpha = 0.05.
Emotional memory task
The memory task in the Cam-CAN database consisted of 120 trials, presented in two blocks46. In brief, every trial started with presentation of a background image for 2.5 s; the background was either positive, negative or neutral. Participants then saw a neutral object superimposed on the background for 7.5 s, during which they were asked to link the item and background by mentally creating a story that combines them. Participants performed a surprise memory test 10 min later.
During the memory test, participants were shown an object and asked to indicate whether or not it had been shown during the study phase (i.e., object memory). For objects indicated as ‘shown’, participants were asked to identify the valence of the background on which the object was superimposed (i.e., associative valence memory), then describe the background scene (i.e., background memory). Participants’ responses to the background memory test were coded to reflect whether participants described correct details, correct gist, incorrect information or no responses were given. The test had 160 trials (120 trials with old stimuli and 40 trials with new stimuli).
The current study used the d’ measure of discriminability84 for the object and the associative valence memory. For the background memory, we computed the proportion of trials where participants could correctly recalled gist. For all memory scores, two memory variables were created: the emotional enhancement effect and the positivity bias. The emotional enhancement effect was obtained by subtracting performance in the neutral condition from the average performance in the positive and negative conditions. The positivity bias measure was computed by subtracting performance in the negative condition from performance in the positive condition.
The Cam-CAN database included a fluid intelligence test, the Cattell Culture Fair Scale 2 Form A85. The test has four subsets of nonverbal intelligence tests. A principal component analysis was performed on the scores of the four subsets to get one composite score of intelligence.
MRI data acquisition
MRI scans were acquired using 3 T Siemens TIM Trio scanner45. Structural T1-weighted images were acquire using the 3D MPRAGE sequence: repetition time (TR) = 2250 ms, echo time (TE) = 2.99 ms, Inversion Time (TI) = 900 ms, flip angle = 9 degrees, GRAPPA acceleration factor = 2, resolution 1.0 mm isotropic. Every participant had one resting-state fMRI scan with an acquisition time of 8 min and 40 s, and a total of 261 volumes. Resting-state BOLD fMRI scans had the following parameters: TR = 1970 ms; TE = 30 ms; flip angle = 78 degrees; slices = 32 of thickness = 23.7 mm; field of view (FOV) = 192 mmx 192 mm; voxel size = 3 mm × 3 mm × 4.44 mm.
We initially processed the raw functional MRI (fMRI) data obtained from the CamCAN database using FMRIB Software Library (FSL)86. Preprocessing included deleting the first two volumes in every scan. Motion correction was then performed on the raw resting-state images using FSL MCFLIRT87 (6 degrees of freedom), where the timeseries were realigned to the middle volume. Three participants (aged 23, 38 and 40) showed translational movement of over 3 mm in one or more volume. We did not exclude participants based on motion cut-off. Field map distortion correction was then applied, before setting high pass filtering cut-off to 100 s, and performing nonlinear registration of brain-extracted T1 images to Montreal Neurological Institute (MNI) space using FSL FNIRT (12 degrees of freedom). Each participant’s T1 structural image was skull/neck stripped using the FSL’s brain extraction tool (BET) and then used to create participant’s specific masks for the white matter, grey matter and cerebrospinal fluid (CSF) using FSL FAST. Although Dubois et al.36 found stronger prediction results when using multimodal surface-based alignment and registration (MSM) compared with MNI, we refrained from using MSM as it excludes subcortical regions, which are relevant for emotional memory16.
We next applied the same denoising steps as included in ‘Pipeline A’ from Dubois et al.36 given that this pipeline had the best prediction performance in predicting personality traits in this study. The pipeline started by z-score normalization of each voxel’s signals. Voxels in the white matter and CSF were then detrended by regressing out the temporal drifts. Next, the mean signals of CSF and white matter voxels were computed and regressed out from grey matter voxels. Motion regression was then performed using translational and rotational and temporal parameters as covariates which were regressed out from the whole-brain through linear regression. Low-pass filtering was performed using a Gaussian kernel with standard deviation of 1 TR. Finally, grey matter voxels were detrended for temporal drifts, followed by a global signal regression. The preprocessing and denoising pipeline scripts used are publicly available (https://github.com/donakand/EmotionalMemory).
The denoised resting-state images were then segmented into 268 nodes48; for each node, we averaged signals in all included voxels for each timepoint to create timeseries for each parcel. A total of 45 participants had missing data in one or more brain nodes; these missing data were restricted to seven nodes: 51, 58, 60, 185, 189, 194 and 202, corresponded to the left and right temporal lobes, located close to the surfaces of the brain48. To keep as many participants as possible, these seven nodes were excluded from the analysis. A connectivity matrix was created by correlating parcels’ time-series using Pearson’s correlation. The connectivity matrix consisted of 33,930 edges (connections) per participant.
Machine learning analyses
Our main machine learning analyses used methods described in Dubois et al.36 Four separate analyses were carried out for different outcome variables: (a) the emotional enhancement effect, (b) the positivity bias, (c) intelligence and (d) age (as exploratory analysis that we did not pre-register). In all models, we included the connectivity from 33,930 edges as predictors. In addition, age, gender, handedness, and intelligence were used as control variables in the emotional enhancement effect and positivity bias analyses. Age, gender and handedness were controlled in the analysis of intelligence. Gender, handedness and intelligence were controlled for in the analysis of age. Similar to Dubois et al.36, the effects of these control variables were regressed out from the outcome measure using multiple linear regression before running a subsequent machine learning analysis. One participant was missing handedness information. The handedness for this participant was replaced by the median handedness value. Motion parameters were not used as control variables in the machine analyses, as motion correction and motion regression were applied to the resting-state scans during preprocessing and denoising (see Supplementary Table 1 for analyses including motion as a control variable). Motion was calculated as the mean translational realignment value. Motion was weakly correlated with age (r = 0.12, p = 0.03) and intelligence (r = − 0.18, p = 0.002), but not significantly correlated with the emotional enhancement effect (r = − 0.04, p = 0.45), positivity effect (r = − 0.05, p = 0.41), or intelligence in younger adults (r = − 0.20, p = 0.06).
For each machine learning analysis, a filtering approach was taken. Only the edges with correlations with the predicted variables with p value < 0.01 were included from the analyses. Next, we generated an Elastic Net model, implemented with Scikit Learn in python version 0.19.288. Similar to Dubois et al.36 to choose optimal parameters, the model’s alpha value was tuned using a grid search of the parameter space, and a three-fold nested cross-validation. The Elastic Net mixing parameter L1 was set as 0.01. However, we also performed exploratory analyses where L1 was chosen through three-fold nested cross-validation. The model was trained and tested using a leave-one-out cross-validation. The model was evaluated using R2. As in Dubois et al.36, in the current study R2 was not the square of the correlation coefficient, but rather determined using Eq. (1). Therefore, R2 could take negative values in case of squared sum of errors larger than that of the null model—a horizontal line through the mean. As in Dubois et al.36, models were also evaluated on the normalized root mean squared deviation (nRMDS), which is the square root of the ratio of the standard deviation of residuals divided by the standard deviation of the observed values, and can be obtained directly from the R2, as in Eq. (2).
To evaluate these results against a null hypothesis, under which the data is not predictive of our variables, and obtain a p value, we generated a null distribution by generating 1000 shuffled permutations of the memory scores in the dataset. We ran our models on every shuffled dataset. The one-tailed p value of the (actual) dataset model was then calculated by computing the number of permutations for which R2 was larger than the dataset’s model, divided by 1000.
To assert the specificity of the results obtained from our preregistered methodology, we conducted further analyses using alternative methods. The exploratory analyses retained the same control variables as our main analyses and were also evaluated using R2. The main analyses were run again while only changing the cross-validation from leave-one-out to 10-Fold cross-validation. This is due to a recent argument suggesting more robustness with k-Fold cross-validation in neuroimaging57. In keeping with the original methodology by Dubois et al.36, we set up the Elastic Net model using a proportion of L1 regularization of 0.01. However, other studies have obtained good prediction using larger L1 ratios52,53, or pure Lasso54,55,56 (L1 = 1). Therefore, we ran the main analyses again, tuning the L1 parameter in a nested threefold cross-validation from values: 0.1, 0.5, 0.7, 0.9, 0.99, 1. The filtering threshold applied before all prediction analyses, was set at 0.01. To check whether the threshold impacted the findings, exploratory analyses were run to predict intelligence for younger adults using filtering thresholds of 0.02, 0.03, 0.04, and 0.05. The highest R-square was obtained using a filtering threshold of 0.05 (Supplementary Table 2). The main analyses were rerun using a filtering threshold of 0.05. Finally, the main analyses were replicated using a different learning algorithm, Random Forest, rather than regularized linear regression. Random Forest is one of the most robust algorithms and has been used in psychology89,90,91. The algorithm does not require assumptions of linearity or collinearity of variables, and has shown good reliability92,93,94. The maximum depth parameter was chosen from values: 5, 10, 20, 40, 50 using a nested threefold cross-validation. Several past studies have successfully used similar parameters to implement Random Forest models in psychological and neuroimaging studies58,59,60,61.
Further analyses investigated whether the null findings were influenced by the exclusion of seven nodes which had missing data from the analyses. The main analyses were rerun while including all the nodes, resulting in 35,778 edges, for the 258 participants who did not have missing data in any node. The analysis followed the main analysis methods of filtering threshold at 0.01, Elastic Net mode with L1 = 0.01, and training the model through leave one-out cross validation.
Data used in the preparation of this work were obtained from the CamCAN repository (available at http://www.mrc-cbu.cam.ac.uk/datasets/camcan/). Users must agree to the terms and conditions and submit an application to access the CamCAN database. The dataset generated and analysed during the current study is available in the OSF repository, https://osf.io/bm98y/files/osfstorage.
Schaefer, A. & Philippot, P. Selective effects of emotion on the phenomenal characteristics of autobiographical memories. Memory 13, 148–160 (2005).
StJacques, P. L. & Levine, B. Ageing and autobiographical memory for emotional and neutral events. Memory 15, 129–144 (2007).
Talarico, J. M. & Rubin, D. C. Flashbulb memories are special after all; In phenomenology, not accuracy. Appl. Cogn. Psychol. Off. J. Soc. Appl. Res. Mem. Cogn. 21, 557–578 (2007).
Kensinger, E. A. Remembering the details: Effects of emotion. Emot. Rev. 1, 99–113 (2009).
Kensinger, E. A. & Corkin, S. Memory enhancement for emotional words: Are emotional words more vividly remembered than neutral words?. Mem. Cogn. 31, 1169–1180 (2003).
LaBar, K. S. & Cabeza, R. Cognitive neuroscience of emotional memory. Nat. Rev. Neurosci. 7, 54–64 (2006).
Mather, M. Emotional arousal and memory binding: An object-based framework. Perspect. Psychol. Sci. 2, 33–52 (2007).
Yonelinas, A. P. & Ritchey, M. The slow forgetting of emotional episodic memories: An emotional binding account. Trends Cogn. Sci. 19, 259–267 (2015).
Talarico, J. M., Labar, K. S. & Rubin, D. C. Emotional intensity predicts autobiographical memory experience. Mem. Cogn. 32, 1118–1132 (2004).
Ferree, N. K. & Cahill, L. Post-event spontaneous intrusive recollections and strength of memory for emotional events in men and women. Conscious. Cogn. 18, 126–134 (2009).
Duyser, F. A. et al. Negative memory bias as a transdiagnostic cognitive marker for depression symptom severity. J. Affect. Disord. 274, 1165–1172 (2020).
Gerritsen, L. et al. Amygdala to hippocampal volume ratio is associated with negative memory bias in healthy subjects. Psychol. Med. 42, 335–343 (2012).
MacQueen, G. & Frodl, T. The hippocampus in major depression: Evidence for the convergence of the bench and bedside in psychiatric research. Mol. Psychiatry 16, 252–264 (2011).
Mather, M. & Carstensen, L. L. Aging and motivated cognition: The positivity effect in attention and memory. Trends Cogn. Sci. 9, 496–502 (2005).
Brassen, S., Gamer, M. & Bchel, C. Anterior cingulate activation is related to a positivity bias and emotional stability in successful aging. Biol. Psychiatry 70, 131–137 (2011).
Murty, V. P., Ritchey, M., Adcock, R. A. & LaBar, K. S. FMRI studies of successful emotional memory encoding: A quantitative meta-analysis. Neuropsychologia 48, 3459–3469 (2010).
Dahlgren, K., Ferris, C. & Hamann, S. Neural correlates of successful emotional episodic encoding and retrieval: An SDM meta-analysis of neuroimaging studies. Neuropsychologia 143, 107495 (2020).
Dolcos, F., LaBar, K. S. & Cabeza, R. Interaction between the amygdala and the medial temporal lobe memory system predicts better memory for emotional events. Neuron 42, 855–863 (2004).
Ritchey, M., Dolcos, F. & Cabeza, R. Role of amygdala connectivity in the persistence of emotional memories over time: An event-related fMRI investigation. Cereb. Cortex 18, 2494–2504 (2008).
Murty, V. P. et al. Age-related alterations in simple declarative memory and the effect of negative stimulus valence. J. Cogn. Neurosci. 21, 1920–1933 (2009).
St. Jacques, P. L., Dolcos, F. & Cabeza, R. Effects of aging on functional connectivity of the amygdala for subsequent memory of negative pictures: A network analysis of functional magnetic resonance imaging data. Psychol. Sci. 20, 74–84 (2009).
Sakaki, M., Nga, L. & Mather, M. Amygdala functional connectivity with medial prefrontal cortex at rest predicts the positivity effect in older adults’ memory. J. Cogn. Neurosci. 25, 1206–1224 (2013).
de Voogd, L. D., Klumpers, F., Fernández, G. & Hermans, E. J. Intrinsic functional connectivity between amygdala and hippocampus during rest predicts enhanced memory under stress. Psychoneuroendocrinology 75, 192–202 (2017).
Kark, S. M. & Kensinger, E. A. Post-encoding amygdala-visuosensory coupling is associated with negative memory bias in healthy young adults. J. Neurosci. 39, 3130–3143 (2019).
Suárez, L. E., Markello, R. D., Betzel, R. F. & Misic, B. Linking structure and function in macroscale brain networks. Trends Cogn. Sci. 24, 302–315 (2020).
Tavor, I. et al. Task-free MRI predicts individual differences in brain activity during task performance. Science 352, 216–220 (2016).
Salami, A., Pudas, S. & Nyberg, L. Elevated hippocampal resting-state connectivity underlies deficient neurocognitive function in aging. Proc. Natl. Acad. Sci. 111, 17654–17659 (2014).
Salami, A., Wahlin, A., Kaboodvand, N., Lundquist, A. & Nyberg, L. Longitudinal evidence for dissociation of anterior and posterior MTL resting-state connectivity in aging: Links to perfusion and memory. Cereb. Cortex 26, 3953–3963 (2016).
Staffaroni, A. M. et al. The longitudinal trajectory of default mode network connectivity in healthy older adults varies as a function of processing speed. J. Neurosci. 38, 2809–2817 (2018).
Fjell, A. M. et al. Brain events underlying episodic memory changes in aging: A longitudinal investigation of structural and functional connectivity. Cereb. Cortex 26, 1272–1286 (2016).
Bleidorn, W. & Hopwood, C. J. Using machine learning to advance personality assessment and theory. Personal. Soc. Psychol. Rev. 23, 190–203 (2019).
Rosenberg, M. D. et al. A neuromarker of sustained attention from whole-brain functional connectivity. Nat. Neurosci. 19, 165–171 (2015).
Fong, A. H. C. et al. Dynamic functional connectivity during task performance and rest predicts individual differences in attention across studies. Neuroimage 188, 14–25 (2019).
Barnes, K. A., Anderson, K. M., Plitt, M. & Martin, A. Individual differences in intrinsic brain connectivity predict decision strategy. J. Neurophysiol. 112, 1838–1848 (2014).
Finn, E. S. et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat. Neurosci. 18, 1664 (2015).
Dubois, J., Galdi, P., Han, Y., Paul, L. K. & Adolphs, R. Resting-state functional brain connectivity best predicts the personality dimension of openness to experience. Personal. Neurosci. 1, 1–21 (2018).
Wu, J., Srinivasan, R., Kaur, A. & Cramer, S. C. Resting-state cortical connectivity predicts motor skill acquisition. Neuroimage 91, 84–90 (2014).
Laird, A. R. et al. Behavioral interpretations of intrinsic connectivity networks. J. Cogn. Neurosci. 23, 4022–4037 (2011).
Smith, S. M. et al. Correspondence of the brain’s functional architecture during activation and rest. Proc. Natl. Acad. Sci. U. S. A. 106, 13040–13045 (2009).
Andreano, J. M., Touroutoglou, A., Dickerson, B. C. & Barrett, L. F. Resting connectivity between salience nodes predicts recognition memory. Soc. Cogn. Affect. Neurosci. 12, 948–955 (2017).
Hedge, C., Powell, G. & Sumner, P. The reliability paradox: Why robust cognitive tasks do not produce reliable individual differences. Behav. Res. Methods 50, 1166–1186 (2018).
Schümann, D., Joue, G., Jordan, P., Bayer, J. & Sommer, T. Test–retest reliability of the emotional enhancement of memory. Memory 28, 49–59 (2020).
Gratton, C. et al. Functional brain networks are dominated by stable group and individual factors, not cognitive or daily variation. Neuron 98, 439-452.e5 (2018).
Taylor, J. R. et al. The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository: Structural and functional MRI, MEG, and cognitive data from a cross-sectional adult lifespan sample. Neuroimage 144, 262–269 (2017).
Shafto, M. A. et al. The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) study protocol: A examination of healthy cognitive ageing. BMC Neurol. https://doi.org/10.1186/s12883-014-0204-1 (2014).
Henson, R. N. et al. Multiple determinants of lifespan memory differences. Sci. Rep. 6, 32527 (2016).
Vergun, S. et al. Characterizing functional connectivity differences in aging adults using machine learning on resting state fMRI data. Front. Comput. Neurosci. 7, 38 (2013).
Shen, X., Tokoglu, F., Papademetris, X. & Constable, R. T. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. Neuroimage 82, 403–415 (2013).
Tsvetanov, K. A. et al. Extrinsic and intrinsic brain network connectivity maintains cognition across the lifespan despite accelerated decay of regional brain activation. J. Neurosci. 36, 3115–3126 (2016).
Grady, C. The cognitive neuroscience of ageing. Nat. Rev. Neurosci. 13, 491–505 (2012).
Cabeza, R., Anderson, N. D., Locantore, J. K. & McIntosh, A. R. Aging gracefully: Compensatory brain activity in high-performing older adults. Neuroimage 17, 1394–1402 (2002).
Liu, Y. S., Chokka, S., Cao, B. & Chokka, P. R. Screening for bipolar disorder in a tertiary mental health centre using EarlyDetect: A machine learning-based pilot study. J. Affect. Disord. Rep. 6, 100215 (2021).
Veretennikova, M. A., Sikorskii, A. & Boivin, M. J. Parameters of stochastic models for electroencephalogram data as biomarkers for child’s neurodevelopment after cerebral malaria. J. Stat. Distrib. Appl. 5, 1–12 (2018).
Fishburn, F. A. et al. Irritability uniquely predicts prefrontal cortex activation during preschool inhibitory control among all temperament domains: A LASSO approach. Neuroimage 184, 68–77 (2019).
Shimizu, Y. et al. Toward probabilistic diagnosis and understanding of depression based on functional MRI data analysis with logistic group LASSO. PLoS ONE 10, e0123524 (2015).
Toiviainen, P., Alluri, V., Brattico, E., Wallentin, M. & Vuust, P. Capturing the musical brain with Lasso: Dynamic decoding of musical features from fMRI data. Neuroimage 88, 170–180 (2014).
Poldrack, R. A., Huckins, G. & Varoquaux, G. Establishment of best practices for evidence for prediction: A review. JAMA Psychiat. 77, 534–540 (2020).
Kim, J., Zhou, Y., Schiavon, S., Raftery, P. & Brager, G. Personal comfort models: Predicting individuals’ thermal preference using occupant heating and cooling behavior and machine learning. Build. Environ. 129, 96–106 (2018).
Han, Y. & Adolphs, R. Estimating the heritability of psychological measures in the Human Connectome Project dataset. PLoS One 15, e0235860 (2020).
Dadi, K. et al. Beyond brain age: Empirically-derived proxy measures of mental health. bioRxiv 1–34 (2020).
Pereira, V., Tavares, F., Mihaylova, P., Mladenov, V. & Georgieva, P. Factor analysis for finding invariant neural descriptors of human emotions. Complexity 2018, 1–8 (2018).
Dörfel, D., Gärtner, A. & Scheffel, C. Resting state cortico-limbic functional connectivity and dispositional use of emotion regulation strategies: A replication and extension study. Front. Behav. Neurosci. 14, 128 (2020).
Jockwitz, C. et al. Influence of age and cognitive performance on resting-state brain networks of older adults in a population-based cohort. Cortex 89, 28–44 (2017).
Infantolino, Z. P., Luking, K. R., Sauder, C. L., Curtin, J. J. & Hajcak, G. Robust is not necessarily reliable: From within-subjects fMRI contrasts to between-subjects comparisons. Neuroimage 173, 146–152 (2018).
Chiu, Y. C., Dolcos, F., Gonsalves, B. D. & Cohen, N. J. On opposing effects of emotion on contextual or relational memory. Front. Psychol. 4, 103 (2013).
Mather, M. & Sutherland, M. R. Arousal-biased competition in perception and memory. Perspect. Psychol. Sci. 6, 114–133 (2011).
Murray, B. D. & Kensinger, E. A. A review of the neural and behavioral consequences for unitizing emotional and neutral information. Front. Behav. Neurosci. 7, 42 (2013).
Sakaki, M., Fryer, K. & Mather, M. Emotion strengthens high-priority memory traces but weakens low-priority memory traces. Psychol. Sci. 25, 387–395 (2014).
Waring, J. D. & Kensinger, E. A. Effects of emotional valence and arousal upon memory trade-offs with aging. Psychol. Aging 24, 412–422 (2009).
Gordon, E. M. et al. Precision functional mapping of individual human brains. Neuron 95, 791-807.e7 (2017).
Anderson, J. S., Ferguson, M. A., Lopez-Larson, M. & Yurgelun-Todd, D. Reproducibility of single-subject functional connectivity measurements. Am. J. Neuroradiol. 32, 548–555 (2011).
Laumann, T. O. et al. Functional system and areal organization of a highly sampled individual human brain. Neuron 87, 657–670 (2015).
Noble, S., Scheinost, D. & Constable, R. T. A decade of test-retest reliability of functional connectivity: A systematic review and meta-analysis. Neuroimage 203, 116157 (2019).
Marek, S. et al. Reproducible brain-wide association studies require thousands of individuals. Nature 603, 654–660 (2022).
Elliott, M. L. et al. General functional connectivity: Shared features of resting-state and task fMRI drive reliable and heritable individual differences in functional brain networks. Neuroimage 189, 516–532 (2019).
Yarkoni, T. & Westfall, J. Choosing prediction over explanation in psychology: Lessons from machine learning. Perspect. Psychol. Sci. 12, 1100–1122 (2017).
Hearne, L. J., Mattingley, J. B. & Cocchi, L. Functional brain networks related to individual differences in human intelligence at rest. Sci. Rep. 6, 1–8 (2016).
Jiang, R. et al. Task-induced brain connectivity promotes the detection of individual differences in brain-behavior relationships. Neuroimage 207, 116370 (2020).
Basten, U., Hilger, K. & Fiebach, C. J. Where smart brains are different: A quantitative meta-analysis of functional and structural brain imaging studies on intelligence. Intelligence 51, 10–27 (2015).
Eavani, H. et al. Heterogeneity of structural and functional imaging patterns of advanced brain aging revealed via machine learning methods. Neurobiol. Aging 71, 41–50 (2018).
Olaya, B., Bobak, M., Haro, J. M. & Demakakos, P. Trajectories of verbal episodic memory in middle-aged and older adults: Evidence from the english longitudinal study of ageing. J. Am. Geriatr. Soc. 65, 1274–1281 (2017).
Greene, A. S., Gao, S., Scheinost, D. & Constable, R. T. Task-induced brain state manipulation improves prediction of individual traits. Nat. Commun. 9, 1–13 (2018).
Faul, F., Erdfelder, E., Buchner, A. & Lang, A.-G. Statistical power analyses using G*Power 3.1: Tests for correlation and regression analyses. Behav. Res. Methods 41, 1149–1160 (2009).
Green, D. M. & Swets, J. A. Signal Detection Theory and Psychophysics. 1, (Wiley, 1966).
Cattell, R. B. Abilities: Their structure, growth, and action. (Houghton Mifflin, 1971).
Smith, S. M. et al. Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 23, S208–S219 (2004).
Glasser, M. F. et al. The minimal preprocessing pipelines for the Human Connectome Project. Neuroimage 80, 105–124 (2013).
Pedregosa, F. et al. Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011).
Parikh, M. N., Li, H. & He, L. Enhancing diagnosis of autism with optimized machine learning models and personal characteristic data. Front. Comput. Neurosci. 13, 9 (2019).
Priya, A., Garg, S. & Tigga, N. P. Predicting anxiety, depression and stress in modern life using machine learning algorithms. Proc. Comput. Sci. 167, 1258–1267 (2020).
Haque, U. M., Kabir, E. & Khanam, R. Detection of child depression using machine learning methods. PLoS One 16, e0261131 (2021).
Dimitriadis, S. I. & Liparas, D. How random is the random forest? Random forest algorithm on the service of structural imaging biomarkers for Alzheimer’s disease: From Alzheimer’s disease neuroimaging initiative (ADNI) database. Neural Regen. Res. 13, 962–970 (2018).
Geurts, P., Irrthum, A. & Wehenkel, L. Supervised learning with decision tree-based methods in computational and systems biology. Mol. Biosyst. 5, 1593–1605 (2009).
Golino, H. F. & Gomes, C. M. A. Random forest as an imputation method for education and psychology research: Its impact on item fit and difficulty of the Rasch model. Int. J. Res. Method Educ. 39, 401–421 (2016).
Data collection and sharing for this project was provided by the Cambridge Centre for Ageing and Neuroscience (CamCAN). CamCAN funding was provided by the UK Biotechnology and Biological Sciences Research Council (Grant number BB/H008217/1), together with support from the UK Medical Research Council and University of Cambridge, UK. This work was supported by grants from the Japan Society for the Promotion of Science (JP16H05959; JH16H06406; JP16H02053), and Leverhulme Trust Research Leadership Award (awarded to Kou Murayama, grant number RL-2016-030). We thank Carolyn McNabb for proofreading the manuscript.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
Kandaleft, D., Murayama, K., Roesch, E. et al. Resting-state functional connectivity does not predict individual differences in the effects of emotion on memory. Sci Rep 12, 14481 (2022). https://doi.org/10.1038/s41598-022-18543-8