Predictive Models for Social Functioning in Healthy Young Adults: A Machine Learning Study Integrating Neuroanatomical, Cognitive, and Behavioral Data

Abstract

Poor social functioning is an emerging public health problem associated with physical and mental health consequences. Developing prognostic tools is critical to identify individuals at risk for poor social functioning and guide interventions. We aimed to inform prediction models of social functioning by evaluating models relying on bio-behavioral data using machine learning. With data from the Human Connectome Project Healthy Young Adult sample (age 22–35, N=1,101), we built Support Vector Regression models to estimate social functioning from variable sets of brain morphology to behavior with increasing complexity: 1) brain-only model, 2) brain-cognition model, 3) cognition-behavioral model, and 4) combined brain-cognition-behavioral model. Predictive accuracy of each model was assessed and the importance of individual variables for model performance was determined. The combined and cognition-behavioral models significantly predicted social functioning, whereas the brain-only and brain-cognition models did not. Negative affect, psychological wellbeing, extraversion, withdrawal, and cortical thickness of the rostral middle-frontal and superior-temporal regions were the most important predictors in the combined model. Results demonstrate that social functioning can be accurately predicted using machine learning methods. Behavioral markers may be more significant predictors of social functioning than brain measures for healthy young adults and may represent important leverage points for preventive intervention.

Introduction

Poor social functioning, characterized by an individual’s lack of social support and the inability to sustain the interpersonal relationships required for meaningful engagement in society, is an emerging public health problem. It is associated with a range of negative physical and mental health consequences,¹ such as early mortality^2,3 and increased risk for cardiovascular,⁴ mood,^5–7 anxiety,⁸ neurocognitive,^9,10 and psychotic disorders.^11,12 Conversely, individuals with adaptive social functioning have demonstrated better overall well-being^13,14 and improved self-management of chronic conditions.^15,16 Impairments in social functioning are also a hallmark of many psychiatric illnesses and could represent a transdiagnostic risk marker for mental health disorders.^17,18

Due to the broad negative consequences of poor social functioning, developing prognostic tools is critical to identify individuals at risk for social decline. Research on the development of such tools is in an early stage, in part because social functioning is a highly complex phenomenon with multiple contributors ranging from neurobiology to behavior (e.g. see citations^19–21). Brain imaging studies have linked structural and functional brain abnormalities to social functioning outcomes in both clinical and non-clinical samples. For example, in individuals at high risk for psychosis, less gray matter density in frontal and limbic areas has been associated with greater functional decline regardless of their transition to psychosis, supporting a transdiagnostic biomarker for social functioning deficits.¹⁸ These trends persist in chronic schizophrenia²² and are partially replicated in bipolar disorder with associations between gray matter volume reductions in the superior and medial prefrontal cortex and social functioning.²³ In non-clinical samples, loneliness and social isolation have been consistently associated with structural brain measures in the prefrontal cortex, insula, hippocampus, amygdala and posterior superior temporal cortex,²⁴ and to altered functional connectivity in distributed brain networks, including reward, fronto-limbic, and default mode networks.^24–28

Social functioning is also dependent on the cognitive and psychological processes necessary for adaptive social behavior. A range of cognitive and social cognitive impairments are linked to poor social functioning in schizophrenia,^29–31 bipolar disorder³² and depression.³³ In non-clinical samples, social cognition predicts overall social functioning³⁴ and supports varied social abilities such as self-regulation in conflict and empathetic engagement with others.^35,36 Additionally, a number of psychological traits such as depression, anxiety, neuroticism, negative affect, externalizing behaviors, and low psychological well-being have been linked to poor social functioning.^37–42

Machine learning methods may hold promise for achieving progress in improving social functioning in clinical and non-clinical populations through development of prognostic tools that could eventually be employed to identify individuals with high risk for social functioning deficits.^43,44 Machine learning methods allow for integrating large numbers of heterogeneous predictors across data modalities (e.g., neuroimaging, cognitive, psychological, and clinical data), avoiding the need to select a small number of predictors and to specify their relationships a priori, allowing for potential identification of pertinent interactions between variables and domains.^44,45 This can be beneficial for data domains that are inherently large, such as neuroimaging modalities, and can result in novel insights about the condition being studied, especially when associations between variables are not fully known (as is often in the case with brain-behavior relationships).^45,46 Given the numerous predictors of social functioning reviewed above, machine learning approaches may be well situated to parse the large, multi-modal predictor space to identify a broad range of predictors that could be leveraged in a predictive model and could be candidate treatment targets.⁴⁵ Additionally, in some cases, machine learning methods may facilitate predictions at the single-subject level due to the ability of some algorithms to uncover complex and potentially more flexible non-linear relationships amongst the data and the use of methods that promote generalization of the model to unseen data, such as robust cross-validation and model evaluation on independent datasets.^43,47

The use of machine learning to predict social functioning has been increasingly explored in clinical populations, most notably psychotic disorders, where initial studies have investigated prediction of functional outcomes from clinical data,^48–53 with smaller proof of concept studies relying on structural brain imaging data.^17,54 Early evidence in clinical high risk for psychosis and recent onset depression samples suggests that combining clinical and brain imaging measures may optimize model performance relative to models relying on only one predictor domain.¹⁷ However, most studies utilizing machine learning to predict social functioning outcomes have been limited by small sample sizes and models that consider the predictor modalities in isolation.^43,53 Further work is needed to understand how the varied predictors of functional outcomes can best be utilized to optimize accuracy and clinical usefulness of models and prognostic tools.

In the current study, we present initial findings to inform the development of prospective prognostic tools predicting social functioning using machine learning. We leveraged the Human Connectome Project Healthy Young Adult (HCP-YA) dataset - which includes multiple domains of cognitive, psychological, and other behavioral measures as well as high-quality neuroimaging data for a sample of over 1000 participants. Machine learning methods were used to evaluate prognostic models for individual social functioning scores, in order to delineate the relative performance of models built from neurobiological, cognitive, psychological and other behavioral domains. Secondly, we investigated which are the most influential predictors to identify potential risk markers and candidate treatment targets for social functioning.

Methods

Data Acquisition

This study utilized data from the HCP-YA sample. Access to the de-identified HCP-YA restricted access dataset was granted from the HCP consortium on June 19, 2019. The de-identified dataset was obtained from the secure HCP server. The University of Minnesota IRB deemed this research exempt and not human subjects research.

Participants

The HCP-YA study enrolled 1,206 participants aged 22–35, including twin and non-twin siblings from 2012 to 2015. Recruitment and study procedures took place at Washington University in St. Louis, Missouri and at the University of Minnesota in Minneapolis, Minnesota. A detailed report of the recruitment procedures has been documented elsewhere.⁵⁵ Inclusion criteria for HCP participants included age 22–35 and the ability to provide valid informed consent. Exclusion criteria included: history of psychiatric disorder, substance use disorder, neurological, endocrine or cardiovascular disease, genetic disorder, head injury, premature birth, history of chemotherapy or immunomodulatory agents, a score of ≤25 on the Folstein Mini-Mental State Exam,⁵⁶ claustrophobia, pregnancy or metal in the body. Participants were excluded from the current study if they did not have structural MRI data available or if they were missing greater than 50% of the cognitive and behavioral data.

Variables

Outcome measure

Social functioning.

We derived the social functioning outcome variable from the Social Relationships scales included in the NIH Emotion Toolbox (NIH-ETB).^57,58 An expert panel commissioned by the NIH Neuroscience Blueprint⁵⁹ led the creation of the NIH-ETB, a set of self-report measures to assess positive and negative aspects of emotional functioning. Based on item response theory, the NIH-ETB uses computer adaptive testing with extensive item banks and has undergone norming and validation. The scales representing the Social Relationships domain have demonstrated strong convergent validity with objective measures of interpersonal support, loneliness, and negative social interactions.⁵⁷

The social functioning composite used in this study was calculated using the formula provided by Babakhanyan et al.,⁶⁰ which includes scores for dimensions of friendship, loneliness, emotional support, instrumental support and perceived rejection. As a computer adaptive measure, the specific items an individual receives for these domains are individualized based on responses to previous items. Each item is assessed on a five-point Likert scale with options ranging from “never” to “always.” The friendship domain assesses perceived engagement with friends and quality of friendships (i.e., “I feel like I am part of a group of friends,” “I feel like I have lots of friends,” and “I get invited to go out and do things with other people”). The loneliness domain assesses perceived social inclusion and feelings of being alone (i.e., “I feel alone and apart from others” and “I feel left out”). The emotional support domain assesses perceived access to support for personal problems and emotions (i.e., “I have someone I trust to talk with about my problems,” and “I feel there are people I can talk to if I am upset”). The instrumental support domain assesses the perceived support network for completing tasks when needed, such as shopping, transportation, and accessing medical care. The perceived rejection domain assesses the frequency of negative social interactions (i.e., how often others “don’t listen when I ask for help,” and “act like they don’t care about me”). A comprehensive set of potential items for each domain is detailed elsewhere.⁵⁸ The summary score formula includes positive and negative factor loadings to reflect the positive (i.e., friendship) and negative (i.e., loneliness) aspects of social functioning.⁶⁰ The individual scales that comprise the composite measure in the HCP-YA dataset are provided as T-scores (mean of 50, standard deviation of 10). In a healthy sample, approximately 16% of individuals are expected to score within the potentially problematic range, defined as one standard deviation below the mean, on the individual scales and composite meausure.⁶⁰ Consitently, the distribution of social functioning scores in our sample follows a normal distribution with approximately 15% in the potentially problematic range.

Predictor variables by domain

Predictor variables are briefly described below; more detailed information for each variable collected by the HCP-YA study can be found online.⁶¹

Neurobiological predictors.

Cortical thickness and surface area for 33 cortical and 8 subcortical regions per hemisphere were obtained from structural MRI. All participants underwent MRI on 3T Siemens scanners using 32 channel head coils. A uniform MRI protocol and processing pipeline using FreeSurfer 5.1 was implemented between sites to control for variability in scanning procedures. Scanning protocols are detailed elsewhere,⁵⁵ and data quality has been ensured by extensive quality assurance procedures.⁵⁵ The automated segmentation of T1 and T2 weighted brain scans, which includes the standard parcellation maps using the Conte69 brain atlas was used. Surface areas were corrected for total intracranial volume.

Cognitive predictors.

Social Cognition.

Our primary measure of social cognition was the Penn Emotion Recognition Task,⁶² a measure of emotion processing which assesses the ability to discriminate emotions on human faces. We included both the correct responses and reaction time as separate variables. In addition, we included two in-scanner tasks: the HCP Theory of Mind task and the HCP Emotion Processing task.⁶³

Cognition.

Cognitive variables included global cognition as measured by the Mini-Mental State Exam⁵⁶ total score and discrete cognitive domains of attention, episodic memory, working memory, language, executive function and processing speed measured by the NIH-Toolbox Cognition Battery and scored as age-adjusted T-scores.

Behavioral predictors.

Psychological variables included the five domain scores from the Costa and McRae Neuroticism/Extroversion/Openness Five Factor Inventory (NEO-FFI)⁶⁴ and the age-adjusted T scores from the Achenbach Adult Self Report (ASR) Syndrome Scales⁶⁵ which includes subscales of anxious/depressed, withdrawn, somatic complaints, thought problems, attention problems, aggression, and rule breaking problems. Negative affect and psychological wellbeing were measured using the composite scores from these factors from the NIH-ETB scales.⁶⁰ Reward processing and self-regulation were measured by the Delay Discounting task.⁶⁶

Substance use variables included self-report measures of total number of alcoholic drinks in the last seven days, typical number of drinks per drinking day in the last 12 months, total times of use of any tobacco products in the last seven days, and total lifetime amount of cannabis use.

Physical functioning variables included the NIH Toolbox endurance, gait speed, dexterity and strength assessments,⁶⁷ the Pittsburg Sleep Quality Questionnaire⁶⁸ total score and Body Mass Index (BMI).

Sensory functioning was operationalized by four measures from the NIH Toolbox that assess sensory and neurological functioning: Words-in-Noise test (measure of audition), Odor Identification test, Regional Taste Intensity test, and the Pain Interference Survey. Vision measures included the Mars Contrast Sensitivity test.⁶⁹

Statistical Analysis

Missing data

Individual participants were excluded if they had incomplete neuroimaging data or were missing 50% or greater of the cognitive or behavioral data. The remaining missing values were imputed using K-Nearest Neighbors (k = 5). As an inclusion criterion for this study was complete structural MRI data, imputation only applied to cognitive and behavioral variables. For the training dataset, imputation was performed separately in each cross-validation fold to prevent information leakage that could lead to overfitting.

Multicollinearity evaluation

We evaluated multicollinearity (Pearson’s r > 0.70) by examining a correlation matrix of all variables. Two variables were highly correlated with NEO-FFI neuroticism: Negative affect (r = 0.72) and the ASR anxiety/depression scale (r = 0.71). Multicollinearity was primarily addressed by using the Recursive Feature Elimination step during model training. Secondarily, multicollinearity was addressed by estimating models with and without NEO-FFI neuroticism. No significant differences in model performance were found, and thus this variable was included in the final model.

Descriptive statistics

Descriptive statistics were calculated for social functioning, cognition, behavioral variables, and demographics, with means (SD) for continuous variables and percentages for categorical variables. All descriptive analyses were calculated in SPSS version 26.⁷⁰

Machine Learning Analysis

Model training and cross validation.

We used a linear support vector regression (SVR)⁷¹ to estimate individual participants’ social functioning score from brain morphological, cognitive, and/or behavioral data. We chose a regression, rather than a classification, approach because regression algorithms avoid the need to set arbitrary thresholds between groups (i.e., creating a social functioning cut-off score for good vs poor functioning), which can lead to loss of information and introduce bias into the model. Due to the NIH-ETB’s intended use in both healthy and clinical populations, it has been suggested that use of cut-points may be too restrictive and could result in overclassifying healthy individuals as having non-problematic functioning.⁵⁷ Social functioning is a complex phenomenon that occurs on a continuum; in the absence of gold-standard thresholds or cut-points for the NIH-ETB, a regression approach allows for predictions to be made on a continuous scale (i.e., predicting where an individual falls on the continuum of social functioning), which can increase model accuracy. ^47,72,73 The SVR is an established approach for uncovering non-linearities in the dataset that would not be uncovered by general linear models and is a robust approach for multi-modal data and large feature sets.

We trained four separate linear SVR models predicting social functioning composite scores: 1) a brain-only model, trained on structural MRI measures only; 2) a brain-cognition model, trained on brain and cognition measures only; 3) a cognition-behavioral model, trained on the cognition and behavioral measures only; and 4) a combined model, trained on the brain, cognition, and behavioral measures. The subsets of variables chosen for our models build in levels of biological to behavioral complexity, allowing for examination of which domains are necessary or sufficient for the prediction of social functioning. Additionally, as cognitive and behavioral measures are less invasive and expensive to obtain relative to brain measures, we tested these models separately because they are more practical for use in most clinical contexts. The machine learning analysis was implemented using the Scikit-learn Python toolbox⁷⁴ which uses the LIBLINEAR library of functions.⁷¹

The full dataset was randomly split into a training and testing dataset, with 10% of participants assigned to the testing dataset (N = 111). In the model training set, a nested five-fold cross validation (CV) procedure was used. To address potential bias due to heritability, samples were split into CV folds such that all related individuals are in the same fold, preventing leakage of information across folds. Prior to training, the data were standardized to have a mean of zero and standard deviation of one. We used Recursive Feature Elimination with the ridge regression estimator as a feature selection step to reduce the feature space. This step was embedded within the CV to determine the optimal number of variables to select for each model. The frequency that a specific variable was selected in the five CV folds was calculated.

The optimal values for the hyperparameters for the linear SVR, including the C-value and epsilon, were determined during the CV procedure using grid search. The epsilon insensitive, or L1, loss function was used. The parameter “dual” was set to false, which is the default when the number of samples is greater than the number of features.

Model testing and evaluation.

Using the hyperparameters learned during the testing and cross validation phase for each model, the learned SVR algorithm was applied to the held-out test dataset to evaluate model performance on unseen data. Model performance was measured based on the coefficient of determination, R², and the mean squared error (MSE) of the learned model. The R² value represents the proportion of variance of the outcome variable that can be explained by the independent variables in the model and is a measure of goodness of fit for the model. A model which always predicts the correct score would have an R² score of 1, and a constant model which always predicts the mean value of the outcome variable would have an R² score of zero. Thus, the R² score can be negative if it performs worse than a constant prediction of the mean. The MSE represents the mean of the squared differences between the actual outcome value and the model-predicted value (lower scores indicate better model performance).

We calculated 95% confidence intervals for the R² and MSE metrics using a bootstrapping procedure with 1000 resamples. Significance was evaluated using the 95% confidence interval of each model’s R² value. Performances of the four models were compared by calculating the 95% confidence interval for the differences in the bootstrapped R² results of two models at a time. P-values (two tailed, alpha = 0.05) were calculated by inverting the confidence interval.

Exploratory analysis of mis-predicted cases

To identify demographic groups or individual profiles for whom the model may not generalize well, we examined trends in the performance of the combined model by examining individual cases in the test dataset. The methods and results of this analysis are detailed in the Supplemental Methods and Results file.

Results

Sample

One hundred and five participants were excluded for missing MRI data or having greater than 50% missing behavioral data. The remaining 1,101 participants met eligibility criteria and were included in the analysis and split into the training (N = 990) and testing (N = 111) datasets. Demographics and social functioning score descriptive statistics for the full analysis sample are provided in Table 1. The social functioning composite score ranged from −13.0 to 15.2 with a mean of 5.6 and standard deviation of 5.0. Descriptive statistics for the individual scales that comprise the social functioning composite score, including the proportion of individuals fall into the potentially problematic range,⁵⁷ are provided in Supplementary Table 1. Descriptive statistics for all cognitive and behavioral variables are provided in Supplementary Table 2.

Table 1.

Demographics and Outcome Measure Descriptive Statistics (N = 1,101)

	Mean (SD) or %
Demographics
Age (years)	28.8 (3.7)
Gender: female	54.4
Education (years)	14.9 (1.8)
Race
American Indian or Alaskan Native	0.2
Asian, Native Hawaiian or Other Pacific Islander	5.7
Black or African American	14.9
White	74.9
More than one	2.5
Unknown or not reported	1.7
Ethnicity: Hispanic or Latino	8.6
Outcome Measure
NIH Emotion Toolbox Social Functioning Composite Score	5.6 (5.0)

Note: SD = standard deviation

Machine Learning Analyses

Model performance

Model performance statistics, including R² and MSE, for all models are provided in Table 2. The combined model and the cognition-behavioral model significantly predicted individual social functioning scores (R² =0.53, 95% CI [0.38, 0.62] for each model). The brain-only model and the brain-cognition model performances were not significant in predicting social functioning scores (R² = 0.06, 95% CI [−0.07, 0.16] and R² = 0.11 95% CI [−0.05, 0.23], respectively). These models performed significantly worse than the combined and cognition-behavioral models. No significant differences in model performance were noted between the brain-only and brain-cognition models or between the combined and cognition-behavioral models (Table 2).

Table 2.

Model Performance Summary Statistics and Model Comparison

Model	R2 (95% CI)	Mean Square Error (MSE) (95% CI)
Brain	0.06 (−0.07, 0.16)	21.36 (16.85, 26.2)
Brain-cognition	0.11 (−0.05, 0.23)	20.16 (15.64, 24.65)
Cognition-behavioral*	0.53 (0.38, 0.62)	10.73 (8.86, 12.73)
Combined brain-cognition-behavioral*	0.53 (0.38, 0.62)	10.68 (8.60, 12.88)
Model Comparison
Model	Point Estimate (95% CI)a	p-value
Brain vs. Brain-cognition	0.05 (−0.06, 0.16)	0.35
Brain vs. Cognition-behavioral	0.47 (0.32, 0.60)	<0.0001
Brain vs. Combined	0.47 (0.33, 0.60)	<0.0001
Brain-Cognition vs. Cognition-behavioral	0.41 (0.27, 0.56)	<0.0001
Brain-Cognition vs. Combined	0.42 (−0.26, 0.57)	<0.0001
Cognition-behavioral vs. Combined	−0.05 (−1.30, 1.20)	0.90

Note:

^*Statistically significant model based on 95% confidence interval

^aPoint estimate and 95% confidence interval represents differences in the bootstrapped R² values of the two models compared

Feature selection and importance

The selected features, or predictor variables, from the best performing models (combined and cognition-behavioral) are provided in Table 3, along with their feature coefficient and the frequency with which they were selected by the SVR algorithm during five-fold cross validation to be included in the model. The weight of the feature coefficient (absolute value) indicates the relative importance of the specific variable to the model. Features selected for the model at high frequencies during cross validation can be interpreted as being more reliable predictors relative to those selected infrequently. The highest weighted and most consistent features in the combined model included negative affect, psychological wellbeing, withdrawn symptoms, extraversion, and the cortical thickness in right and left rostral middle frontal gyri and the left superior temporal gyrus. The highest weighted and most consistent features in the cognition-behavioral model included negative affect, psychological wellbeing, withdrawn symptoms, extraversion, agreeableness, and aggression symptoms. A correlation table presenting the relationship between the variables in cognition-behavioral model and the social functioning composite are provided in Supplementary Table 3.

Table 3.

Feature Importance for Combined and Cognition-behavioral Models

Combined brain-cognitive-behavioral model
Feature	Feature Coefficient	CV Frequency (max = 5)
Negative affect	–1.53	5
Psychological well-being	1.37	5
ASR withdrawn symptoms	–0.82	5
NEO-FFI extraversion	0.76	5
R rostral middle frontal thickness	–0.56	4
L superior temporal thickness	0.49	5
L supramarginal thickness	–0.45	1
L rostral middle frontal thickness	0.42	3
L temporal pole area	–0.37	1
NEO-FFI agreeableness	0.32	1
R superior parietal area	0.31	1
R inferior parietal thickness	0.25	1
Cognition-behavioral model
Feature	Feature Coefficient	CV Frequency (max = 5)
Negative Affect	–1.71	5
Psychological Well-being	1.33	5
ASR Withdrawn symptoms	–0.85	5
NEO-FFI Extraversion	0.76	5
NEO-FFI Agreeableness	0.37	4
ASR Aggression symptoms	0.35	4
Education level	0.19	1

Note: ASR = Achenbach Adult Self Report scale; NEO-FFI = Neuroticism, Extraversion, Openness Five Factor Inventory; L = left; R = right

Discussion

In this study, we leveraged a large dataset of integrated biological, cognitive, and behavioral data and machine learning to predict social functioning in a large sample of healthy young adults. We found that only the combined and cognition-behavioral models significantly predicted social functioning scores, and these models performed equally well. That is, adding neuroanatomical data to the cognition-behavioral models did not enhance model performance. Further, the performance of the cognition-behavioral model was driven only by psychological variables, without significant influence from cognition. Our findings thus suggest that behavioral variables in the psychological domain that capture aspects of negative affect, psychological well-being, and personality traits, may be more robust predictors of social functioning than cognitive or neuroanatomical measures in healthy young adults, and may inform future development of prospective individual prediction models. While combining neuroanatomical measures with cognitive and behavioral measures did not improve model accuracy, potential neuroanatomical regions that may be important to the prediction of social functioning when considered along with behavioral variables were identified in the combined model. Below we highlight key findings in relation to existing literature.

Predictive models of social functioning from integrated bio-behavioral data

The combined brain-cognition-behavioral model significantly predicted social functioning, however it performed at an equal accuracy as the cognition-behavioral model, suggesting no additional benefit of including neuroanatomical measures. These models performed at similar accuracies when compared to existing machine learning research seeking to predict mental health outcomes,^75–77 and thus the more cost-effective models relying on behavioral data may be useful tools in identifying those at risk of social functioning deficits if applied to longitudinal data.

We found that the brain-only and brain-cognition models performed very poorly in the prediction of social functioning. This was unexpected given 1) previous studies in psychiatric populations that have successfully used support vector machines to predict functional outcomes from neuroanatomical data in first episode psychosis,¹⁷ early depression,¹⁷ and bipolar disorder ²³ and 2) research consistently linking cognitive and social cognitive impairments to social functioning in psychiatric disorders.^29,78–80 In healthy controls who are expected to have less variance in brain structure, however, these relationships may not be evident without much larger sample sizes to detect differences.⁸¹ The limited variance in cortical thickness and gray matter volumes in a healthy population likely contributed to the inability of the machine learning algorithm to discriminately use brain features in prediction. This is supported by Sartori et al,²³ who found that neuroanatomical data led to successful prediction of social functioning in bipolar disorder, but not in the healthy control sample. Likewise, recent large scale replicability studies linking personality features and other psychological variables to brain morphometry in healthy adults have failed to detect associations, which has been attributed to limited variance and small effect sizes even in samples greater than 1000 participants.^81–83 Limited variance and overall higher performance in cognitive functioning (i.e. ceiling effects) in our healthy population likely also contributed to the poor performance of the brain-cognition model.

Salient variables and data modalities for the prediction of social functioning

In the combined and cognition-behavioral models, the most predictive variables included negative affect, psychological well-being, withdrawn symptoms, extraversion and agreeableness personality traits, and aggression symptoms. This is not surprising, as these measures – especially psychological well-being, withdrawn symptoms and extraversion - have some conceptual overlap with perceived social functioning. These results are in line with a large body of work linking emotion, personality features, and psychological well-being to social functioning across healthy and clinical populations.^42,84 For example, negative affect is associated with the perceived quality and quantity of social relationships in multiple life stages, from childhood to older age.^84,85 Negative affect has also been identified as a strong predictor of social functioning in schizophrenia and bipolar disorder, even after accounting for cognitive deficits and other psychiatric symptoms.⁸⁶ Social psychology research shows that positive emotions may facilitate better social functioning through fostering and maintaining adaptive relationships and promoting positive evaluation of social interactions, while negative emotions relate to distancing from groups, negative bias in social attributions, and avoidant behavior.^87,88 Longitudinal studies taking a developmental perspective have also identified personality traits, particularly extraversion and neuroticism, as primary drivers of social functioning and well-being.^84,89 In turn, these personality traits are linked to positive and negative emotional experiences and overall psychological well-being.⁸⁴

Taken together with previous research, our results suggest that the domains of negative and positive affect, psychological well-being, and personality characteristics may be useful for identifying those at risk for poor functional outcomes and may point to potential intervention targets, which are amenable to psychological therapies,^90–93 to improve social outcomes and longer-term well-being. Future longitudinal work could investigate whether there are distinct developmental trajectories of social functioning based on individual characteristics of negative affect, wellbeing, and personality traits. Individual characteristics may be associated with distinct trajectories of persistent social functioning deficits as opposed to time limited impairments that are more responsive to intervention, as has been demonstrated in disorders of social functioning such as antisocial behavior.⁹⁴

Due to the large body of evidence linking cognitive and social cognitive impairments to social functioning in psychiatric disorders,^29,78–80 it was an unexpected finding that no cognitive variables were selected in our best performing models. While the importance of cognition could have been masked by the more heavily weighted variables in our combined and cognition-behavioral models, the brain-cognition model failed to significantly predict social functioning, which suggests that cognition was not highly predictive of social functioning in this healthy sample of young adults. As discussed above, this may be attributed to limited variance in cognitive functioning in a healthy young adult sample. Additionally, the construct of social functioning used in this study was very broad; the primary social cognition measure of emotion recognition may instead be a better predictor of specific aspects of social functioning.^29,95,96

Interestingly, while the brain-only and brain-cognition models failed to significantly predict social functioning, in the combined model cortical thickness of the middle frontal and superior temporal regions were consistently selected as model variables. This could suggest that while neuroanatomical measures may not be necessary (i.e., do not provide additional value over behavioral variables) in the prediction of social functioning in healthy young adults, when considered alongside behavioral variables these regions may be investigated for playing a potential – but seemingly very modest - role in social functioning. These represent important hubs of the social brain and their structure (i.e., relative size) and function has been linked to higher order social cognitive processes such as theory of mind and emotion recognition.^20,97–100 The structure and function of these brain regions have been consistently linked to social functioning deficits in schizophrenia.²² Additionally, machine learning studies predicting functional outcomes from neuroanatomical data have found these regions to be important for their prediction models in bipolar disorder, clinical high risk, and recent onset depression samples.^17,23 Thus, it is possible that structural measures of the middle frontal and temporal regions may serve as a transdiagnostic risk marker for social functioning across the spectrum of health to psychiatric illness. However, due to the poor performance of the brain-only model and the lower feature weights of these regions relative to the behavioral variables in the combined model, these results must be interpreted with caution. Future research is needed to characterize any relationship between these brain measures and social functioning in healthy samples.

Limitations and Future Research

The results of this study must be interpreted in the context of several limitations. Although efforts were made by the HCP research consortium to build a representative sample of participants based on U.S. Census data,⁵⁵ the majority of young adult participants were non-Hispanic whites, and their average education attainment was relatively high. Therefore, the generalizability of results to diverse racial, cultural, and socioeconomic groups may be limited. Many measures used in this study – including the social functioning outcome score - were self-reports, which may be less reliable than objective measures. However, use of high quality and psychometrically sound self-report measures is becoming increasingly important in health research seeking to promote patient centered outcomes that reflect perceived quality of life.^101–103 Some of the measures included in our models may not be expected to meaningfully contribute to the prediction of social functioning in a healthy young adult sample, such as the Mini-Mental State Exam. We relied on a data-driven feature selection method to eliminate variables that did not contribute to model performance, in which the machine learning algorithm learns which variables are predictive. Previous research has shown that self-control plays a role in predicting other aspects of longitudinal social functioning (i.e., wealth and criminal activity).¹⁰⁴ Our study was limited in not including a measure of self-control. Further, our models did not include sociodemographic measures that could influence social functioning. Although the top performing models had statistically significant performance, social functioning scores were mis-predicted in nearly 20% of the test sample.

This study analyzed cross sectional data in a proof-of-concept manner. Clinically useful models must predict longitudinal outcomes and future research should extend our models to the prediction of social functioning at later timepoints using baseline data. This study is limited by not developing a cutoff score that is clinically meaningful for screening individuals who have or are at risk for poor social functioning. Further work is needed to develop and validate such cut-off scores in healthy and clinical populations. Additionally, due to the relatively small proportion of individuals in this healthy sample with potentially problematic social functioning,⁵⁷ further replication is needed in samples with higher incidence of poor social functioning, such as may be seen in clinical populations, to confirm whether these models can accurately identify those at risk for clinically meaningful poor social functioning.

Given the enormous burden of social functioning deficits on mental and physical health and the lack of treatment options for persistent social functioning deficits across psychiatric disorders, identifying individuals who are at risk in early adulthood – if not earlier -- and addressing potential treatment targets is an urgent research need. Models predicting longitudinal outcomes that demonstrate superiority over clinician prediction¹⁷ are needed to promote clinical utility.

Conclusions

Our findings indicate that machine learning models relying on psychological and behavioral measures – but not neuroanatomical measures – may be useful for predicting social functioning and provide directions for future prospective research. While a renewed focus on understanding and leveraging biological underpinnings of complex behaviors such as social functioning characterizes the state of the science, our results suggest that prognostic models utilizing less costly and more readily available psychological and behavioral data may be optimal. Further, the study of psychological constructs may be more impactful in understanding the potential intervention points for poor social functioning that can be targeted to improve an individual’s long-term well-being.