Machine learning models for diagnosis and risk prediction in eating disorders, depression, and alcohol use disorder

This study uses machine learning models to uncover diagnostic and risk prediction markers for eating disorders (EDs), major depressive disorder (MDD), and alcohol use disorder (AUD). Utilizing case-control samples (ages 18-25 years) and a longitudinal population-based sample (n=1,851), the models, incorporating diverse data domains, achieved high accuracy in classifying EDs, MDD, and AUD from healthy controls. The area under the receiver operating characteristic curves (AUC-ROC [95% CI]) reached 0.92 [0.86-0.97] for AN and 0.91 [0.85-0.96] for BN, without relying on body mass index as a predictor. The classification accuracies for MDD (0.91 [0.88-0.94]) and AUD (0.80 [0.74-0.85]) were also high. Each data domain emerged as accurate classifiers individually, with personality distinguishing AN, BN, and their controls with AUC-ROCs ranging from 0.77 to 0.89. The models demonstrated high transdiagnostic potential, as those trained for EDs were also accurate in classifying AUD and MDD from healthy controls, and vice versa (AUC-ROCs, 0.75-0.93). Shared predictors, such as neuroticism, hopelessness, and symptoms of attention-deficit/hyperactivity disorder, were identified as reliable classifiers. For risk prediction in the longitudinal population sample, the models exhibited moderate performance (AUC-ROCs, 0.64-0.71), highlighting the potential of combining multi-domain data for precise diagnostic and risk prediction applications in psychiatry.

Neuroanatomical markers of psychotic experiences in adolescents: A machine-learning approach in a longitudinal population-based sample

It is important to identify accurate markers of psychiatric illness to aid early prediction of disease course. Subclinical psychotic experiences (PEs) are important risk factors for later mental ill-health and suicidal behaviour. This study used machine learning to investigate neuroanatomical markers of PEs in early and later stages of adolescence. Machine learning using logistic regression using Elastic Net regularization was applied to T1-weighted and diffusion MRI data to classify adolescents with subclinical psychotic experiences vs. controls across 3 timepoints (Time 1:11-13 years, n = 77; Time 2:14-16 years, n = 56; Time 3:18-20 years, n = 40). Neuroimaging data classified adolescents aged 11-13 years with current PEs vs. controls returning an AROC of 0.62, significantly better than a null model, p = 1.73^e-29. Neuroimaging data also classified those with PEs at 18-20 years (AROC = 0.59;P = 7.19^e-10) but performance was at chance level at 14-16 years (AROC = 0.50). Left hemisphere frontal regions were top discriminant classifiers for 11-13 years-old adolescents with PEs, particularly pars opercularis. Those with future PEs at 18-20 years-old were best distinguished from controls based on left frontal regions, right-hemisphere medial lemniscus, cingulum bundle, precuneus and genu of the corpus callosum (CC). Deviations from normal adolescent brain development in young people with PEs included an acceleration in the typical pattern of reduction in left frontal thickness and right parietal curvature, and accelerated progression of microstructural changes in right white matter and corpus callosum. These results emphasise the importance of multi-modal analysis for understanding adolescent PEs and provide important new insights into early phenotypes for psychotic experiences.

Stress-primed secretory autophagy promotes extracellular BDNF maturation by enhancing MMP9 secretion

The stress response is an essential mechanism for maintaining homeostasis, and its disruption is implicated in several psychiatric disorders. On the cellular level, stress activates, among other mechanisms, autophagy that regulates homeostasis through protein degradation and recycling. Secretory autophagy is a recently described pathway in which autophagosomes fuse with the plasma membrane rather than with lysosomes. Here, we demonstrate that glucocorticoid-mediated stress enhances secretory autophagy via the stress-responsive co-chaperone FK506-binding protein 51. We identify the matrix metalloproteinase 9 (MMP9) as one of the proteins secreted in response to stress. Using cellular assays and in vivo microdialysis, we further find that stress-enhanced MMP9 secretion increases the cleavage of pro-brain-derived neurotrophic factor (proBDNF) to its mature form (mBDNF). BDNF is essential for adult synaptic plasticity and its pathway is associated with major depression and posttraumatic stress disorder. These findings unravel a cellular stress adaptation mechanism that bears the potential of opening avenues for the understanding of the pathophysiology of stress-related disorders.

Brain event-related potentials predict individual differences in inhibitory control

Stop-signal reaction time (SSRT), the time needed to cancel an already-initiated motor response, quantifies individual differences in inhibitory control. Electrophysiological correlates of SSRT have primarily focused on late event-related potential (ERP) components over midline scalp regions from successfully inhibited stop trials. SSRT is robustly associated with the P300, there is mixed evidence for N200 involvement, and there is little information on the role of early ERP components. Here, machine learning was first used to interrogate ERPs during both successful and failed stop trials from 64 scalp electrodes at 4 ms resolution (n = 148). The most predictive model included data from both successful and failed stop trials, with a cross-validated Pearson's r of 0.32 between measured and predicted SSRT, significantly higher than null models. From successful stop trials, spatio-temporal features overlapping the N200 in right frontal areas and the P300 in frontocentral areas predicted SSRT, as did early ERP activity (<200 ms). As a demonstration of the reproducibility of these findings, the application of this model to a separate dataset of 97 participants was also significant (r = 0.29). These results show that ERPs during failed stops are relevant to SSRT, and that both early and late ERP activity contribute to individual differences in SSRT. Notably, the right lateralized N200, which predicted SSRT here, is not often observed in neurotypical adults. Both the ascending slope and peak of the P300 component predicted SSRT. These results were replicable, both within the training sample and when applied to ERPs from a separate dataset.

Brain-predicted age difference score is related to specific cognitive functions: a multi-site replication analysis

Brain-predicted age difference scores are calculated by subtracting chronological age from 'brain' age, which is estimated using neuroimaging data. Positive scores reflect accelerated ageing and are associated with increased mortality risk and poorer physical function. To date, however, the relationship between brain-predicted age difference scores and specific cognitive functions has not been systematically examined using appropriate statistical methods. First, applying machine learning to 1359 T1-weighted MRI scans, we predicted the relationship between chronological age and voxel-wise grey matter data. This model was then applied to MRI data from three independent datasets, significantly predicting chronological age in each dataset: Dokuz Eylül University (n = 175), the Cognitive Reserve/Reference Ability Neural Network study (n = 380), and The Irish Longitudinal Study on Ageing (n = 487). Each independent dataset had rich neuropsychological data. Brain-predicted age difference scores were significantly negatively correlated with performance on measures of general cognitive status (two datasets); processing speed, visual attention, and cognitive flexibility (three datasets); visual attention and cognitive flexibility (two datasets); and semantic verbal fluency (two datasets). As such, there is firm evidence of correlations between increased brain-predicted age differences and reduced cognitive function in some domains that are implicated in cognitive ageing.

Embracing diversity and inclusivity in an academic setting: Insights from the Organization for Human Brain Mapping

Scientific research aims to bring forward innovative ideas and constantly challenges existing knowledge structures and stereotypes. However, women, ethnic and cultural minorities, as well as individuals with disabilities, are systematically discriminated against or even excluded from promotions, publications, and general visibility. A more diverse workforce is more productive, and thus discrimination has a negative impact on science and the wider society, as well as on the education, careers, and well-being of individuals who are discriminated against. Moreover, the lack of diversity at scientific gatherings can lead to micro-aggressions or harassment, making such meetings unpleasant, or even unsafe environments for early career and underrepresented scientists. At the Organization for Human Brain Mapping (OHBM), we recognized the need for promoting underrepresented scientists and creating diverse role models in the field of neuroimaging. To foster this, the OHBM has created a Diversity and Inclusivity Committee (DIC). In this article, we review the composition and activities of the DIC that have promoted diversity within OHBM, in order to inspire other organizations to implement similar initiatives. Activities of the committee over the past four years have included (a) creating a code of conduct, (b) providing diversity and inclusivity education for OHBM members, (c) organizing interviews and symposia on diversity issues, and (d) organizing family-friendly activities and providing childcare grants during the OHBM annual meetings. We strongly believe that these activities have brought positive change within the wider OHBM community, improving inclusivity and fostering diversity while promoting rigorous, ground-breaking science. These positive changes could not have been so rapidly implemented without the enthusiastic support from the leadership, including OHBM Council and Program Committee, and the OHBM Special Interest Groups (SIGs), namely the Open Science, Student and Postdoc, and Brain-Art SIGs. Nevertheless, there remains ample room for improvement, in all areas, and even more so in the area of targeted attempts to increase inclusivity for women, individuals with disabilities, members of the LGBTQ+ community, racial/ethnic minorities, and individuals of lower socioeconomic status or from low and middle-income countries. Here, we present an overview of the DIC's composition, its activities, future directions and challenges. Our goal is to share our experiences with a wider audience to provide information to other organizations and institutions wishing to implement similar comprehensive diversity initiatives. We propose that scientific organizations can push the boundaries of scientific progress only by moving beyond existing power structures and by integrating principles of equity and inclusivity in their core values.

The biological classification of mental disorders (BeCOME) study: a protocol for an observational deep-phenotyping study for the identification of biological subtypes

BACKGROUND

A major research finding in the field of Biological Psychiatry is that symptom-based categories of mental disorders map poorly onto dysfunctions in brain circuits or neurobiological pathways. Many of the identified (neuro) biological dysfunctions are "transdiagnostic", meaning that they do not reflect diagnostic boundaries but are shared by different ICD/DSM diagnoses. The compromised biological validity of the current classification system for mental disorders impedes rather than supports the development of treatments that not only target symptoms but also the underlying pathophysiological mechanisms. The Biological Classification of Mental Disorders (BeCOME) study aims to identify biology-based classes of mental disorders that improve the translation of novel biomedical findings into tailored clinical applications.

METHODS

BeCOME intends to include at least 1000 individuals with a broad spectrum of affective, anxiety and stress-related mental disorders as well as 500 individuals unaffected by mental disorders. After a screening visit, all participants undergo in-depth phenotyping procedures and omics assessments on two consecutive days. Several validated paradigms (e.g., fear conditioning, reward anticipation, imaging stress test, social reward learning task) are applied to stimulate a response in a basic system of human functioning (e.g., acute threat response, reward processing, stress response or social reward learning) that plays a key role in the development of affective, anxiety and stress-related mental disorders. The response to this stimulation is then read out across multiple levels. Assessments comprise genetic, molecular, cellular, physiological, neuroimaging, neurocognitive, psychophysiological and psychometric measurements. The multilevel information collected in BeCOME will be used to identify data-driven biologically-informed categories of mental disorders using cluster analytical techniques.

DISCUSSION

The novelty of BeCOME lies in the dynamic in-depth phenotyping and omics characterization of individuals with mental disorders from the depression and anxiety spectrum of varying severity. We believe that such biology-based subclasses of mental disorders will serve as better treatment targets than purely symptom-based disease entities, and help in tailoring the right treatment to the individual patient suffering from a mental disorder. BeCOME has the potential to contribute to a novel taxonomy of mental disorders that integrates the underlying pathomechanisms into diagnoses.

TRIAL REGISTRATION

Retrospectively registered on June 12, 2019 on ClinicalTrials.gov (TRN: NCT03984084).

Inhibitory-control event-related potentials correlate with individual differences in alcohol use

Impulsivity is a multidimensional construct that is related to different aspects of alcohol use, abuse, and dependence. Inhibitory control, one facet of impulsivity, can be assayed using the stop-signal task (SST) and quantified behaviorally via the stop-signal reaction time (SSRT) and electrophysiologically using event-related potentials (ERPs). Research on the relationship between alcohol use and SSRTs, and between alcohol use and inhibitory-control ERPs, is mixed. Here, adult alcohol users (n = 79), with a wide range of scores on the Alcohol Use Disorders Identification Test (AUDIT), completed the SST under electroencephalography (EEG) (70% of participants had AUDIT total scores greater than or equal to 8). Other measures, including demographic, self-report, and task-based measures of impulsivity, personality, and psychological factors, were also recorded. A machine-learning method with penalized linear regression was used to correlate individual differences in alcohol use with impulsivity measures. Four separate models were tested, with out-of-sample validation used to quantify performance. ERPs alone statistically predicted alcohol use (cross-validated r = 0.28), with both early and late ERP components contributing to the model (larger N2, but smaller P3, amplitude). Behavioral data from a wide range of impulsivity measures were also associated with alcohol use (r = 0.37). SSRT was a relatively weak statistical predictor, whereas the Stroop interference effect was relatively strong. The addition of nonimpulsivity behavioral measures did not improve the correlation (r = 0.34) and was similar when ERPs were combined with non-ERP data (r = 0.29). These findings show that inhibitory control ERPs are robustly correlated individual differences in alcohol use.

Quantifying performance of machine learning methods for neuroimaging data

Machine learning is increasingly being applied to neuroimaging data. However, most machine learning algorithms have not been designed to accommodate neuroimaging data, which typically has many more data points than subjects, in addition to multicollinearity and low signal-to-noise. Consequently, the relative efficacy of different machine learning regression algorithms for different types of neuroimaging data are not known. Here, we sought to quantify the performance of a variety of machine learning algorithms for use with neuroimaging data with various sample sizes, feature set sizes, and predictor effect sizes. The contribution of additional machine learning techniques - embedded feature selection and bootstrap aggregation (bagging) - to model performance was also quantified. Five machine learning regression methods - Gaussian Process Regression, Multiple Kernel Learning, Kernel Ridge Regression, the Elastic Net and Random Forest, were examined with both real and simulated MRI data, and in comparison to standard multiple regression. The different machine learning regression algorithms produced varying results, which depended on sample size, feature set size, and predictor effect size. When the effect size was large, the Elastic Net, Kernel Ridge Regression and Gaussian Process Regression performed well at most sample sizes and feature set sizes. However, when the effect size was small, only the Elastic Net made accurate predictions, but this was limited to analyses with sample sizes greater than 400. Random Forest also produced a moderate performance for small effect sizes, but could do so across all sample sizes. Machine learning techniques also improved prediction accuracy for multiple regression. These data provide empirical evidence for the differential performance of various machines on neuroimaging data, which are dependent on number of sample size, features and effect size.

Individual differences in learning from probabilistic reward and punishment predicts smoking status

INTRODUCTION

The ability to update reward and punishment contingencies is a fundamental aspect of effective decision-making, requiring the ability to successfully adapt to the changing demands of one's environment. In the case of nicotine addiction, research has predominantly focused on reward- and punishment-based learning processes among current smokers relative to non-smokers, whereas less is known about these processes in former smokers.

METHODS

In a total sample of 105 students, we used the Probabilistic Selection Task to examine differences in reinforcement learning among 41 current smokers, 29 ex-smokers, and 35 non-smokers. The PST was comprised of a training and test phase that allowed for the comparison of learning from positive versus negative feedback.

RESULTS

The test phase of the Probabilistic Selection Task significantly predicted smoking status. Current and non-smokers were classified with moderate accuracy, whereas ex-smokers were typically misclassified as smokers. Lower rates of learning from rewards were associated with an increased likelihood of being a smoker or an ex-smoker compared with being a non-smoker. Higher rates of learning from punishment were associated with an increased likelihood of being a smoker relative to non-smoker. However, learning from punishment did not predict ex-smoker status.

CONCLUSIONS

Current smokers and ex-smokers were less likely to learn from rewards, supporting the hypothesis that deficient reward processing is a feature of chronic addiction. In addition, current smokers were more sensitive to punishment than ex-smokers, contradicting some recent findings.

A Combination of Impulsivity Subdomains Predict Alcohol Intoxication Frequency

BACKGROUND

Impulsivity, broadly characterized as the tendency to act prematurely without foresight, is linked to alcohol misuse in college students. However, impulsivity is a multidimensional construct and different subdomains likely underlie different patterns of alcohol misuse. Here, we quantified the association between alcohol intoxication frequency and alcohol consumption frequency and choice, action, cognitive, and trait domains of impulsivity.

METHODS

University student drinkers (n = 106) completed a battery of demographic and alcohol-related items, as well as self-report and task-based measures indexing different facets of impulsivity. Two orthogonal latent factors, intoxication frequency and alcohol consumption frequency, were generated. Their validity was demonstrated with respect to adverse consequences of alcohol use. Machine learning with penalized regression and feature selection was then utilized to predict intoxication and alcohol consumption frequency using all impulsivity subdomains. Out-of-sample validation was used to quantify model performance.

RESULTS

Impulsivity measures alone were significant predictors of intoxication frequency, but not consumption frequency. Propensity for increased intoxication frequency was characterized by increased trait impulsivity, including the Disinhibition subscale of the Sensation Seeking Scale, Attentional and Non-planning subscales of the Barratt Impulsiveness Scale, increased task-based cognitive impulsivity (response time variability), and increased choice impulsivity (steeper delay discounting on a delay discounting questionnaire). A model combining impulsivity domains with other risk factors (gender; nicotine, cannabis, and other drug use; executive functioning; and learning processes) was also significant but did not outperform the model comprising of impulsivity alone.

CONCLUSIONS

Intoxication frequency, but not consumption frequency, was characterized by a number of impulsivity subdomains.

Neuromarkers for Mental Disorders: Harnessing Population Neuroscience

Despite abundant research into the neurobiology of mental disorders, to date neurobiological insights have had very little impact on psychiatric diagnosis or treatment. In this review, we contend that the search for neuroimaging biomarkers-neuromarkers-of mental disorders is a highly promising avenue toward improved psychiatric healthcare. However, many of the traditional tools used for psychiatric neuroimaging are inadequate for the identification of neuromarkers. Specifically, we highlight the need for larger samples and for multivariate analysis. Approaches such as machine learning are likely to be beneficial for interrogating high-dimensional neuroimaging data. We suggest that broad, population-based study designs will be important for developing neuromarkers of mental disorders, and will facilitate a move away from a phenomenological definition of mental disorder categories and toward psychiatric nosology based on biological evidence. We provide an outline of how the development of neuromarkers should occur, emphasizing the need for tests of external and construct validity, and for collaborative research efforts. Finally, we highlight some concerns regarding the development, and use of, neuromarkers in psychiatric healthcare.

The potential of neuroimaging for identifying predictors of adolescent alcohol use initiation and misuse

BACKGROUND AND AIMS

Dysfunction in brain regions underlying impulse control, reward processing and executive function have been associated previously with adolescent alcohol misuse. However, identifying pre-existing neurobiological risk factors, as distinct from changes arising from early alcohol-use, is difficult. Here, we outline how neuroimaging data can identify the neural predictors of adolescent alcohol-use initiation and misuse by using prospective longitudinal studies to follow initially alcohol-naive individuals over time and by neuroimaging adolescents with inherited risk factors for alcohol misuse.

METHOD

A comprehensive narrative of the literature regarding neuroimaging studies published between 2010 and 2016 focusing on predictors of adolescent alcohol use initiation and misuse.

FINDINGS

Prospective, longitudinal neuroimaging studies have identified pre-existing differences between adolescents who remained alcohol-naive and those who transitioned subsequently to alcohol use. Both functional and structural grey matter differences were observed in temporal and frontal regions, including reduced brain activity in the superior frontal gyrus and temporal lobe, and thinner temporal cortices of future alcohol users. Interactions between brain function and genetic predispositions have been identified, including significant association found between the Ras protein-specific guanine nucleotide releasing factor 2 (RASGRF2) gene and reward-related striatal functioning.

CONCLUSIONS

Neuroimaging predictors of alcohol use have shown modest utility to date. Future research should use out-of-sample performance as a quantitative measure of a predictor's utility. Neuroimaging data should be combined across multiple modalities, including structural information such as volumetrics and cortical thickness, in conjunction with white-matter tractography. A number of relevant neurocognitive systems should be assayed; particularly, inhibitory control, reward processing and executive functioning. Combining a rich magnetic resonance imaging data set could permit the generation of neuroimaging risk scores, which could potentially yield targeted interventions.

Ventral Striatum Connectivity During Reward Anticipation in Adolescent Smokers

Substance misusers, including adolescent smokers, often have reduced reward system activity during processing of non-drug rewards. Using a psychophysiological interaction approach, we examined functional connectivity with the ventral striatum during reward anticipation in a large (N = 206) sample of adolescent smokers. Increased smoking frequency was associated with (1) increased connectivity with regions involved in saliency and valuation, including the orbitofrontal cortex and (2) reduced connectivity between the ventral striatum and regions associated with inhibition and risk aversion, including the right inferior frontal gyrus. These results demonstrate that functional connectivity during reward processing is relevant to adolescent addiction.