Our Research Projects

The Center for Suicide Research and Prevention (CSRP) is a multidisciplinary, practice-based center aimed at advancing and implementing innovative new suicide prevention research. It will support the development, deployment, and evaluation of practice-ready and clinically focused interventions aimed at improving the identification and effective treatment of patients at risk of suicide. This is a collaborative effort between researchers, clinicians, and stakeholders at the Center for Precision Psychiatry, Mass General Brigham and Harvard University.

To advance prediction and prevention of suicide, a leading cause of death, we need more precise understanding of who needs what intervention and when through investigation of proximal risk factors, and impulsivity is an ideal candidate. The proposed research will use digital monitoring of patients at high risk for suicide to conduct the first intensive longitudinal assessment of multiple components of impulsivity and suicidal urges in real time among individuals at high risk for suicide. Findings will improve understanding of impulsivity as a transdiagnostic short-term risk factor for suicide, laying the groundwork for development of targeted interventions for suicide prevention, and may be generalizable to other impulsive behaviors.

Aim 1 will involve secondary data analysis of a digital monitoring study of individuals presenting to an emergency department with suicidal thoughts to analyze real time associations between impulsivity, suicidal urges, and ability to resist suicidal urges. We will test whether state impulsivity is predictive beyond the effect of trait impulsivity.

In Aim 2, we will conduct a digital monitoring study of 140 individuals hospitalized for suicidal thoughts to assess multiple components of state impulsivity using self-report, mobile tasks, and passive phone data, and we will test specific associations with suicidal urges and ability to resist them in real time.

In Aim 3, we will compare group-level, subgroup-level, and personalized models of these data using a combination of inferential statistics (network modeling) and predictive analytics (machine learning). 

Major depressive disorder (MDD) is one of the leading causes of disability worldwide. While it is known that MDD is caused by a combination of genetic, environmental and psychological factors, the pathophysiology of MDD remains unclear. Prior work of brain morphology and functional brain networks in MDD produced mixed or even contradictory results, likely due to (i) limited sample sizes of most neuroimaging studies of MDD, and (ii) the heterogeneity of MDD phenotypes across studies and cohorts.

Leveraging the largest available and growing datasets with structural and functional MRI scans,1–3 genome-wide genomic data and whole-exome sequencing, along with socioeconomic and environmental factors from the UK Biobank (UKBB), as well as recent work that defined and imputed a range of MDD phenotypes, from shallow (e.g., based on self-reported depression or health records of seeking care for depression) to deep (e.g., based on structured clinical interviews, questionnaires and clinical diagnostic criteria), we now have the unprecedented opportunity to (i) clarify the neuroanatomical and functional bases of depression, (ii) identify robust brain-based clusters, or neurotypes, within depression, and (iii) probe the common and rare genetic overlap between neuroimaging features and depression and their clinical implications.

We propose to build on and expand our existing program, supported by the Tommy Fuss Fund, that links genomics, brain structure and function, and psychiatric disorders, to pursue the following specific aims.

Aim 1: Clarify the neuroanatomical and functional underpinnings of depression across the phenotyping spectrum
Aim 2:Identify and validate brain-based biotypes of depression based on participant- level clustering of neuroimaging features
Aim 3: Examine the role of common and rare genetic variation in depression phenotypes and biotypes, as well as their overlap with other aspects of mental and physical health.

Depression is a common, debilitating mental disorder with a prevalence over 260 million patients world-wide and associated with 43 million years lived with disability in 2017 ¹. Depression is also frequently associated with other physical and mental health outcomes, such as hypertensive diseases, metabolic disorders, and neurological diseases (i.e., sleep disorders, migraine, etc.) ². Currently, treatment choices for depression include antidepressant medications, psychotherapies, and other non-pharmacological treatments such as electroconvulsive therapy (ECT), repetitive transcranial magnetic stimulation (rTMS) and deep brain stimulation (DBS). Despite these multiple therapeutic options, a substantial fraction of patients with depression experience limited benefit. It has been estimated that about one-third of patients treated for major depressive disorder (MDD) do not respond satisfactorily to antidepressants (ADs) ³. Specifically, those patients who do not respond to multiple antidepressants or other treatments (such as ECT) are often referred to as having treatment-resistant depression (TRD). While TRD is widely recognized as an important clinical problem, there is ongoing debate about the definition of TRD in clinical practice and trials. There is also a lack of systematic examination of clinical features using real-world data for the existing TRD definitions ⁴. The lack of a clear, consensus definition hinders the diagnosis, clinical management, and novel therapeutic development of TRD. In many respects, the problem of TRD is one for which “precision medicine” approaches to prevention and early intervention could be transformative. The precision medicine framework aims to improve diagnosis, prevention, and treatment by accounting for individual differences in biology, lifestyle, and environment ⁵. At present, we also have no robust methods to predict risk of TRD and variability in treatment response for depression in general. Characterizing the heterogeneity and clinical characteristics of TRD patients and potential risk factors/predictors of TRD could enable much more effective strategies for drug development for depression and more specifically for TRD. The goal of this project is to utilize large-scale resources of electronic health records (EHRs) and genomics to identify TRD (and depression) patient cohorts by leveraging biobanks across multiple countries/health systems and build patient stratification tools that can estimate individualized risk of TRD.

Aim 1:Develop automated TRD phenotyping algorithm and identify patient cohorts in MGB RPDR, FinnGen and UK Biobank.
Aim 2: Identify predictors and build machine learning models for TRD prediction in MGBB, FinnGen and UK Biobank.

In this second phase of the project "Linking Genomics, Brain Structure, and the Development of Psychiatric Disorders", we propose the following aims:

Aim 1: Create a more detailed map of the underlying genetic relationships that influence brain structure
Aim 2: Identify and prioritize genes and patterns of brain gene expression that contribute most strongly to each of the 4 genomic factors underlying 11 psychiatric disorders
Aim 3: Identify potential drug compounds that target genes contributing to transdiagnostic genomic factors

In this study, we propose to apply state-of-the-art artificial intelligence methods to predict bipolar disorder in the large Partners Healthcare System which includes EHR data for more than 5 million patients spanning 15-20 years of data. Our goal is to develop and validate a risk prediction algorithm focused on early-onset (before age 25) bipolar disorder that can identify individuals who either are at high risk of developing the disorder or who have undiagnosed bipolar disorder. 

Suicide is the tenth leading cause of death in the United States, accounting for more than 40,000 deaths annually. Despite ongoing efforts to reduce the burden of suicide and suicidal behavior, rates have remained relatively constant over the past half century. Attempts to predict suicidal behavior have relied almost exclusively on self-reporting of suicidal thoughts and intentions. This is problematic because of well-known reporting biases and the fact that many people at high risk are motivated to deny suicidal thoughts to avoid hospitalization. Even though the majority of all suicide decedents have contact with a healthcare professional in the month before their death, suicide risk is rarely detected in such cases. Efforts to identify risk factors have also been stymied by the fact that suicide is a low-base rate event so that very large samples are needed to test the complex combinations of factors that are likely to contribute to risk. The widespread adoption of longitudinal electronic health records (EHRs) has created a powerful but still under-utilized resource for detecting and predicting important health outcomes. In prior work using machine learning methods to analyze structured EHR data, we have developed predictive models that detect up to 45% of first-episode suicidal behavior, on average 3 years in advance. Here we aim to systematically extend and improve our EHR prediction methods in a large healthcare system (N = 4.6 million patients) by incorporating 1) external public record datasets (LexisNexis SocioEconomic Health Attribute data) that include environmental, socioeconomic, and life event information; 2) natural language processing (NLP) to leverage unstructured EHR text, including text-based scores that capture RDoC domains; 3) a novel method of deriving temporal risk envelopes to capture the time-dependent effects of individual risk factors; and 4) clinical risk trajectories that incorporate ordered temporal sequences of risk factors. We will systematically compare the performance of each of these approaches to identify optimal strategies for enhancing risk surveillance and prediction in healthcare settings. Completion of these aims would represent a crucial step towards novel, clinically deployable, and potentially transformative tools for improving outcomes for those at risk for suicide and suicidal behavior.

Psychosis is a major public health challenge, with approximately 100,000 adolescents and young adults in the US experiencing a first episode of psychosis (FEP) every year. Early intervention following FEP is critical for achieving improved outcomes, yet treatment of FEP is often delayed between 1 and 3 years in the US due to delays in detection and referral. The World Health Organization has advocated shortening the duration of untreated psychosis (DUP) to three months or less. The goal of this study is to develop and validate a universal EHR-based screening tool for early detection of FEP across large clinical populations in diverse healthcare settings. In order to maximize the impact and generalizability of the tool across a wide range of healthcare settings, we will rely only on coded medical information collected in the course of care and thus widely available in EHRs. The tool will be developed and validated with data from three diverse health systems that cover over 8 million patients spanning a wide range of demographic, socioeconomic and ethnic backgrounds: Partners Healthcare System, Boston Children’s Hospital, and Boston Medical Center. The study will be conducted by a closely collaborating interdisciplinary team of clinical specialists, psychosis researchers, and risk modeling experts based at these health systems and Harvard Medical School, with extensive experience in treating psychosis patients, and developing strategies for detecting FEP and EHR-based risk screening tools for early detection of various clinical conditions. Our preliminary studies show that EHR-based risk models can be used to sensitively and specifically detect FEP cases, on average 2 years before the first psychosis diagnosis appears in their EHR. Our specific aims include: 1. Define a robust cross-site case definition for FEP that relies only on information commonly available in EHRs and validate it through expert chart review; 2. Train and validate a predictive model for early detection of FEP based on large samples of patient data from the three sites; 3. Develop and validate FEP early detection models for key subpopulations, including patients receiving care at mental health clinics, adolescent medicine outpatient programs, and substance abuse treatment programs; and 4. Engage clinical stakeholders in the process of developing a prototype clinician-facing EHR-based risk screening tool for FEP, and release it as an open source SMART App, enabling further validation and clinical integration across a wide range of healthcare settings. Completion of these aims would provide a novel, clinically deployable, and potentially transformative tool for improving the trajectory of those affected with psychosis and reducing the burden and costs of untreated illness.

Suicide is the 2nd leading cause of death among adolescents, and, alarmingly, rates have continued to increase over the past decade. With the current dearth of empirically supported psychosocial interventions for adolescent suicidal behavior, there is a great need to identify processes underlying risk for this outcome. Interpersonal negative life events (NLEs) have been linked to short-term risk for suicidal behavior (i.e., suicide attempts and deaths). However, how or why these interpersonal NLEs confer risk is currently unclear. Understanding the moderators and mechanisms of the link between interpersonal NLEs and suicidal behavior is crucial in order to identify when adolescents are most at risk and to develop interventions that can target these mechanisms and reduce suicidal behavior during these periods of risk. Interpersonal NLEs may lead to suicidal behavior through their impact negative emotional states (e.g., agitation) and social connectedness, and further magnified by rejection sensitivity, emotion regulation difficulties, and sleep disturbance. Although past studies have found these constructs to be linked with suicidal behavior, this research is limited in three key ways: (1) Most prior studies have examined long time periods (months and years), which lack the temporal resolution to examine these mechanisms on the time scale in which they unfold (hours and days). We will intensively (over hours and days) examine how interpersonal NLEs relate to suicidal behavior. (2) Little is known about the impact of the social system, such as the family unit, and what key collaterals (e.g., parents) are observing in the short-term leading up to their adolescent’s suicidal behavior. By including parental reports, this study will reveal parents’ awareness of their adolescents’ suicide risk and potentially identify how parents can help them during these high-risk times. (3) Most past work has been insufficiently powered to examine links with suicidal behavior (instead focusing on suicidal thoughts). We propose to examine how and why interpersonal NLEs confer risk for suicidal behavior using smartphone-based ecological momentary assessment (EMA) and wearable sensors (i.e., wrist actigraphy) among 600 adolescents and their parents recruited from three diverse sites (Brown, Old Dominion, Rutgers) during the month after discharge from acute psychiatric care. We will address three aims: Aim 1. Examine how interpersonal (particularly familial and peer) NLEs increase risk for suicidal behavior through increased negative emotion and decreased social connectedness. Aim 2. Examine how rejection sensitivity moderates the link from interpersonal NLEs to negative emotionality/decreased social connectedness. Aim 3. Examine how sleep disturbance and emotion regulation difficulties moderate links from increased negative emotions and decreased social connectedness to suicidal behavior. This work will significantly advance understanding of proximal risk for suicidal behavior in youth and how the field assesses and treats suicidal adolescents.

Learn more >

Abstract: Self-injurious behaviors occur at alarmingly high rates among adolescents, with suicide ranking as the second leading cause of death among those ages 15-24. A history of prior self-injury, including both nonsuicidal self-injury and suicidal self-injury (e.g., suicide attempts), has been consistently found to be the strongest predictor of future suicidal behavior, with evidence suggesting that the more severe such behaviors are, the greater the risk for future self-injury. Importantly, however, our current means of assessing severity of prior self-injury is almost entirely reliant on self-report, despite the fact that self-injury frequently leaves tangible physical markings. Although applications of machine learning in medical image analysis are growing exponentially, none have attempted to augment suicide risk detection through automated analysis of self-directed tissue damage. Leveraging computer vision to automatically assess images of tissue damage has the potential to obviate complete reliance on subjective patient report of self-injury severity characteristics. Thus, the objective of this proposal is to utilize computer vision techniques to automate the assessment of hypothesized self-injury visual severity indicators, learn new visual severity indicators, and determine the utility of these visual signals in predicting prospective suicide attempt risk. Community adolescents ages 16 to 18 years old will be recruited on Facebook and Instagram if they have currently visible physical marking(s) secondary to self-injury. Participants will securely upload images of markings secondary to intentional self-injury. A subset of participants will be followed longitudinally for three months to assess prospective suicide attempts. We will employ deep convolutional neural networks, a class of artificial neural networks, to develop algorithms to detect severity indices of self-injury and to examine their accuracy in predicting short-term prospective suicide risk. We will assess the generalizability of a subset of algorithms by applying them to a separate clinical sample of psychiatrically hospitalized adolescents ages 16 to 18 years old. This proof-of-concept study will set the stage to determine the feasibility of pursuing our long-term goal of integrating this technology into psychiatric care entry-points (e.g., emergency departments, inpatient units) to assess whether this technology can augment current suicide risk assessment models and in turn, serve as a clinical decision-support tool to help clinicians assess suicide risk. This research is significant in that it aligns with the NIMH/National Action Alliance for Suicide Prevention’s Prioritized Research Agenda for Suicide Prevention’s Aspirational Goal 2 of determining suicide risk in diverse populations and settings using feasible and effective assessment approaches, and Goal 3 of finding novel ways to assess for imminent suicide risk, given that our target prediction period is three months.

Abstract: The objective of this proposal is to extend ongoing research supported by a National Institute of Mental Health-funded R21 aimed at utilizing computer vision techniques to automate the assessment of self-injury visual severity indicators and to determine the utility of these visual signals in predicting suicide risk. The R21 will fund the recruitment of a large sample of adolescents recruited online and followed prospectively. The R21 will also fund the recruitment of a sample of 50 psychiatrically hospitalized adolescents to be assessed at one time point. This Young Investigator Grant will fund the recruitment of an additional 50 psychiatrically hospitalized adolescents ages 13-18 with a history of self-injury who have currently visible physical marking(s) secondary to self-injury. This Young Investigator Grant will also fund the prospective assessment of suicidal behavior across the full clinical sample of 100 adolescents. One month after psychiatric hospitalization discharge, adolescents will be assessed for prospective engagement in suicide attempts. Deep convolutional neural networks will be applied to the images to detect severity indices of self-injury and to examine their accuracy in predicting prospective suicide attempt risk in this high-risk clinical sample. This study will help assess the feasibility of pursuing our long-term goal of integrating this technology into psychiatric care entry-points to assess whether it can improve suicide risk assessment models and in turn, serve as a clinical decision-support tool.

Abstract: Suicide is the second leading cause of death among adolescents and the rates have doubled since 2000. The rise in suicide rates is due in part to a failure to identify short-term risk factors for suicidal thoughts and behaviors. Additionally, most existing research relies primarily on group-level methodological approaches to suicide risk assessment; intraindividual suicide risk processes are largely neglected. Because of the lack of knowledge on short-term and individual-level suicide risk, it remains unclear when and how to intervene with the individual adolescents who need it most. Three particularly promising observable, state-sensitive, temporally delimited, and modifiable proximal indicators of suicide risk among adolescents are social engagement, sleep, and physical activity. Although acute changes in these behavioral processes are often denoted as imminent “behavioral warning signs” of suicide, most existing research has examined these behavioral factors only as distal predictors and correlates of suicide risk. Importantly, their contribution to short-term risk using individual-level approaches (i.e., fully idiographic, “n-of-1” methods) remains unknown. The Candidate’s proposed K23’s overarching goal is to employ mobile sensing and actigraphy to assess whether objectively and passively measured acute behavioral changes from typical patterns of social engagement, sleep, and physical activity indicate proximal risk for increases in suicidal ideation using idiographic n-of-1 models in high-risk adolescents. It further aims to characterize the intraindividual network structure of these behavioral factors and suicidal ideation to enhance suicide risk assessment and guide intervention. Adolescents (N=100) admitted to an inpatient or partial hospital program due to acute suicide risk will be recruited. For a period of 3 months, wearable actigraphs will be used to assess adolescents’ sleep and physical activity and mobile sensing will be used to assess adolescents’ digital social engagement and patterns of movement to approximate additional indices of physical activity. Once-daily mobile surveys will be used to assess suicidal ideation. The proposed training plan complements the Candidate’s research plan and will facilitate training in: conducting translational digital health research in high-risk adolescents, passive mobile sensing, passive adolescent sleep and physical activity assessment via actigraphy, and advanced computational approaches to person-specific intensive data modeling. A team of leading scholars will provide expert mentorship to facilitate the Candidate’s training goals within the highly resourced environment of the Alpert Medical School of Brown University and Rhode Island Hospital. The proposed study will promote the Candidate’s long-term career goal to employ low-burden and scalable methods of assessment to develop personalized risk models that will improve the proximal prediction of suicide risk and inform intervention for youth. Through the execution of this research and training plan, the Candidate will be positioned to become a leader in the field of adolescent suicide.