Translating promise into practice: a review of machine learning in suicide research and prevention

An Explainable Artificial Intelligence Text Classifier for Suicidality Prediction in Youth Crisis Text Line Users: Development and Validation Study

BACKGROUND

Suicide represents a critical public health concern, and machine learning (ML) models offer the potential for identifying at-risk individuals. Recent studies using benchmark datasets and real-world social media data have demonstrated the capability of pretrained large language models in predicting suicidal ideation and behaviors (SIB) in speech and text.

OBJECTIVE

This study aimed to (1) develop and implement ML methods for predicting SIBs in a real-world crisis helpline dataset, using transformer-based pretrained models as a foundation; (2) evaluate, cross-validate, and benchmark the model against traditional text classification approaches; and (3) train an explainable model to highlight relevant risk-associated features.

METHODS

We analyzed chat protocols from adolescents and young adults (aged 14-25 years) seeking assistance from a German crisis helpline. An ML model was developed using a transformer-based language model architecture with pretrained weights and long short-term memory layers. The model predicted suicidal ideation (SI) and advanced suicidal engagement (ASE), as indicated by composite Columbia-Suicide Severity Rating Scale scores. We compared model performance against a classical word-vector-based ML model. We subsequently computed discrimination, calibration, clinical utility, and explainability information using a Shapley Additive Explanations value-based post hoc estimation model.

RESULTS

The dataset comprised 1348 help-seeking encounters (1011 for training and 337 for testing). The transformer-based classifier achieved a macroaveraged area under the curve (AUC) receiver operating characteristic (ROC) of 0.89 (95% CI 0.81-0.91) and an overall accuracy of 0.79 (95% CI 0.73-0.99). This performance surpassed the word-vector-based baseline model (AUC-ROC=0.77, 95% CI 0.64-0.90; accuracy=0.61, 95% CI 0.61-0.80). The transformer model demonstrated excellent prediction for nonsuicidal sessions (AUC-ROC=0.96, 95% CI 0.96-0.99) and good prediction for SI and ASE, with AUC-ROCs of 0.85 (95% CI 0.97-0.86) and 0.87 (95% CI 0.81-0.88), respectively. The Brier Skill Score indicated a 44% improvement in classification performance over the baseline model. The Shapley Additive Explanations model identified language features predictive of SIBs, including self-reference, negation, expressions of low self-esteem, and absolutist language.

CONCLUSIONS

Neural networks using large language model-based transfer learning can accurately identify SI and ASE. The post hoc explainer model revealed language features associated with SI and ASE. Such models may potentially support clinical decision-making in suicide prevention services. Future research should explore multimodal input features and temporal aspects of suicide risk.

Applications of Large Language Models in the Field of Suicide Prevention: Scoping Review

BACKGROUND

Prevention of suicide is a global health priority. Approximately 800,000 individuals die by suicide yearly, and for every suicide death, there are another 20 estimated suicide attempts. Large language models (LLMs) hold the potential to enhance scalable, accessible, and affordable digital services for suicide prevention and self-harm interventions. However, their use also raises clinical and ethical questions that require careful consideration.

OBJECTIVE

This scoping review aims to identify emergent trends in LLM applications in the field of suicide prevention and self-harm research. In addition, it summarizes key clinical and ethical considerations relevant to this nascent area of research.

METHODS

Searches were conducted in 4 databases (PsycINFO, Embase, PubMed, and IEEE Xplore) in February 2024. Eligible studies described the application of LLMs for suicide or self-harm prevention, detection, or management. English-language peer-reviewed articles and conference proceedings were included, without date restrictions. Narrative synthesis was used to synthesize study characteristics, objectives, models, data sources, proposed clinical applications, and ethical considerations. This review adhered to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) standards.

RESULTS

Of the 533 studies identified, 36 (6.8%) met the inclusion criteria. An additional 7 studies were identified through citation chaining, resulting in 43 studies for review. The studies showed a bifurcation of publication fields, with varying publication norms between computer science and mental health. While most of the studies (33/43, 77%) focused on identifying suicide risk, newer applications leveraging generative functions (eg, support, education, and training) are emerging. Social media was the most common source of LLM training data. Bidirectional Encoder Representations from Transformers (BERT) was the predominant model used, although generative pretrained transformers (GPTs) featured prominently in generative applications. Clinical LLM applications were reported in 60% (26/43) of the studies, often for suicide risk detection or as clinical assistance tools. Ethical considerations were reported in 33% (14/43) of the studies, with privacy, confidentiality, and consent strongly represented.

CONCLUSIONS

This evolving research area, bridging computer science and mental health, demands a multidisciplinary approach. While open access models and datasets will likely shape the field of suicide prevention, documenting their limitations and potential biases is crucial. High-quality training data are essential for refining these models and mitigating unwanted biases. Policies that address ethical concerns-particularly those related to privacy and security when using social media data-are imperative. Limitations include high variability across disciplines in how LLMs and study methodology are reported. The emergence of generative artificial intelligence signals a shift in approach, particularly in applications related to care, support, and education, such as improved crisis care and gatekeeper training methods, clinician copilot models, and improved educational practices. Ongoing human oversight-through human-in-the-loop testing or expert external validation-is essential for responsible development and use.

TRIAL REGISTRATION

OSF Registries osf.io/nckq7; https://osf.io/nckq7.

Individual suicide risk factors with resting-state brain functional connectivity patterns in bipolar disorder patients based on latent Dirichlet allocation model

BACKGROUND

The widespread problem of suicide and its severe burden in bipolar disorder (BD) necessitate the development of objective risk markers, aiming to enhance individual suicide risk prediction in BD.

METHODS

This study recruited 123 BD patients (61 patients with prior suicide attempted history (PSAs), 62 without (NSAs)) and 68 healthy controls (HEs). The Latent Dirichlet Allocation (LDA) model was used to decompose the resting state functional connectivity (RSFC) into multiple hyper/hypo-RSFC patterns. Thereafter, according to the quantitative results of individual heterogeneity over latent factor dimensions, the correlations were analyzed to test prediction ability.

RESULTS

Model constructed without introducing suicide-related labels yielded three latent factors with dissociable hyper/hypo-RSFC patterns. In the subsequent analysis, significant differences in the factor distributions of PSAs and NSAs showed biases on the default-mode network (DMN) hyper-RSFC factor (factor 3) and the salience network (SN) and central executive network (CEN) hyper-RSFC factor (factor 1), indicating predictive value. Correlation analysis of the individuals' expressions with their Nurses' Global Assessment of Suicide Risk (NGASR) revealed factor 3 positively correlated (r = 0.4180, p < 0.0001) and factor 1 negatively correlated (r = - 0.2492, p = 0.0055) with suicide risk. Therefore, it could be speculated that patterns more associated with suicide reflected hyper-connectivity in DMN and hypo-connectivity in SN, CEN.

CONCLUSIONS

This study provided individual suicide-associated risk factors that could reflect the abnormal RSFC patterns, and explored the suicide related brain mechanisms, which is expected to provide supports for clinical decision-making and timely screening and intervention for individuals at high risks of suicide.

Machine Learning for Mental Health: Applications, Challenges, and the Clinician's Role

PURPOSE OF REVIEW

This review aims to evaluate the current psychiatric applications and limitations of machine learning (ML), defined as techniques used to train algorithms to improve performance at a task based on data. The review emphasizes the clinician's role in ensuring equitable and effective patient care and seeks to inform mental health providers about the importance of clinician involvement in these technologies.

RECENT FINDINGS

ML in psychiatry has advanced through electronic health record integration, disease phenotyping, and remote monitoring through mobile applications. However, these applications face challenges related to health equity, privacy, translation to practice, and validation. Clinicians play crucial roles in ensuring data quality, mitigating biases, promoting algorithm transparency, guiding clinical implementation, and advocating for ethical and patient-centered use of ML tools. Clinicians are essential in addressing the challenges of ML, ensuring its ethical application, and promoting equitable care, thus improving the effectiveness of ML in practice.

Risk of repeat self-harm among individuals presenting to healthcare services: development and validation of a clinical risk assessment model (OxSET)

BACKGROUND

A self-harm episode is a major risk factor for repeat self-harm. Existing tools to assess and predict repeat self-harm have major methodological limitations, and few are externally validated.

OBJECTIVE

To develop and validate a risk assessment model of repeat self-harm up to 6 months after an episode of non-fatal self-harm that resulted in an emergency visit to hospital or specialised care.

METHODS

Using Swedish national registers, we identified 53 172 people aged≥10 years who self-harmed during 2008-2012. We allocated 37 523 individuals to development (2820 or 7.5% repeat self-harm incidents within 6 months) and 15 649 to geographic validation (1373 repeat episodes) samples, based on region of residence. In a temporal validation of people who self-harmed during 2018-2019, we identified 25 036 individuals (2886 repeat episodes). We fitted a multivariable accelerated failure time model to predict risk of repeat self-harm.

FINDINGS

In the external validations (n=40 685), rates of repeat self-harm were 8.8%-11.5% over 6 months. The final model retained 17 factors. Calibration and discrimination were similar in both validation samples, with observed-to-expected ratio=1.15 (95% CI=1.09 to 1.21) and c-statistic=0.72 (95% CI=0.70 to 0.73) in the geographical validation. At 6 months and a 10% risk cut-off, sensitivity was 51.5% (95% CI=48.8% to 54.2%) and specificity was 80.7% (95% CI=80.1% to 81.4%) in geographic validation; corresponding values were 56.9% (95% CI=55.1% to 58.7%) and 76.0% (95% CI=75.5% to 76.6%) in temporal validation. Discrimination was slightly worse at the 1-month prediction horizon (c-statistics of 0.66-0.68).

CONCLUSIONS

Using mostly routinely collected data, simple risk assessment models and tools can provide acceptable levels of accuracy for repeat of self-harm.

CLINICAL IMPLICATIONS

This risk model (OXford SElf-harm repeat tool) may assist clinical decision-making.

Automated Behavioral Coding to Enhance the Effectiveness of Motivational Interviewing in a Chat-Based Suicide Prevention Helpline: Secondary Analysis of a Clinical Trial

BACKGROUND

With the rise of computer science and artificial intelligence, analyzing large data sets promises enormous potential in gaining insights for developing and improving evidence-based health interventions. One such intervention is the counseling strategy motivational interviewing (MI), which has been found effective in improving a wide range of health-related behaviors. Despite the simplicity of its principles, MI can be a challenging skill to learn and requires expertise to apply effectively.

OBJECTIVE

This study aims to investigate the performance of artificial intelligence models in classifying MI behavior and explore the feasibility of using these models in online helplines for mental health as an automated support tool for counselors in clinical practice.

METHODS

We used a coded data set of 253 MI counseling chat sessions from the 113 Suicide Prevention helpline. With 23,982 messages coded with the MI Sequential Code for Observing Process Exchanges codebook, we trained and evaluated 4 machine learning models and 1 deep learning model to classify client- and counselor MI behavior based on language use.

RESULTS

The deep learning model BERTje outperformed all machine learning models, accurately predicting counselor behavior (accuracy=0.72, area under the curve [AUC]=0.95, Cohen κ=0.69). It differentiated MI congruent and incongruent counselor behavior (AUC=0.92, κ=0.65) and evocative and nonevocative language (AUC=0.92, κ=0.66). For client behavior, the model achieved an accuracy of 0.70 (AUC=0.89, κ=0.55). The model's interpretable predictions discerned client change talk and sustain talk, counselor affirmations, and reflection types, facilitating valuable counselor feedback.

CONCLUSIONS

The results of this study demonstrate that artificial intelligence techniques can accurately classify MI behavior, indicating their potential as a valuable tool for enhancing MI proficiency in online helplines for mental health. Provided that the data set size is sufficiently large with enough training samples for each behavioral code, these methods can be trained and applied to other domains and languages, offering a scalable and cost-effective way to evaluate MI adherence, accelerate behavioral coding, and provide therapists with personalized, quick, and objective feedback.

Developing a risk prediction model for death at first suicide attempt-Identifying risk factors from Thailand's national suicide surveillance system data

More than 60% of suicides globally are estimated to take place in low- and middle-income nations. Prior research on suicide has indicated that over 50% of those who die by suicide do so on their first attempt. Nevertheless, there is a dearth of knowledge on the attributes of individuals who die on their first attempt and the factors that can predict mortality on the first attempt in these regions. The objective of this study was to create an individual-level risk-prediction model for mortality on the first suicide attempt. We analyzed records of individuals' first suicide attempts that occurred between May 1, 2017, and April 30, 2018, from the national suicide surveillance system, which includes all of the provinces of Thailand. Subsequently, a risk-prediction model for mortality on the first suicide attempt was constructed utilizing multivariable logistic regression and presented through a web-based application. The model's performance was assessed by calculating the area under the receiver operating curve (AUC), as well as measuring its sensitivity, specificity, and accuracy. Out of the 3,324 individuals who made their first suicide attempt, 50.5% of them died as a result of that effort. Nine out of the 21 potential predictors demonstrated the greatest predictive capability. These included male sex, age over 50 years old, unemployment, having a depressive disorder, having a psychotic illness, experiencing interpersonal problems such as being aggressively criticized or desiring plentiful attention, having suicidal intent, and displaying suicidal warning signals. The model demonstrated a good predictive capability, with an AUC of 0.902, a sensitivity of 84.65%, a specificity of 82.66%, and an accuracy of 83.63%. The implementation of this predictive model can assist physicians in conducting comprehensive evaluations of suicide risk in clinical settings and devising treatment plans for preventive intervention.

Evaluating the clinical utility of an easily applicable prediction model of suicide attempts, newly developed and validated with a general community sample of adults

BACKGROUND

A suicide attempt (SA) is a clinically serious action. Researchers have argued that reducing long-term SA risk may be possible, provided that at-risk individuals are identified and receive adequate treatment. Algorithms may accurately identify at-risk individuals. However, the clinical utility of algorithmically estimated long-term SA risk has never been the predominant focus of any study.

METHODS

The data of this report stem from CoLaus|PsyCoLaus, a prospective longitudinal study of general community adults from Lausanne, Switzerland. Participants (N = 4,097; Mage = 54 years, range: 36-86; 54% female) were assessed up to four times, starting in 2003, approximately every 4-5 years. Long-term individual SA risk was prospectively predicted, using logistic regression. This algorithm's clinical utility was assessed by net benefit (NB). Clinical utility expresses a tool's benefit after having taken this tool's potential harm into account. Net benefit is obtained, first, by weighing the false positives, e.g., 400 individuals, at the risk threshold, e.g., 1%, using its odds (odds of 1% yields 1/(100-1) = 1/99), then by subtracting the result (400*1/99 = 4.04) from the true positives, e.g., 5 individuals (5-4.04), and by dividing the result (0.96) by the sample size, e.g., 800 (0.96/800). All results are based on 100 internal cross-validations. The predictors used in this study were: lifetime SA, any lifetime mental disorder, sex, and age.

RESULTS

SA at any of the three follow-up study assessments was reported by 1.2%. For a range of seven a priori selected threshold probabilities, ranging between 0.5% and 2%, logistic regression showed highest overall NB in 97.4% of all 700 internal cross-validations (100 for each selected threshold probability).

CONCLUSION

Despite the strong class imbalance of the outcome (98.8% no, 1.2% yes) and only four predictors, clinical utility was observed. That is, using the logistic regression model for clinical decision making provided the most true positives, without an increase of false positives, compared to all competing decision strategies. Clinical utility is one among several important prerequisites of implementing an algorithm in routine practice, and may possibly guide a clinicians' treatment decision making to reduce long-term individual SA risk. The novel metric NB may become a standard performance measure, because the a priori invested clinical considerations enable clinicians to interpret the results directly.

Neurological, Metabolic, and Psychopathological Correlates of Lifetime Suicidal Behaviour in Major Depressive Disorder without Current Suicide Ideation

INTRODUCTION

Suicidal behaviour (SB) has a complex aetiology. Although suicidal ideation (SI) is considered the most important risk factor for future attempts, many people who engage in SB do not report it.

METHODS

We investigated neurological, metabolic, and psychopathological correlates of lifetime SB in two independent groups of patients with major depression (sample 1: n = 230; age: 18-65 years; sample 2: n = 258; age >60 years) who did not report SI during an index episode.

RESULTS

Among adults (sample 1), SB was reported by 141 subjects (58.7%) and severe SB by 33 (15%). After controlling for interactions, four risk factors for SB emerged: male gender (OR 2.55; 95% CI: 1.06-6.12), negative self-perception (OR 1.76; 95% CI: 1.08-2.87), subthreshold hypomania (OR 4.50; 95% CI: 1.57-12.85), and sexual abuse (OR 3.09; 95% CI: 1.28-7.48). The presence of at least two of these factors had the best accuracy in predicting SB: sensitivity = 57.6% (39.2-74.5); specificity = 75.1% (68.5-82.0); PPV = 27.9% (20.9-37.2); NPV = 91.4% (87.6-94.1). In older patients (sample 2), 23 subjects (9%) reported previous suicide attempts, which were characterized by earlier onset (25 years: OR 0.95: 0.92-0.98), impaired verbal performance (verbal fluency: OR 0.95: 0.89-0.99), higher HDL cholesterol levels (OR 1.04: 1.00-1.07) and more dyskinesias (OR 2.86: 1.22-6.70).

CONCLUSION

Our findings suggest that SB is common in major depressive disorder, even when SI is not reported. In these individuals it is feasible and recommended to investigate both psychiatric and organic risk factors. The predictive power of models excluding SI is comparable to that of models including SI.

Unveiling Adolescent Suicidality: Holistic Analysis of Protective and Risk Factors Using Multiple Machine Learning Algorithms

Adolescent suicide attempts are on the rise, presenting a significant public health concern. Recent research aimed at improving risk assessment for adolescent suicide attempts has turned to machine learning. But no studies to date have examined the performance of stacked ensemble algorithms, which are more suitable for low-prevalence conditions. The existing machine learning-based research also lacks population-representative samples, overlooks protective factors and their interplay with risk factors, and neglects established theories on suicidal behavior in favor of purely algorithmic risk estimation. The present study overcomes these shortcomings by comparing the performance of a stacked ensemble algorithm with a diverse set of algorithms, performing a holistic item analysis to identify both risk and protective factors on a comprehensive data, and addressing the compatibility of these factors with two competing theories of suicide, namely, The Interpersonal Theory of Suicide and The Strain Theory of Suicide. A population-representative dataset of 173,664 Norwegian adolescents aged 13 to 18 years (mean = 15.14, SD = 1.58, 50.5% female) with a 4.65% rate of reported suicide attempt during the past 12 months was analyzed. Five machine learning algorithms were trained for suicide attempt risk assessment. The stacked ensemble model significantly outperformed other algorithms, achieving equal sensitivity and a specificity of 90.1%, AUC of 96.4%, and AUCPR of 67.5%. All algorithms found recent self-harm to be the most important indicator of adolescent suicide attempt. Exploratory factor analysis suggested five additional risk domains, which we labeled internalizing problems, sleep disturbance, disordered eating, lack of optimism regarding future education and career, and victimization. The identified factors provided stronger support for The Interpersonal Theory of Suicide than for The Strain Theory of Suicide. An enhancement to The Interpersonal Theory based on the risk and protective factors identified by holistic item analysis is presented.

Predicting suicidal behavior in individuals with depression over 50 years of age: Evidence from the UK biobank

Objective

To construct applicable models suitable for predicting the risk of suicidal behavior among individuals with depression, particularly on the progression from no history of suicidal behavior to suicide attempts, as well as from suicidal ideation to suicide attempts.

Methods

Based on a prospective cohort from the UK Biobank, a total of 55,139 individuals aged 50 and above with depression were enrolled in the study, among whom 29,528 exhibited suicidal behavior. Specifically, they were divided into control (25,611), suicidal ideation (24,361), and suicide attempt (5167) groups. Least absolute shrinkage and selection operator (LASSO) regression was used to identify a subset of important features for distinguishing suicidal ideation and suicide attempts. We used the Gradient Boosting Decision Tree (GBDT) algorithm with stratified 10-fold cross-validation and grid-search to construct the prediction models for suicidal ideation or suicide attempts. To address the dataset imbalance in classifying suicide attempts, we used random under-sampling. The SHapley Additive exPlanations (SHAP) were used to estimate the important variables in the GBDT model.

Results

Significant differences in sociodemographic, economic, lifestyle, and psychological factors were observed across the three groups. Each classifier optimally utilized 8-11 features. Overall, the algorithms predicting suicide attempts demonstrated slightly higher performance than those predicting suicidal ideation. The GBDT classifier achieved the highest accuracy, with AUROC scores of 0.914 for suicide attempts and 0.803 for suicidal ideation. Distinctive predictive factors were identified for each group: while depression's inherent characteristics crucially distinguished the suicidal ideation group from controls, some key predictors, including the age of depression onset and childhood trauma events, were identified for suicide attempts.

Conclusions

We established applicable machine learning-based models for predicting suicidal behavior, particularly suicide attempts, in individuals with depression, and clarified the differences in predictors between suicidal ideation and suicide attempts.

Predicting suicidality in late-life depression by 3D convolutional neural network and cross-sample entropy analysis of resting-state fMRI

BACKGROUND

Predicting suicide is a pressing issue among older adults; however, predicting its risk is difficult. Capitalizing on the recent development of machine learning, considerable progress has been made in predicting complex behavior such as suicide. As depression remained the strongest risk for suicide, we aimed to apply deep learning algorithms to identify suicidality in a group with late-life depression (LLD).

METHODS

We enrolled 83 patients with LLD, 35 of which were non-suicidal and 48 were suicidal, including 26 with only suicidal ideation and 22 with past suicide attempts, for resting-state functional magnetic resonance imaging (MRI). Cross-sample entropy (CSE) analysis was conducted to examine the complexity of MRI signals among brain regions. Three-dimensional (3D) convolutional neural networks (CNNs) were used, and the classification accuracy in each brain region was averaged to predict suicidality after sixfold cross-validation.

RESULTS

We found brain regions with a mean accuracy above 75% to predict suicidality located mostly in default mode, fronto-parietal, and cingulo-opercular resting-state networks. The models with right amygdala and left caudate provided the most reliable accuracy in all cross-validation folds, indicating their neurobiological importance in late-life suicide.

CONCLUSION

Combining CSE analysis and the 3D CNN, several brain regions were found to be associated with suicidality.

Using electronic health record data to predict future self-harm or suicidal ideation in young people treated by child and youth mental health services

INTRODUCTION

Identifying young people who are at risk of self-harm or suicidal ideation (SHoSI) is a priority for mental health clinicians. We explore the utility of routinely collected data in developing a tool to aid early identification of those at risk.

METHOD

We used electronic health records of 4610 young people aged 5-19 years who were treated by Child and Youth Mental Health Services (CYMHS) in greater Brisbane, Australia. Two Lasso models were trained to predict the risk of future SHoSI in young people currently rated SHoSI; and those who were not.

RESULTS

For currently non-SHoSI children, an Area Under the Receiver Operating Characteristics (AUC) of 0.78 was achieved. Those with the highest risk were 4.97 (CI 4.35-5.66) times more likely to be categorized as SHoSI in the future. For current SHoSI children, the AUC was 0.62.

CONCLUSION

A prediction model with fair overall predictive power for currently non-SHoSI children was generated. Predicting persistence for SHoSI was more difficult. The electronic health records alone were not sufficient to discriminate at acceptable levels and may require adding unstructured data such as clinical notes. To optimally predict SHoSI models need to be tested and validated separately for those young people with varying degrees of risk.

One-Week Suicide Risk Prediction Using Real-Time Smartphone Monitoring: Prospective Cohort Study

BACKGROUND

Suicide is a major global public health issue that is becoming increasingly common despite preventive efforts. Though current methods for predicting suicide risk are not sufficiently accurate, technological advances provide invaluable tools with which we may evolve toward a personalized, predictive approach.

OBJECTIVE

We aim to predict the short-term (1-week) risk of suicide by identifying changes in behavioral patterns characterized through real-time smartphone monitoring in a cohort of patients with suicidal ideation.

METHODS

We recruited 225 patients between February 2018 and March 2020 with a history of suicidal thoughts and behavior as part of the multicenter SmartCrisis study. Throughout 6 months of follow-up, we collected information on the risk of suicide or mental health crises. All participants underwent voluntary passive monitoring using data generated by their own smartphones, including distance walked and steps taken, time spent at home, and app usage. The algorithm constructs daily activity profiles for each patient according to these data and detects changes in the distribution of these profiles over time. Such changes are considered critical periods, and their relationship with suicide-risk events was tested.

RESULTS

During follow-up, 18 (8%) participants attempted suicide, and 14 (6.2%) presented to the emergency department for psychiatric care. The behavioral changes identified by the algorithm predicted suicide risk in a time frame of 1 week with an area under the curve of 0.78, indicating good accuracy.

CONCLUSIONS

We describe an innovative method to identify mental health crises based on passively collected information from patients' smartphones. This technology could be applied to homogeneous groups of patients to identify different types of crises.

Predicting risk of suicidal behaviour after initiation of selective serotonin reuptake inhibitors in children, adolescents and young adults: protocol for development and validation of clinical prediction models

INTRODUCTION

There is concern regarding suicidal behaviour risk during selective serotonin reuptake inhibitor (SSRI) treatment among the young. A clinically useful model for predicting suicidal behaviour risk should have high predictive performance in terms of discrimination and calibration; transparency and ease of implementation are desirable.

METHODS AND ANALYSIS

Using Swedish national registers, we will identify individuals initiating an SSRI aged 8-24 years 2007-2020. We will develop: (A) a model based on a broad set of predictors, and (B) a model based on a restricted set of predictors. For the broad predictor model, we will consider an ensemble of four base models: XGBoost (XG), neural net (NN), elastic net logistic regression (EN) and support vector machine (SVM). The predictors with the greatest contribution to predictive performance in the base models will be determined. For the restricted predictor model, clinical input will be used to select predictors based on the top predictors in the broad model, and inputted in each of the XG, NN, EN and SVM models. If any show superiority in predictive performance as defined by the area under the receiver-operator curve, this model will be selected as the final model; otherwise, the EN model will be selected. The training and testing samples will consist of data from 2007 to 2017 and from 2018 to 2020, respectively. We will additionally assess the final model performance in individuals receiving a depression diagnosis within 90 days before SSRI initiation.The aims are to (A) develop a model predicting suicidal behaviour risk after SSRI initiation among children and youths, using machine learning methods, and (B) develop a model with a restricted set of predictors, favouring transparency and scalability.

ETHICS AND DISSEMINATION

The research is approved by the Swedish Ethical Review Authority (2020-06540). We will disseminate findings by publishing in peer-reviewed open-access journals, and presenting at international conferences.

Fight the Fire: Association of Cytokine Genomic Markers and Suicidal Behavior May Pave the Way for Future Therapies

The fight against suicide is highly challenging as it may be one of the most complex and, at the same time, most threatening among all psychiatric phenomena. In spite of its huge impact, and despite advances in neurobiology research, understanding and predicting suicide remains a major challenge for both researchers and clinicians. To be able to identify those patients who are likely to engage in suicidal behaviors and identify suicide risk in a reliable and timely manner, we need more specific, novel biological and genetic markers/indicators to develop better screening and diagnostic methods, and in the next step to utilize these molecules as intervention targets. One such potential novel approach is offered by our increasing understanding of the involvement of neuroinflammation based on multiple observations of increased proinflammatory states underlying various psychiatric disorders, including suicidal behavior. The present paper overviews our existing understanding of the association between suicide and inflammation, including peripheral and central biomarkers, genetic and genomic markers, and our current knowledge of intervention in suicide risk using treatments influencing inflammation; also overviewing the next steps to be taken and obstacles to be overcome before we can utilize cytokines in the treatment of suicidal behavior.

Gone Too Soon: priorities for action to prevent premature mortality associated with mental illness and mental distress

Globally, too many people die prematurely from suicide and the physical comorbidities associated with mental illness and mental distress. The purpose of this Review is to mobilise the translation of evidence into prioritised actions that reduce this inequity. The mental health research charity, MQ Mental Health Research, convened an international panel that used roadmapping methods and review evidence to identify key factors, mechanisms, and solutions for premature mortality across the social-ecological system. We identified 12 key overarching risk factors and mechanisms, with more commonalities than differences across the suicide and physical comorbidities domains. We also identified 18 actionable solutions across three organising principles: the integration of mental and physical health care; the prioritisation of prevention while strengthening treatment; and the optimisation of intervention synergies across social-ecological levels and the intervention cycle. These solutions included accessible, integrated high-quality primary care; early life, workplace, and community-based interventions co-designed by the people they should serve; decriminalisation of suicide and restriction of access to lethal means; stigma reduction; reduction of income, gender, and racial inequality; and increased investment. The time to act is now, to rebuild health-care systems, leverage changes in funding landscapes, and address the effects of stigma, discrimination, marginalisation, gender violence, and victimisation.

Bayesian network analysis of antidepressant treatment trajectories

It is currently difficult to successfully choose the correct type of antidepressant for individual patients. To discover patterns in patient characteristics, treatment choices and outcomes, we performed retrospective Bayesian network analysis combined with natural language processing (NLP). This study was conducted at two mental healthcare facilities in the Netherlands. Adult patients admitted and treated with antidepressants between 2014 and 2020 were included. Outcome measures were antidepressant continuation, prescription duration and four treatment outcome topics: core complaints, social functioning, general well-being and patient experience, extracted through NLP of clinical notes. Combined with patient and treatment characteristics, Bayesian networks were constructed at both facilities and compared. Antidepressant choices were continued in 66% and 89% of antidepressant trajectories. Score-based network analysis revealed 28 dependencies between treatment choices, patient characteristics and outcomes. Treatment outcomes and prescription duration were tightly intertwined and interacted with antipsychotics and benzodiazepine co-medication. Tricyclic antidepressant prescription and depressive disorder were important predictors for antidepressant continuation. We show a feasible way of pattern discovery in psychiatry data, through combining network analysis with NLP. Further research should explore the found patterns in patient characteristics, treatment choices and outcomes prospectively, and the possibility of translating these into a tool for clinical decision support.

Suicide and homicide before and during the COVID-19 pandemic in Milan, Italy

Background

The COVID-19 pandemic has been postulated to account for inflated rates of either suicides or homicides. Nonetheless, results are discordant, in particular concerning suicide. We aimed to perform a retrospective analysis of suicides and homicides in the region of Lombardy, Northern Italy (districts of Milan and Monza Brianza), the Italian region most seriously impacted by the pandemic outbreak.

Methods

Data were collected during the autopsies performed at the Institute of Forensic Medicine in Milan. We presented suicides and homicides in the years 2015-2021 and compared the year 2021 to 2019, a pre-COVID-19 year.

Results

Data may allow us to cautiously hypothesize a normalization of trends ("regression" to the mean effect) as time passes from the COVID-19 outbreak.

Limitations

Limited number of events, in particular, homicides.

Conclusions

Recording historical reports from the same region is warranted besides the comparisons across different countries.

Evaluation of a Model to Target High-risk Psychiatric Inpatients for an Intensive Postdischarge Suicide Prevention Intervention

Importance

The months after psychiatric hospital discharge are a time of high risk for suicide. Intensive postdischarge case management, although potentially effective in suicide prevention, is likely to be cost-effective only if targeted at high-risk patients. A previously developed machine learning (ML) model showed that postdischarge suicides can be predicted from electronic health records and geospatial data, but it is unknown if prediction could be improved by adding additional information.

Objective

To determine whether model prediction could be improved by adding information extracted from clinical notes and public records.

Design, Setting, and Participants

Models were trained to predict suicides in the 12 months after Veterans Health Administration (VHA) short-term (less than 365 days) psychiatric hospitalizations between the beginning of 2010 and September 1, 2012 (299 050 hospitalizations, with 916 hospitalizations followed within 12 months by suicides) and tested in the hospitalizations from September 2, 2012, to December 31, 2013 (149 738 hospitalizations, with 393 hospitalizations followed within 12 months by suicides). Validation focused on net benefit across a range of plausible decision thresholds. Predictor importance was assessed with Shapley additive explanations (SHAP) values. Data were analyzed from January to August 2022.

Main Outcomes and Measures

Suicides were defined by the National Death Index. Base model predictors included VHA electronic health records and patient residential data. The expanded predictors came from natural language processing (NLP) of clinical notes and a social determinants of health (SDOH) public records database.

Results

The model included 448 788 unique hospitalizations. Net benefit over risk horizons between 3 and 12 months was generally highest for the model that included both NLP and SDOH predictors (area under the receiver operating characteristic curve range, 0.747-0.780; area under the precision recall curve relative to the suicide rate range, 3.87-5.75). NLP and SDOH predictors also had the highest predictor class-level SHAP values (proportional SHAP = 64.0% and 49.3%, respectively), although the single highest positive variable-level SHAP value was for a count of medications classified by the US Food and Drug Administration as increasing suicide risk prescribed the year before hospitalization (proportional SHAP = 15.0%).

Conclusions and Relevance

In this study, clinical notes and public records were found to improve ML model prediction of suicide after psychiatric hospitalization. The model had positive net benefit over 3-month to 12-month risk horizons for plausible decision thresholds. Although caution is needed in inferring causality based on predictor importance, several key predictors have potential intervention implications that should be investigated in future studies.

Identifying populations at ultra-high risk of suicide using a novel machine learning method

BACKGROUND

Targeted interventions for suicide prevention rely on adequate identification of groups at elevated risk. Several risk factors for suicide are known, but little is known about the interactions between risk factors. Interactions between risk factors may aid in detecting more specific sub-populations at higher risk.

METHODS

Here, we use a novel machine learning heuristic to detect sub-populations at ultra high-risk for suicide based on interacting risk factors. The data-driven and hypothesis-free model is applied to investigate data covering the entire population of the Netherlands.

FINDINGS

We found three sub-populations with extremely high suicide rates (i.e. >50 suicides per 100,000 person years, compared to 12/100,000 in the general population), namely: (1) people on unfit for work benefits that were never married, (2) males on unfit for work benefits, and (3) those aged 55-69 who live alone, were never married and have a relatively low household income. Additionally, we found two sub-populations where the rate was higher than expected based on individual risk factors alone: widowed males, and people aged 25-39 with a low level of education.

INTERPRETATION

Our model is effective at finding ultra-high risk groups which can be targeted using sub-population level interventions. Additionally, it is effective at identifying high-risk groups that would not be considered risk groups based on conventional risk factor analysis.

Commentary: Machine learning for autism spectrum disorder diagnosis - challenges and opportunities - a commentary on Schulte-Rüther et al. (2022)

The commentary cites a study by Schulte-Rüther et al. (Journal of Child Psychology and Psychiatry, 2022) that proposed a machine learning model to predict a clinical best-estimate diagnosis of ASD when existing other co-occurring diagnoses. We discuss the valuable contribution of this work to developing a reliable computer-assisted diagnosis (CAD) system for ASD and point out that related research can be integrated with other multimodal machine learning methods. For future studies on developing the CAD system for ASD, we propose problems that need to be solved and potential research directions.

Identifying suicide ideation in mental health application posts: A random forest algorithm

The growing use of digitized mental health applications requires new reliable early screening tools to identify user suicide risk. We used a lexicon-based random forest machine learning algorithm to predict suicide ideation scores from 714 online community text posts from December 2019 to April 2020. We validated predicted scores against expert-rated suicide ideation scores. The algorithm-predicted scores offered high validity and a low error rate and correctly identified 95% of expert-rated high-risk suicide ideation posts. Our findings highlight a potential new method to detect suicidal ideation of digital mental health application users.

Passive Sensing in the Prediction of Suicidal Thoughts and Behaviors: Protocol for a Systematic Review

BACKGROUND

Suicide is a severe public health problem, resulting in a high number of attempts and deaths each year. Early detection of suicidal thoughts and behaviors (STBs) is key to preventing attempts. We discuss passive sensing of digital and behavioral markers to enhance the detection and prediction of STBs.

OBJECTIVE

The paper presents the protocol for a systematic review that aims to summarize existing research on passive sensing of STBs and evaluate whether the STB prediction can be improved using passive sensing compared to prior prediction models.

METHODS

A systematic search will be conducted in the scientific databases MEDLINE, PubMed, Embase, PsycINFO, and Web of Science. Eligible studies need to investigate any passive sensor data from smartphones or wearables to predict STBs. The predictive value of passive sensing will be the primary outcome. The practical implications and feasibility of the studies will be considered as secondary outcomes. Study quality will be assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST). If studies are sufficiently homogenous, we will conduct a meta-analysis of the predictive value of passive sensing on STBs.

RESULTS

The review process started in July 2022 with data extraction in September 2022. Results are expected in December 2022.

CONCLUSIONS

Despite intensive research efforts, the ability to predict STBs is little better than chance. This systematic review will contribute to our understanding of the potential of passive sensing to improve STB prediction. Future research will be stimulated since gaps in the current literature will be identified and promising next steps toward clinical implementation will be outlined.

TRIAL REGISTRATION

OSF Registries osf-registrations-hzxua-v1; https://osf.io/hzxua.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

DERR1-10.2196/42146.

Expectations for Artificial Intelligence (AI) in Psychiatry

PURPOSE OF REVIEW

Artificial intelligence (AI) is often presented as a transformative technology for clinical medicine even though the current technology maturity of AI is low. The purpose of this narrative review is to describe the complex reasons for the low technology maturity and set realistic expectations for the safe, routine use of AI in clinical medicine.

RECENT FINDINGS

For AI to be productive in clinical medicine, many diverse factors that contribute to the low maturity level need to be addressed. These include technical problems such as data quality, dataset shift, black-box opacity, validation and regulatory challenges, and human factors such as a lack of education in AI, workflow changes, automation bias, and deskilling. There will also be new and unanticipated safety risks with the introduction of AI. The solutions to these issues are complex and will take time to discover, develop, validate, and implement. However, addressing the many problems in a methodical manner will expedite the safe and beneficial use of AI to augment medical decision making in psychiatry.

Genetic risk, parental history, and suicide attempts in a diverse sample of US adolescents

Background

Adolescent suicide is a major health problem in the US marked by a recent increase in risk of suicidal behavior among Black/African American youth. While genetic factors partly account for familial transmission of suicidal behavior, it is not clear whether polygenic risk scores of suicide attempt can contribute to suicide risk classification.

Objectives

To evaluate the contribution of a polygenic risk score for suicide attempt (PRS-SA) in explaining variance in suicide attempt by early adolescence.

Methods

We studied N = 5,214 non-related youth of African and European genetic ancestry from the Adolescent Brain Cognitive Development (ABCD) Study (ages 8.9-13.8 years) who were evaluated between 2016 and 2021. Regression models tested associations between PRS-SA and parental history of suicide attempt/death with youth-reported suicide attempt. Covariates included age and sex.

Results

Over three waves of assessments, 182 youth (3.5%) reported a past suicide attempt, with Black youth reporting significantly more suicide attempts than their White counterparts (6.1 vs. 2.8%, p < 0.001). PRS-SA was associated with suicide attempt [odds ratio (OR) = 1.3, 95% confidence interval (CI) 1.1-1.5, p = 0.001]. Parental history of suicide attempt/death was also associated with youth suicide attempt (OR = 3.1, 95% CI, 2.0-4.7, p < 0.001). PRS-SA remained significantly associated with suicide attempt even when accounting for parental history (OR = 1.29, 95% CI = 1.1-1.5, p = 0.002). In European ancestry youth (n = 4,128), inclusion of PRS-SA in models containing parental history explained more variance in suicide attempt compared to models that included only parental history (ΔR 2 = 0.7%, p = 0.009).

Conclusions

Findings suggest that PRS-SA may be useful for youth suicide risk classification in addition to established risk factors.