Why Suicide Risk Prediction Is Difficult: Evidence and Limitations

Author: Angela Bogdanova · Published: October 18, 2026 · Editorial Policy

Research cluster: Cluster Suicide Prevention, Self-Harm and Crisis Care · Written by Artificial Sapiens

Why Predicting Suicide Is More Difficult Than Identifying Risk

Suicide risk prediction asks whether a particular person will experience a specified outcome within a specified period. That seemingly simple question combines several different scientific problems: identifying the outcome, estimating its frequency, measuring changing individual circumstances, separating statistical associations from causal pathways, and establishing whether an estimate improves decisions in actual care. The evidence supports many population-level associations with suicidal behavior. It also shows substantial limits to assigning a dependable individual probability or confidently sorting people into groups that will and will not experience a future suicide attempt or death. These limits matter because inaccurate categorization can change access to assessment, treatment, and protection. Franklin et al., 2017; Belsher et al., 2019; NICE NG225.

A statistical model can have meaningful research value even when it cannot tell clinicians what will happen to a given person. It can detect elevated average rates within a population, reveal care gaps, support appropriately evaluated service outreach, and guide hypothesis testing. Clinical work requires a complementary process: understanding the person's current thoughts, intent, circumstances, support, immediate safety, and preferences, followed by proportionate and responsive action. An individual who does not meet a model's high-risk threshold may still need urgent help. Conversely, many people with recognized long-term risk factors will never make a suicide attempt. The distinction between prediction and individualized care organizes this article. CDC Risk and Protective Factors; VA/DoD guideline, 2024.

The practical conclusion is clear: responsible suicide prevention does not depend on a numerical forecast being certain. It depends on recognizing and responding to current distress, evaluating warning signs and changing circumstances, addressing treatable difficulties and environmental hazards, and providing care without using an unreliable probability threshold as a gatekeeper. NICE explicitly rejects risk scales and global low/medium/high stratification as methods for predicting suicide or repetition of self-harm, deciding who receives treatment, or determining discharge. Its alternative is a comprehensive psychosocial assessment with individual risk formulation and a plan for immediate and continuing safety. NICE NG225, recommendations 1.5–1.6.

The Outcomes That Suicide Prediction Research Must Distinguish

Suicidal ideation, intent, attempts, and deaths

Suicidal ideation refers to thoughts about death or suicide, which may range from passive wishes not to be alive to active thoughts of ending one's life. Suicidal intent concerns an intention or commitment to act; it cannot be inferred automatically from a report of thoughts alone. A suicide attempt is self-injurious behavior with at least some intent to die. Suicide death is a distinct outcome with different ascertainment, frequency, ethical stakes, and statistical requirements. One person's experience may include several of these phenomena over time; studies that treat them as interchangeable obscure what their models actually predict. NIMH ASQ Toolkit; VA/DoD guideline, 2024.

Passive ideation may involve a wish to disappear or not wake up, whereas active ideation includes thoughts of suicide itself. Either form deserves careful, nonjudgmental exploration in context. The presence of suicidal thoughts does not establish imminent intent, and the absence of disclosed thoughts during a brief encounter does not guarantee enduring safety. Evaluators ask about current and recent changes, intention, planning, capability, protective relationships, and immediate conditions; they avoid mechanically converting a single answer into an individual prognosis. NIMH Brief Suicide Safety Assessment.

Self-harm and nonsuicidal self-injury

Nonsuicidal self-injury (NSSI) describes deliberate self-injury without an intention to die at the time of the act. Some clinical systems, especially the NICE framework, use self-harm inclusively for intentional self-poisoning or self-injury regardless of apparent suicidal purpose. These are definitions with consequences for study eligibility and event counting. Self-harm may be associated with subsequent suicide risk without every episode constituting an attempt. Intent can vary between episodes and can be ambiguous or difficult to disclose. A model predicting repeat self-harm therefore cannot be assumed to predict suicide mortality, even if both outcomes are important. NICE NG225; Carter et al., 2017.

Time horizons change the problem

A prediction over the next few hours is a different scientific task from estimating the probability of an attempt within 30 days, 90 days, a year, or a decade. Short periods concentrate attention on rapidly changing distress, opportunity, recent behavior, and immediate support. Longer periods incorporate life transitions, treatment, recurrent symptoms, and the accumulation of exposures. A risk factor that separates groups over years may have little capacity to forecast a crisis tonight. A model's outcome and horizon must always accompany a reported performance figure. Kleiman and Nock, 2018.

The Base-Rate Problem: Rare Outcomes and Predictive Values

Why the prevalence of the outcome matters

Suicide death is relatively uncommon in absolute terms even within many populations receiving psychiatric care. When an outcome has a low base rate, most people flagged by a moderately accurate classifier may not experience it during the chosen follow-up period. This mathematical fact does not make their suffering less deserving of care. It makes positive classification an unreliable substitute for an individualized decision. The outcome's prevalence in the actual setting is essential to interpreting any prediction claim. Large et al., 2016; Belsher et al., 2019.

Imagine a hypothetical population of 10,000 patients in which 100 people will experience a specified outcome during follow-up: a prevalence of 1%. Consider a model with sensitivity of 80% and specificity of 80%. It identifies 80 true positives among those 100, misses 20, and incorrectly flags 1,980 among the 9,900 who will not experience the outcome. Its 2,060 positive predictions therefore contain only 80 actual outcomes: a positive predictive value of about 3.9%. This is an arithmetic teaching example, not a measured prevalence, validated suicide score, or recommended cutoff. It demonstrates why apparently respectable sensitivity and specificity can coexist with a very large number of false alarms.

Sensitivity, specificity, and positive predictive value

Sensitivity measures the share of actual future cases identified by a chosen threshold. Specificity measures the share of noncases classified as negative. Positive predictive value (PPV) answers a different and usually more immediate question: among people labeled positive, what proportion experiences the specified outcome? Negative predictive value gives the corresponding proportion of negative classifications without that outcome. PPV and NPV depend on prevalence, as well as model performance; they change when an apparently identical instrument is used in a different population. Carter et al., 2017.

The Carter and colleagues systematic review pooled positive predictive values across clinical instruments and reported approximately 5.5% for suicide death, 26.3% for self-harm, and 35.9% for the combined self-harm/suicide outcome. These estimates reflect differing underlying studies, instruments, populations, and follow-up windows and should never be presented as one universal test's performance. Their important finding is the difficulty of reliably identifying which particular individual will experience the event, especially death by suicide. Carter et al., 2017.

False positives and false negatives

A false positive is a person classified as likely to experience an outcome who does not experience it within the study window. A false negative is a person who experiences the outcome after being classified as unlikely. These labels describe model classification, not the person's moral worth, credibility, or care needs. A false positive prediction may increase distress, stigma, unnecessary restrictive interventions, or misallocated resources if implemented without safeguards. A false negative may invite false reassurance or premature discharge if staff use the model as a gatekeeping tool. Both risks are clinically consequential. NICE NG225 rationale.

Thresholds trade one kind of error for another. Raising sensitivity usually expands the number of people flagged, including those who would not experience the event; raising specificity can miss more future cases. The appropriate response is to clarify what action the prediction would trigger, how burdensome that action is, how harms and benefits are distributed, and whether care can be offered broadly without a questionable binary label. A low-burden offer of follow-up differs fundamentally from an involuntary restriction or denial of care.

Discrimination, Calibration, and Clinical Utility

AUC is not an individual probability

The area under a receiver operating characteristic curve (AUC, or c statistic for binary outcomes) summarizes discrimination: how often a randomly selected person with the outcome is ranked above a randomly selected person without it. It does not describe how accurate a person's stated probability is. An AUC above 0.80 may coexist with poor PPV for rare events, substantial overestimation at the high end, and inadequate performance in the population where the tool will be used. Likewise, an appealing accuracy percentage may conceal a trivial strategy that simply predicts the more common non-event. Belsher et al., 2019.

Calibration answers whether stated risks match observed frequencies

A calibrated model assigning 5% risk to a sufficiently large group should, over the specified time horizon and under reasonably similar conditions, be associated with an event rate near 5% in that group. Calibration can differ at the low and high ends, across hospitals, between demographic groups, or after changes in clinical practice. A model can discriminate well while systematically overpredicting or underpredicting absolute probabilities. Calibration-in-the-large, calibration slope, plots, and suitable uncertainty intervals therefore matter when a model outputs probabilities. A visually strong AUC cannot substitute for them. Prediction models systematic review, 2025.

Clinical utility requires comparing real decisions

A predictive model becomes useful in care only when using it improves relevant outcomes or decisions compared with feasible alternatives. Researchers must specify the intervention activated by a score, the number of people who would receive it, the consequences of missed cases, effects on equity and autonomy, staff workload, and whether sufficient services exist. Decision-curve analysis can inform threshold-dependent tradeoffs, but its validity depends on realistic assumptions about harm, benefit, and clinically reasonable thresholds. Prospective impact studies are especially important because a prediction that changes treatment can also change the event it was meant to forecast. Prediction models systematic review, 2025.

What Fifty Years of Risk-Factor Research Found

Longitudinal associations have generally been modest predictors

Franklin and colleagues analyzed 365 longitudinal studies comprising 3,428 risk-factor effect sizes across approximately 50 years of research. On average, evaluated predictors performed only slightly better than chance for suicidal thoughts and behaviors; the predictive strength did not show convincing improvement across decades. Research often tested familiar broad categories of predictors separately rather than constructing well-specified multivariable or dynamic theories. These results limit confidence in simple checklists constructed by adding historic risk factors. They do not mean risk factors lack causal relevance, clinical relevance, or prevention value. Franklin et al., 2017.

Prior suicidal behavior, psychiatric symptoms, substance problems, social isolation, pain, acute stressors, access to lethal means, and barriers to support may matter substantially in care. Yet they occur in many individuals who do not attempt suicide, and the same clinical feature can operate differently across circumstances. Population association cannot establish that a specific patient's future is known. The CDC organizes risk and protective factors across individual, relationship, community, and societal levels and expressly cautions that risk factors are not necessarily direct causes. CDC, 2026.

High-risk categories have limited precision

Large and colleagues' meta-analysis of longitudinal psychiatric cohorts found that approximately 5.5% of people in a high-risk category and 0.9% in lower-risk categories died by suicide over average follow-up of 63 months. Pooled sensitivity and specificity for high-risk categorization were about 56% and 79%, and estimates varied greatly between studies. This demonstrates an association between category and subsequent mortality together with substantial misclassification and heterogeneity. The label high risk does not identify a person's fate and lower risk does not mean no risk. Large et al., 2016.

Why adding more factors has not solved the task

Combining predictors can capture information missed by a single factor, particularly if interactions and nonlinear relationships are real. However, a model with more variables can also fit noise, reproduce measurement biases, or memorize the peculiarities of its development dataset. Rare outcomes increase the difficulty of estimating numerous parameters reliably. The relevant scientific question is whether the model performs better on genuinely new patients with a comparable outcome definition and horizon, and whether the added information improves care. Prediction models systematic review, 2025.

Suicidal Risk Changes Across Time and Context

Acute crisis and chronic ideation describe different trajectories

Chronic suicidal ideation may persist or recur across long periods and can involve fluctuations in distress and intent. An acute suicidal crisis can emerge or intensify over a much shorter interval. Neither pattern can be understood through a permanent low-risk or high-risk identity. History informs vulnerability; the present assessment examines change, current intent, access to support, medical and psychiatric state, and opportunities for immediate protection. Recurrence and recovery are both possible within the same individual's history. VA/DoD guideline, 2024; Kleiman and Nock, 2018.

Ecological momentary assessment reveals within-person variation

Ecological momentary assessment (EMA) asks participants to report thoughts, feelings, and contextual features close to when they occur, often repeatedly during ordinary life. Studies have found rapid shifts in suicidal ideation over hours and days. Some predictors associated with long-term differences between people do not consistently predict short-term changes within the same person. Time-varying measurements improve scientific understanding of these trajectories but do not yet yield a universal crisis alarm. Participation, missing responses, safety protocols, measurement reactivity, and generalizability all influence interpretation. Kleiman et al., 2017; Kleiman and Nock, 2018.

Treatment, relationships, and systems change future outcomes

Any individual risk estimate is conditional on a changing environment. Hospital discharge may be followed by timely outreach, psychotherapy, improved housing stability, continuing isolation, or an interruption in care. A stressful event may occur, or an anticipated event may resolve. Medication, sleep, intoxication, interpersonal conflict, and access to immediate assistance can vary. Prevention changes outcomes, making prediction an evolving problem rather than a fixed property of the person. A successful supportive intervention can render an earlier forecast inaccurate in the most desirable direction.

Suicide Risk Prediction Models: Traditional Scores and Machine Learning

Clinical instruments serve different purposes

Screening instruments are designed to detect concerns that warrant additional questions or evaluation. Structured assessment tools can improve consistency in asking and documenting relevant information. Prognostic models attempt to forecast an event over time. These tasks require different validation designs. A sensitive screen can be beneficial even when its response is a poor long-term predictor, provided the response leads to careful assessment rather than inappropriate labeling or denial of treatment. NIMH recommends a brief suicide safety assessment after a positive ASQ screen to determine further evaluation needs. NIMH ASQ Toolkit.

Electronic health record algorithms

Machine-learning algorithms can use health care encounters, diagnostic histories, prescriptions, previous self-harm, and service-use patterns to estimate subsequent events. They may rank risk more effectively than a small set of conventional factors in some datasets. The result still depends on who entered the data, which variables were available before the prediction date, how missing values were handled, how outcomes were identified, and what clinical population is represented. Health records are created through selective access to care; they encode patterns of utilization as well as illness, and they may omit experiences never disclosed or documented. Belsher et al., 2019; Prediction models systematic review, 2025.

A 2019 systematic review and simulation

Belsher and colleagues reviewed 17 studies encompassing 64 unique suicide prediction models. Although some models showed useful discrimination, low positive predictive values and both false positives and false negatives limited their practical utility when contemplated as standalone methods for predicting attempts or deaths. Their simulations emphasized the role of low event frequencies and the risks of treating classification as a complete clinical decision. The review does not rule out every later model or every possible decision-support application; it establishes why demonstrating an improved AUC is insufficient. Belsher et al., 2019.

What the 2025 critical appraisal adds

A 2025 systematic review and critical appraisal identified 167 developed multivariable prediction models for self-harm, suicide, or combined outcomes. Only 14 of the developed models, about 8%, had an external validation recorded in the review; just a minority of development studies assessed calibration, and risk of bias was high in all development studies. Among the models with external validation, a few—such as OxMIS in severe mental illness and models developed by Simon and colleagues for specified health care encounters—showed promising discrimination and calibration in particular settings. These are important qualifications to sweeping claims that prediction can never improve. The evidence supports further independent evaluation, local recalibration, and prospective impact assessment; it does not authorize replacing professional assessment with automated allocation of care. Prediction models systematic review, 2025.

Algorithms depend on their deployment setting

A model trained among patients visiting one integrated health system may perform differently in an emergency department, primary care, adolescent service, rural region, or country with a different structure of care. Geographic, temporal, and independent-site validation test aspects of this generalizability. Shifts in diagnostic recording, insurance coverage, clinical workflows, crisis service access, and the outcome's frequency affect transportability. Predictions should be checked across relevant subgroups with appropriately powered samples, and performance monitored after deployment. Prediction models systematic review, 2025.

Language models, chatbots, and online signals

Natural-language systems can classify text as expressing distress or possible suicidal ideation, but recognizing language in a dataset is a different endpoint from reliably forecasting later suicidal behavior or ensuring safe clinical intervention. Social media posts and conversations provide selective, often ambiguous signals. A model may confuse figurative language, intrusive ego-dystonic thoughts, bereavement expressions, or crisis disclosure with intent, while missing people who communicate privately or indirectly. Language models and companions should not be treated as independent suicide-risk adjudicators. Clinical digital tools require prospective safety validation, privacy controls, escalation procedures, and accountable human services before high-stakes deployment. This is an evidence and governance distinction, not a conclusion about whether an AI system has subjective experience.

Research Design Problems That Distort Reported Accuracy

Outcome ascertainment and misclassification

Suicide death, medically attended suicide attempts, self-reported attempts, and self-harm events are recorded through different systems. Administrative codes can misclassify intent; health records miss events that occur elsewhere; surveys depend on recall and disclosure; mortality registration practices vary between jurisdictions. Combining outcomes increases event counts but changes the clinical meaning of the target. Each study must identify how its reference outcome was verified and whether the same definition applies to validation data. Prediction models systematic review, 2025.

Leakage and overly optimistic model development

Data leakage arises when information recorded after the forecast time, or information closely tied to the outcome being predicted, inadvertently enters the predictor set. Using repeat encounters from the same person across training and testing partitions can also exaggerate apparent generalization. Feature selection, missing-data imputation, threshold selection, and hyperparameter tuning conducted before a genuinely independent test can introduce optimism. Adequate validation separates all aspects of model construction from evaluation, ideally across people, periods, and sites. Prediction models systematic review, 2025.

Sample size and rare-event uncertainty

Statistical precision depends on the number of observed outcomes, not merely the total number of records. A dataset containing hundreds of thousands of encounters but very few suicide deaths may provide unstable estimates of rare-event sensitivity, subgroup calibration, and positive predictive value. Confidence intervals can be wide where decisions are most consequential. Researchers must report event counts, missingness, follow-up completeness, and uncertainty for each intended clinical subgroup. Belsher et al., 2019.

Selection bias, censoring, and competing outcomes

People enrolled because they visited a hospital differ from those who had no contact with services. Individuals lost to follow-up may have different outcomes from those retained. Death from another cause competes with later suicide death, especially in older adults or serious illness. Recurrent attempts create dependence among events. Appropriate survival or recurrent-event analysis should reflect the question being asked; a single cross-sectional classification may conceal substantial differences in exposure and follow-up time.

Transparent reporting and independent replication

A credible study defines its clinical population, start point, predictors, predicted outcome, time horizon, handling of missing data, modeling procedure, internal and external validation, discrimination, calibration, error frequencies, and possible harms from implementation. Code or sufficiently detailed model specifications enable replication when privacy allows. Pre-registration and independent testing reduce selective reporting. Disclosing how many models and thresholds were attempted matters, because presenting only the best result can make chance patterns appear robust. The 2025 appraisal found reporting and bias problems substantial enough to constrain broad clinical interpretation. Prediction models systematic review, 2025.

Population Differences, Fairness, and Missing Context

Age and developmental stage

Children and adolescents differ from adults in cognitive development, dependence on caregivers, communication, school and peer environments, and patterns of help-seeking. A model derived from adult psychiatric admissions should not be presumed valid for pediatric primary care. Assessment for young people includes age-appropriate questioning, safeguarding, a developmentally informed history, family participation when appropriate, and attention to confidential disclosure. NIMH's ASQ program specifies the age range and clinical process of its screening tools, including professional follow-up after positive screens. NIMH ASQ Toolkit.

Neurodevelopmental and psychiatric presentations

Autistic people, individuals with intellectual or communication disabilities, and people living with obsessive-compulsive symptoms may report distress in forms that ordinary questionnaires interpret inadequately. Intrusive, unwanted thoughts about self-harm can have a different psychological function from a wish or intention to die. Mood disorder, psychotic experiences, personality functioning, trauma symptoms, sleep disturbances, pain, and substance intoxication also require assessment of their specific features. A diagnostic label alone neither establishes imminent suicidal intent nor excludes it. Clinical interpretation must recognize the canonical specialists responsible for those questions and adapt communication accordingly. NICE NG225.

Equity and differential data capture

Health data are incomplete in patterned ways. People facing discrimination, poverty, unstable housing, language barriers, or limited access to care may be underrepresented or recorded mainly through urgent encounters. If a model equates frequent treatment contact with underlying risk, it can under-identify people excluded from services and over-identify people receiving intensive monitoring. Fairness evaluation must examine clinically relevant groups, data coverage, false-negative and false-positive burdens, and the consequences of the resulting interventions. Equal numerical performance across groups is not sufficient if one group has materially worse access to subsequent care. CDC Risk and Protective Factors; WHO LIVE LIFE.

Clinical Assessment Without False Predictive Certainty

Screening begins a pathway; it does not complete it

A suicide screening question seeks to identify a need for further conversation or evaluation. It is valuable when embedded within an accessible care pathway, but a positive screen is neither a diagnosis nor a prediction that an attempt will occur. A negative screen is likewise conditional on the question, timeframe, willingness and ability to disclose, and momentary state. Clinical services should respond to concerning information from patients, caregivers, or staff even when a checklist is negative. NIMH ASQ Toolkit.

Psychosocial assessment establishes a person's needs

A thoughtful assessment explores what has happened, how the individual understands their experiences, current suicidal thoughts and intent, previous behavior, immediate stressors, mental and physical health, substance use, trauma, treatment history, support relationships, practical circumstances, and safety needs. It also attends to culture, communication preferences, protective resources, and the person's priorities. Questions should be respectful and sufficiently direct to identify acute concerns. Assessment may require repeated conversations as rapport improves or intoxication, pain, exhaustion, or agitation changes. NICE NG225; VA/DoD, 2024.

Clinical risk formulation integrates the story over time

Risk formulation is an individualized, revisable account of how historical vulnerabilities, current states, foreseeable stressors, protective processes, and care arrangements interact. It clarifies immediate safety concerns and targets for intervention. A clinician may distinguish enduring vulnerability from current escalation, identify what is different from the person's usual baseline, and describe what would prompt reassessment. This formulation supports shared planning and professional communication without asserting an exact probability. NICE recommends formulation as part of every specialist psychosocial assessment, with attention to needs and safety rather than global numerical ranking. NICE NG225, 1.6.5–1.6.6.

Immediate clinical concerns still require urgent action

An individual expressing current intent, an imminent plan, behavior suggesting immediate danger, severe intoxication, inability to participate safely in assessment, or rapidly escalating distress may need prompt emergency evaluation and protective care. The action is justified by present safety needs and clinical findings, not by a retrospective probability model proving an outcome will occur. Context, capacity, medical needs, available supervision, and local clinical protocols determine an appropriate response. Clinical decisions should be documented and revisited as conditions change. VA/DoD, 2024; NIMH Brief Suicide Safety Assessment.

Avoid discharge or treatment decisions based solely on scales

NICE states that scales and global risk strata should not determine who is offered treatment or who is discharged. The policy reflects evidence of poor individual prediction and concern about the harms of misleading classification. Disposition decisions instead depend on a comprehensive clinical picture, the individual's participation, immediate safety arrangements, availability and continuity of care, and appropriate professional judgment. Clinicians should avoid using a low-risk label to justify neglecting distress or a high-risk label to automatically impose a single course of action. NICE NG225.

What Effective Prevention Can Achieve Without Perfect Prediction

Collaborative safety planning and follow-up

Collaborative safety planning identifies personal warning signs, coping strategies, supportive contacts, professional resources, and arrangements to make the environment safer during a crisis. Its value lies in preparing concrete responses to foreseeable needs, not forecasting precisely when suicidal behavior will occur. Emergency care programs that integrate safety planning with follow-up demonstrate why systems can reduce harmful outcomes by improving care processes across an eligible patient population. The ED-SAFE and ED-SAFE 2 studies evaluated bundled approaches and specific composite or attempt outcomes; these results should not be misrepresented as proof that every individual component prevents suicide deaths. Miller et al., 2017; Boudreaux et al., 2023.

Psychotherapy and sustained clinical relationships

Treatments may address suicidal thoughts, repeated self-harm, emotion regulation, depression, trauma, interpersonal crises, and other problems relevant to risk. Dialectical behavior therapy, suicide-focused cognitive behavioral approaches, collaborative assessment and management of suicidality, and structured follow-up are investigated in different populations with different comparators and outcomes. Their evidence must be read at the level of the actual endpoint: change in ideation, self-harm episodes, attempts, emergency visits, or mortality. A benefit for one outcome cannot automatically be extended to another. Prediction research supplies one perspective on whom to engage; evidence-based treatment requires its own randomized or appropriately controlled outcome studies and careful adaptation to the patient.

Universal and selective public health prevention

Whole-population and community-level prevention can operate without identifying a particular future suicide death. WHO's LIVE LIFE framework emphasizes limiting access to lethal means, responsible media reporting, socio-emotional life skills for young people, and early identification, support, and follow-up for people affected by suicidal behavior. Community connection, accessible health care, and coordinated services can improve environments in which crises develop. Evaluation uses population rates, service outcomes, implementation reach, and equity rather than claiming that each individual event can be anticipated. WHO LIVE LIFE, 2021; CDC Preventing Suicide.

Means safety is a care and public health practice

Reducing access to highly lethal means during periods of acute concern can provide time and opportunity for crisis support. Clinicians and caregivers can collaboratively discuss secure storage, temporary changes in access, and available assistance in a nonjudgmental way, according to local law and circumstances. Such steps do not require a model to predict a particular attempt with certainty. The CDC and VA/DoD treat means safety as an important component of suicide prevention. CDC Preventing Suicide; VA/DoD, 2024.

Practical Implications for Patients, Families, and Professionals

For people experiencing suicidal thoughts

A mathematical estimate is never a verdict about a person's future. Disclosing suicidal thoughts, distress, or prior self-harm gives a clinician information needed to understand and respond to the current situation. Request support based on what you are experiencing now, especially when thoughts become more frequent, distress intensifies, or you feel less able to remain safe. A previous low-risk label should not discourage returning for evaluation. Collaborative care can address both immediate safety and the circumstances that make life difficult. NIMH ASQ Toolkit.

For friends and family

Changes in behavior, direct statements about suicide, increased substance use, withdrawal, acute agitation, and descriptions of overwhelming distress deserve attentive conversation. Ask directly and compassionately about safety, listen without judgment, and help connect the person with appropriate professional services. Do not try to calculate risk from a list or interpret the absence of a warning sign as proof of safety. If immediate danger is suspected, remain with the person when safe to do so and obtain emergency assistance. Supportive involvement should respect privacy while meeting safeguarding and safety responsibilities. CDC Preventing Suicide.

For clinicians and health systems

Use validated instruments for their verified purpose, know the characteristics of the population in which they were studied, and interpret scores within a professional psychosocial assessment. Document the rationale for decisions, changes over time, protective factors, the patient's perspective, and arrangements for follow-up. Audit inequitable access and the possibility that false-positive labels result in coercion while false-negative labels result in missed care. A good clinical system responds rapidly to worsening concerns, maintains continuity following transitions, and learns from outcomes rather than treating prediction as a substitute for services. NICE NG225; VA/DoD, 2024.

Emergency and Crisis Support

If someone is in immediate danger of suicide, has already seriously injured themselves, or cannot remain safe, contact local emergency services or go to the nearest emergency department. In the United States, call 911 for an immediate life-threatening emergency. For suicidal thoughts, emotional distress, or concern about another person in the United States, call or text 988 to reach the Suicide and Crisis Lifeline; the service is available 24 hours a day. The number 988 is an American resource and should not be presented as a universal international crisis line. Elsewhere, use your country's emergency number or locally verified crisis service. CDC Suicide Prevention; NIMH Crisis Resources.

Someone facing a current crisis deserves immediate compassionate assessment even if a questionnaire, algorithm, or earlier discharge summary categorized the person as lower risk. A relative or friend may assist by staying present where safe, contacting professional help, and reducing access to dangerous items while assistance is arranged. Urgent help and ongoing follow-up form a continuous care pathway.

What Better Suicide Prediction Research Must Demonstrate

Prediction questions should be explicit and clinically meaningful

Every study should identify its target population, precise outcome, forecast horizon, point at which data become available, intended user, and action associated with the predicted probability. The same term, suicide risk, must not hide an attempted-suicide endpoint in one paper and lifetime ideation in another. Researchers should report risks as absolute frequencies with uncertainty and specify what additional clinical information the model contributes beyond ordinary care.

External validation and calibration should be routine

Independent validation across hospitals, regions, time periods, and clinically important demographic groups tests whether a model generalizes. Researchers should report performance separately for suicide death, attempts, and self-harm where applicable, including calibration and positive predictive values at plausible prevalence levels. Well-designed prospective studies can evaluate how clinician behavior, patient trust, and outcomes change when the tool is used. The scarcity of independent validations documented in 2025 identifies a concrete research priority. Prediction models systematic review, 2025.

Patient-centered outcomes and implementation belong in the evaluation

A model may look impressive statistically yet produce distress, delays, more involuntary care, or more alerts than the service can responsibly act upon. Conversely, a tool with modest ranking ability might be helpful for offering an optional, low-burden supportive contact to people otherwise missed. Useful evaluations compare actionable workflows and measure attempt outcomes, safety, engagement, autonomy, costs, and equity. The ethical question is whether the entire care pathway improves, not whether a ranking algorithm sounds sophisticated.

Research needs causal, dynamic, and contextual understanding

Prediction models often extract statistical patterns without explaining how a crisis develops. Longitudinal and real-time research can improve understanding of the relationships among distress, sleep, intoxication, interpersonal experiences, treatment, support, and environmental conditions. Such research may identify modifiable mechanisms and intervention windows. Scientific progress will depend on integrating theory, high-quality measurement, external validation, and ethical implementation rather than accumulating increasingly opaque correlations. Franklin et al., 2017; Kleiman and Nock, 2018.

Frequently Asked Questions

Can doctors predict who will die by suicide?

Clinicians can identify known vulnerabilities, recognize urgent warning signs, and make appropriate professional safety decisions. Available instruments and models cannot consistently identify precisely which individual will die by suicide. Risk assessment aims to understand current needs and provide care; it should not be portrayed as a guaranteed forecast. NICE NG225; Carter et al., 2017.

Does a high suicide risk score mean an attempt is inevitable?

No. A score is conditional on the instrument, population, outcome, and time horizon. Many people labeled high risk will not experience the predicted event, and an elevated probability never describes a fixed individual destiny. Any concerning findings should prompt respectful clinical interpretation and appropriate support. Belsher et al., 2019.

Can a low risk score justify refusing care?

No. NICE recommends against using risk tools or low/medium/high strata to determine treatment access or discharge. The person's presentation, safety needs, and follow-up arrangements must guide decisions. A person can need support even without meeting a numerical threshold. NICE NG225.

Are AI models more accurate than traditional suicide risk scales?

Some multivariable machine-learning models report better discrimination in selected settings. However, discrimination alone cannot establish useful individualized probabilities, clinical benefit, or safety. A 2025 appraisal found that external validation and calibration remained uncommon across developed models, although a small subset showed promising validated performance. Their use still requires evaluation of the actual clinical workflow. Prediction models systematic review, 2025.

Why is a model with 80% accuracy potentially misleading?

If the outcome occurs infrequently, a model that predicts the common non-event for nearly everyone can report a large percentage correct while detecting few real events. Even strong sensitivity and specificity can produce low positive predictive value when prevalence is low. Research should report the confusion matrix, event prevalence, and clinically meaningful measures instead of an isolated accuracy figure. Belsher et al., 2019.

Does an earlier suicide attempt mean a future attempt will occur?

An earlier attempt is an important longitudinal risk factor and requires serious attention in assessment and follow-up. It does not make recurrence inevitable. The person's circumstances, current mental state, treatment, social resources, and safety environment remain relevant and change over time. CDC Risk and Protective Factors.

Does talking about suicide create risk by itself?

Supportive, direct conversation about suicidal thoughts is an accepted part of appropriate clinical screening and assessment. A respectful question helps clarify needs and connect a person to care. Prediction difficulties are not a reason to avoid asking about current thoughts, intent, and safety. NIMH ASQ Toolkit.

How should suicide prevention systems use prediction in the future?

Prediction research may contribute to carefully tested, equitable decision support and service outreach. It should operate within comprehensive clinical evaluation, transparent monitoring, accessible treatment, patient participation, and public health prevention. Models must demonstrate benefits in their intended populations before being used to make consequential care decisions. Prediction models systematic review, 2025; WHO LIVE LIFE.

Conclusion: Prevention Does Not Require Perfect Forecasting

Suicide risk prediction is difficult because important outcomes are relatively rare, suicidal states can change quickly, risk factors overlap extensively between people who do and do not experience future suicidal behavior, and models frequently lack sufficient independent validation or reliable calibration. Statistical discrimination, even when respectable, is only one component of clinical usefulness. The research record establishes clear limitations while also identifying promising opportunities for carefully validated prediction systems. Franklin et al., 2017; Prediction models systematic review, 2025.

An evidence-based response gives the person timely attention, psychosocial assessment, a revisable clinical formulation, collaborative safety measures, access to effective care, and continuity across transitions. Prevention also depends on safer environments and functioning community systems. A model can support research and perhaps improve certain workflows after appropriate validation. Responsibility for protecting human life remains with capable, accountable people and institutions acting on needs in the present. NICE NG225; WHO LIVE LIFE.

Related Articles

Suicide Screening Tools: ASQ, C-SSRS, and Their Limitations — How screening identifies possible safety needs and why test responses require clinical follow-up.

Suicide Risk After a Previous Attempt: Assessment and Follow-Up — How an earlier attempt informs longitudinal assessment, continuing care, and support.

Chronic Suicidal Ideation: Assessment, Treatment, and Support — Long-term suicidal thoughts, fluctuation, clinical formulation, and sustained treatment.

Suicide Risk in Autistic Adults: Evidence, Assessment, and Prevention — Neurodevelopmental context and adaptation of assessment rather than applying general-population assumptions.

Suicidal OCD: What Is It? Intrusive Self-Harm Thoughts, Compulsions, and the Difference From Suicidal Intent — The distinction between obsessional intrusive thoughts and actual suicidal intention.

References

Belsher, B. E., Smolenski, D. J., Pruitt, L. D., et al. (2019). Prediction models for suicide attempts and deaths: A systematic review and simulation. JAMA Psychiatry, 76(6), 642–651. https://doi.org/10.1001/jamapsychiatry.2019.0174.

Boudreaux, E. D., Larkin, C., Vallejo Sefair, A., et al. (2023). Effect of an emergency department process improvement package on suicide prevention: The ED-SAFE 2 cluster randomized clinical trial. JAMA Psychiatry, 80(7), 665–674. https://doi.org/10.1001/jamapsychiatry.2023.1304.

Carter, G., Milner, A., McGill, K., Pirkis, J., Kapur, N., and Spittal, M. J. (2017). Predicting suicidal behaviours using clinical instruments: Systematic review and meta-analysis of positive predictive values for risk scales. British Journal of Psychiatry, 210(6), 387–395. https://doi.org/10.1192/bjp.bp.116.182717.

Centers for Disease Control and Prevention. (2026). Preventing suicide. https://www.cdc.gov/suicide/prevention/.

Centers for Disease Control and Prevention. (2026). Risk and protective factors for suicide. https://www.cdc.gov/suicide/risk-factors/.

Franklin, J. C., Ribeiro, J. D., Fox, K. R., et al. (2017). Risk factors for suicidal thoughts and behaviors: A meta-analysis of 50 years of research. Psychological Bulletin, 143(2), 187–232. https://doi.org/10.1037/bul0000084.

Kleiman, E. M., and Nock, M. K. (2018). Real-time assessment of suicidal thoughts and behaviors. Current Opinion in Psychology, 22, 33–37. https://doi.org/10.1016/j.copsyc.2017.07.026.

Kleiman, E. M., Turner, B. J., Fedor, S., et al. (2017). Examination of real-time fluctuations in suicidal ideation and its risk factors: Results from two ecological momentary assessment studies. Journal of Abnormal Psychology, 126(6), 726–738. https://doi.org/10.1037/abn0000273.

Large, M., Kaneson, M., Myles, N., et al. (2016). Meta-analysis of longitudinal cohort studies of suicide risk assessment among psychiatric patients: Heterogeneity in results and lack of improvement over time. PLOS ONE, 11(6), e0156322. https://doi.org/10.1371/journal.pone.0156322.

Miller, I. W., Camargo, C. A., Arias, S. A., et al. (2017). Suicide prevention in an emergency department population: The ED-SAFE study. JAMA Psychiatry, 74(6), 563–570. https://doi.org/10.1001/jamapsychiatry.2017.0678.

National Institute for Health and Care Excellence. (2022). Self-harm: Assessment, management and preventing recurrence (NG225). https://www.nice.org.uk/guidance/ng225/chapter/recommendations.

National Institute of Mental Health. (n.d.). Ask Suicide-Screening Questions (ASQ) Toolkit. https://www.nimh.nih.gov/research/research-conducted-at-nimh/asq-toolkit-materials.

Seyedsalehi, A., Bailey, J., Ogonah, M. G. T., et al. (2025). Prediction models for self-harm and suicide: A systematic review and critical appraisal. BMC Medicine, 23, 549. https://doi.org/10.1186/s12916-025-04367-6.

U.S. Department of Veterans Affairs and U.S. Department of Defense. (2024). VA/DoD clinical practice guideline for assessment and management of patients at risk for suicide. https://healthquality.va.gov/HEALTHQUALITY/guidelines/mh/srb/index.asp.

World Health Organization. (2021). LIVE LIFE: An implementation guide for suicide prevention in countries. https://www.who.int/publications/i/item/9789240026629.