Is it possible to predict inpatient suicide in psychiatric hospitals?

Background

Suicide is a leading cause of death among patients with severe psychiatric disorders. Although the majority of suicides occur outside of hospitals, suicide by psychiatric inpatients, while rare, occurs at rates substantially higher than those observed in the general population, with some reports suggesting rates up to 50 times higher in psychiatric inpatient settings than in the general population. Because inpatient suicide is often perceived as preventable, it carries significant clinical and medico-legal implications, and may lead to scrutiny of clinical decision-making and standards of care. Despite this, the identification of patients at imminent risk during hospitalization remains a major challenge. This study compared three approaches applied electronic medical records (EMR): qualitative analysis, machine learning (ML), and review by senior psychiatrists.

Methods

We conducted a multicenter matched case-control study using EMRs from three psychiatric hospitals in Israel. The sample included 29 inpatients who died by suicide during hospitalization and 87 matched controls. For each case, three controls were matched by hospital, age, sex, and diagnosis. Data were extracted from the final month of hospitalization, or the entire admission when shorter. Qualitative analysis used a structured coding framework derived from clinical documentation. ML models were applied to de-identified physician-authored notes. Two senior psychiatrists, blinded to outcome, reviewed the EMRs and provided binary suicide-risk classifications and 5-point risk ratings. Since the study used a matched case-control design with an enriched proportion of suicide cases, prevalence-adjusted PPVs were calculated using the observed sensitivity and specificity adjusted by the prevalence of 0.04% per admission, based on the inpatient suicide rate in the source population.

Results

The qualitative analysis did not identify meaningful differences between suicide cases and controls; only a trend-level finding suggested lower inpatient cooperation among cases. ML performance was modest across models, with low prevalence-adjusted positive predictive values. In the analysis by senior psychiatrists, the positive predictive value of psychiatrist classifications remained low, at 0.12%, indicating limited clinical utility for identifying individual patients who would die by suicide during hospitalization. Inter-rater agreement between psychiatrists was low. Cohen’s kappa for the binary classification was 0.13, and agreement across the 5-point ordinal ratings was similarly low (kappa=0.093), indicating limited consistency in clinical judgment.

Conclusions

Using three completely different analytic approaches, no method yielded clinically useful discrimination between future psychiatric inpatient suicide cases and matched controls. The qualitative analysis did not identify meaningful between-group differences, machine learning models showed only modest performance with low prevalence-adjusted positive predictive values, and psychiatrist-based assessments showed low positive predictive value and low agreement. Suicide during psychiatric hospitalization is impossible to predict at the individual level. These findings likely reflect both the limitations of record-based data and the inherent difficulty of predicting such a rare and dynamic outcome. Accordingly, risk assessment in inpatient psychiatry should not be understood as a problem that can be solved by reviewing records alone, whether by humans or by algorithms.

References

No references