TL;DR
This study leverages electronic health record (EHR) data from over 2.4 million emergency department (ED) visits to predict suicide risk in the six months following discharge. The researchers found that while incorporating a patient's recent history of mental health service use significantly improves prediction accuracy, simple binary indicators (e.g., "did they have a prior visit?") are just as effective as complex models that track the specific timing or "trajectory" of those visits.
What Problem Does This Paper Address?
The period immediately following an ED visit for mental health is a high-risk window for suicide. However, identifying which patients are at the highest risk remains a "vexing challenge." Standard clinical screening often falls short; research shows that 33% to 50% of patients who later attempt suicide actually denied having suicidal thoughts during their initial ED evaluation.
Clinicians need better tools to stratify risk. Previous research suggested that the pattern of healthcare use—such as an escalating frequency of visits or a sudden drop-off in care—might be a "digital phenotype" for suicide risk. This paper tests whether modeling these complex temporal patterns actually provides a meaningful boost in predictive power over simpler methods.
Methods
The authors conducted a retrospective cohort study using logistic regression to predict non-fatal suicidal events within 180 days of an ED discharge. They built a "base" model using demographics and current ED diagnoses, then layered on six different ways (permutations) of characterizing the patient's mental health service use over the preceding 180 days:
- Binary Indicator: Any mental health-related encounter (Yes/No).
- Aggregate Count: Total number of encounters.
- Monthly Counts: A set of six predictors for each month's visit volume.
- Weighted Slope: A slope calculated using an exponential decay function to give more weight to recent visits.
- Monthly Presence: Six binary indicators (one for each month).
- Interactions: A complex model looking at all 64 possible combinations of monthly service use patterns.
Model performance was measured using the Area Under the Receiver Operating Curve (AUC), where 1.0 is a perfect prediction and 0.5 is no better than a coin flip.
Datasets
The study utilized the Optum Labs Data Warehouse (OLDW), a massive database containing de-identified EHR data for over 200 million enrollees across the U.S.
- Scope: 2,445,597 ED mental health episodes involving 1,341,253 unique individuals.
- Timeframe: September 2015 to September 2022 (aligned with the U.S. adoption of ICD-10 coding).
- Population: Individuals aged 10 years and older.
Key Insights and Ideas
- Simplicity Scales: The most surprising finding was that "granular" modeling (tracking slopes or complex interactions of visits) did not significantly outperform a simple binary indicator of prior service use. For clinical implementation, this is a major win: it means effective risk tools can be built without needing computationally expensive temporal modeling.
- Acute Care is a Massive Red Flag: The strongest predictor of a future event was a prior inpatient or ED visit specifically for self-harm or suicidal ideation (Odds Ratio = 5.45).
- The "Aha" Moment on AUC: Moving from a model based only on demographics (AUC 0.63) to one that includes ED diagnosis and prior service history (AUC 0.83) represents a substantial leap in clinical utility. An AUC of 0.83 is considered "excellent" for behavioral health outcomes and outperforms many standard screening tools like the C-SSRS (which some studies peg at an AUC of ~0.59).
- Diagnosis Matters: Beyond prior history, certain diagnoses during the index ED visit were strong predictors, particularly personality disorders (OR 4.49) and psychotic disorders (OR 3.86).
Key Results
- Incidence: 7.2% (n=176,000) of the ED episodes resulted in a non-fatal suicidal event within 180 days.
- Predictive Power:
- Demographics only: 0.63 AUC
- Demographics + ED Diagnosis: 0.76 AUC
- Full Model (including prior service use): 0.83 AUC
- The "Trajectory" Comparison: The AUC for the simplest prior-use model (any event in 6 months) was 0.795, while the most complex interaction model only reached 0.806 (when added to demographics/diagnosis). This marginal gain suggests that the fact of recent acute care is more important than the specific timing of it.
Limitations and Future Work
- Data Bias: The Optum dataset, while large, may not perfectly represent the uninsured or those treated in systems outside the Optum network.
- Coding Limitations: The study relied on ICD-10 codes, which can sometimes conflate suicide attempts with non-suicidal self-injury (NSSI).
- Fatalities: The study focused on non-fatal events; predicting completed suicides remains a separate, albeit related, challenge.
- Future Direction: The authors suggest these findings should be used to develop "clinical decision support tools"—automated alerts within the EHR that help ED physicians identify high-risk patients who might otherwise be missed during a standard interview.
Summary updated