Determining Distinct Suicide Attempts From Recurrent Electronic Health Record Codes: Classification Study.

Area of Science:

• Medical Informatics
• Public Health
• Psychiatry

Background:

• Prior suicide attempts are a significant risk factor for future suicidal behavior.
• Electronic health record (EHR) data is increasingly used for suicide attempt risk prediction.
• Data leakage from diagnostic codes for prior attempts can inflate model performance.

Purpose of the Study:

• To develop an automated rule for identifying distinct suicide attempt events using diagnostic codes.
• To address data leakage in EHR-based suicide risk prediction models.

Main Methods:

• Reviewed 1015 pairs of suicide attempt diagnostic codes from EHRs within 90 days.
• Chart reviewers determined clinical setting, method, and inter-code interval for each pair.
• Calculated the probability that the second code represented a distinct suicide attempt event.

Main Results:

• 82.3% of code pairs were nonindependent, referring to the same event.
• Distinct suicide attempts were identified only 3.3% of the time when the second code was non-ED.
• A positive predictive value of 0.90 was achieved for distinct attempts when the second code was in an ED at least 5 days after the first.

Conclusions:

• EHR-based suicide risk models are susceptible to bias from non-distinct suicide attempt codes.
• A rule was derived: ED codes ≥5 days after a prior attempt reliably indicate new events.
• This rule can minimize bias in suicide risk prediction models using prior attempt data.