A systematic review of validated suicide outcome classification in observational studies

BACKGROUND: Suicidal outcomes, including ideation, attempt, and completed suicide, are an important drug safety issue, though few epidemiological studies address the accuracy of suicidal outcome ascertainment. Our primary objective was to evaluate validated methods for suicidal outcome classification in electronic health care database studies. METHODS: We performed a systematic review of PubMed and EMBASE to identify studies that validated methods for suicidal outcome classification published 1 January 1990 to 15 March 2016. Abstracts and full texts were screened by two reviewers using prespecified criteria. Sensitivity, specificity, and predictive value for suicidal outcomes were extracted by two reviewers. Methods followed PRISMA-P guidelines, PROSPERO Protocol: 2016: CRD42016042794. RESULTS: We identified 2202 citations, of which 34 validated the accuracy of measuring suicidal outcomes using International Classification of Diseases (ICD) codes or algorithms, chart review or vital records. ICD E-codes (E950-9) for suicide attempt had 2-19% sensitivity, and 83-100% positive predictive value (PPV). ICD algorithms that included events with 'uncertain' intent had 4-70% PPV. The three best-performing algorithms had 74-92% PPV, with improved sensitivity compared with E-codes. Read code algorithms had 14-68% sensitivity and 0-56% PPV. Studies estimated 19-80% sensitivity for chart review, and 41-97% sensitivity and 100% PPV for vital records. CONCLUSIONS: Pharmacoepidemiological studies measuring suicidal outcomes often use methodologies with poor sensitivity or predictive value or both, which may result in underestimation of associations between drugs and suicidal behaviour. Studies should validate outcomes or use a previously validated algorithm with high PPV and acceptable sensitivity in an appropriate population and data source.