Predictors of suicide risk in adolescent and young adults (AYA) with cancer.

Abstract

11519 Background: Whilst AYA accounts for a small proportion of annual cancer cases, the complexity of this subgroup places them at a disproportionately high risk of psychological distress and poor mental health. This study aimed to identify patient and disease characteristics associated with suicide in AYA patients. Methods: A retrospective analysis of AYA (aged 15-39) between the years 1973 to 2015 from the Surveillance, Epidemiology, and End Result Research database were included. A training and test set were used to develop a model predicting suicide. The SMOTE sampling algorithm was used for the training set due to severe class imbalances. A random forest model was trained with 10 clinical features (cancer subtype, age, sex, race, marital status, diagnostic year, number of in situ/malignant tumours, chemotherapy, surgery, and radiation therapy) with recursive feature elimination (RFE) via 10-fold cross validation. Results: There were 139,394 AYA included with 974 (0.7%) having a documented suicide or self/inflicted injury yielding an incidence of 1499 suicides / 100,000 person years. The standardized mortality ratio (SMR) for AYA was 34.4 (95% CI: 32-37)(Aged 15-24: 47.4 (39-57); Aged 25-39: 32.3 (29-35)). Suicide rates increased over time (1973-1980: SMR = 18.0 (15-22); 2001-2015: SMR = 127.0 (102-155)). Univariate analyses observed high rates of suicide in: females (43.8 (38-51)); single/unknown relationship status (47.1 (42-52)); “other” race (85.2 (52-132)) and by cancer type: leukaemia (77.1 (42-129)), soft tissue sarcoma (73.7 (57-94)), unspecified malignancy (71.0 (9-256)) and brain (63.6 (43-91)). We randomly assigned AYAs to a training set (n = 97,576) and test set (n = 41,818), stratified by cause of death. All 10 clinical features were retained with RFE. The prediction model achieved an AUC of 0.59, accuracy of 0.78, sensitivity of 0.40 and specificity of 0.78. Conclusions: Risk of suicide in AYA is high compared to normative data, further heightened in females, single/unknown relationship and certain tumour subtypes. The prediction model performed poorly, potentially related to the exclusion of mental health variables, which are not collected by the SEER dataset. Our findings suggest that dedicated psychosocial supports and targeted mental health assessments are critical components of care for AYA.