Abstract
BACKGROUND: Auditory verbal hallucinations (AVHs) are a core symptom of psychosis but their prevalence in the general population ranges from 5-28 %.
METHODS: This study aimed to identify the most strongly associated predictors of AVH from 42 sociodemographic, environmental, biological, and psychological and health measure predictors in two samples from the UK Biobank; a clinical sample of individuals diagnosed with psychosis (n = 304 individuals who have experienced AVH, n = 634 non-AVH) and a non-clinical sample from the general population (n = 1772 AVH, n = 130,464 non-AVH). A third, combined sample was used to identify key predictors for distinguishing those with and without psychosis among AVH-only samples. The study employed the machine learning XGBoost classification algorithm to identify important predictors of AVH in these samples.
RESULTS: All three models performed reasonably well with balanced accuracies of 66-78 %. There were consistent results for both the non-clinical and clinical samples where passive suicidal ideation, measures of distress, traumatic experiences, anxiety and educational attainment were important predictors of AVH in both samples. In addition, childhood trauma, depression and smoking cannabis were predictors of AVH in the non-clinical population sample, while self-harm, sleep deprivation and polygenic risk score (PRS) for bipolar disorder were predictors of AVH in the clinical sample. Among the AVH-only sample, mental distress, seeking professional help, self-harm, PRS for schizophrenia, and employment status were predictors of psychosis.
CONCLUSIONS: While many predictors identified may be a result of experiencing AVHs, ML analysis indicates that traumatic experiences may be useful temporal predictors of AVHs, confirming previous hypotheses in the literature.