Machine learning-based measurement of delusional dreaming
This study investigated the prevalence of delusion sentiment in dreams using machine learning-based measurement. Classification models were developed by training the SVM (support-vector machine) algorithm with 841 words relevant to grandiose delusions and 978 words relating to persecutory delusions. They were then utilized to score grandiose and persecutory sentiment in 2611 dreams primarily obtained from an open source, including dreams reported by American, Chinese, German, and Peruvian people. The classification accuracy of the SVM model for detecting grandiose words was 86.4%, that for detecting persecutory words being 97.6%. The prevalence rates of dream reports being classified by the SVM algorithms as grandiose and persecutory dreams in the entire dream collection were 12.2% and 11.2%, respectively. Overall, around a quarter of dreams exhibited delusional content, which is more prevalent than the epidemiological estimate of psychosis in waking life – that is, approximately 0.3% worldwide. Given its fine-grained scoring, cost-efficiency, and absence of subjective judgment, the SVM method can be a useful tool for coding delusional sentiment in dreams.