P1.45. HALLUCINATIONS AS A CONSEQUENCE OF PREDICTIVE CODING AND BAYESIAN INFERENCE DISRUPTIONS: A SYSTEMATIC REVIEW ACROSS PSYCHOSIS, PSYCHEDELICS, AND SENSORY DEPRIVATION WITH MACHINE LEARNING PERSPECTIVES
Aakanksha Aakanksha1, Indranath Chatterjee2
1 Tbilisi State Medical University, Department of Neurology, 33 Vazha Pshavela Avenue, Tbilisi 0159, Georgia
2 Manchester Metropolitan University, Department of Computing and Mathematics, Manchester M1 5GD, United Kingdom.
INTRODUCTION: Hallucinations are perceptions without any external stimuli, which may arise from disrupted predictive coding, which is a neurocomputational framework deeply rooted in Bayesian inference. In this, the brain integrates prior beliefs with sensory input to minimize errors in prediction. When this mechanism breaks down due to very strong priors in psychosis, hypopriors in psychedelic states, and lack of input in sensory deprivation, hallucinations occur. Despite differences, these states share inferential dysfunction. ML can help us classify and predict the hallucinatory state by detecting patterns in EEG, fMRI, and other data, such as behavioural data.
AIM(S): This systematic review aims to identify and analyze studies exploring predictive coding and Bayesian inference disruptions in hallucinations across different states, compare how these mechanisms vary, and then evaluate the possible role of ML in classifying or predicting hallucinatory phenomena.
METHOD(S): A search was conducted on PubMed and Scopus (2020-2025) using PRISMA guidelines. Inclusion criteria covered empirical, computational, or ML studies focused on hallucinations that involved Bayesian and predictive mechanisms. A total of 46 studies met the inclusion criteria, which are included in this systematic review.
RESULTS: Studies show predictive coding disruptions as the main reason for hallucinations. In schizophrenia (impaired precision and a reduction in mismatch negativity), Psychedelics (hypopriors with unstable percepts), and Sensory deprivation trigger spontaneous top-down predictions. ML models using EEG and fMRI features like N100 suppression and effective connectivity may achieve high classification accuracy of hallucinatory states. However, inter-condition robustness remains limited.
CONCLUSIONS: This review sports a unified Bayesian framework underlying hallucinatory states. The integration of Bayesian inference with ML offers a promising path towards diagnostic biomarkers and precision interventions across altered states.
FINANCIAL SUPPORT: No financial support was received for this study.