Research Summary

Artificial Intelligence Identifies Biomarker for Psychosis in Alzheimer Disease

Miranda Manier, BA

Key Highlights

  • A novel Alzheimer’s Disease Psychosis Network (ADPN) distinguished psychotic from non-psychotic Alzheimer disease with 77% accuracy.
  • The ADPN also separated psychotic Alzheimer’s disease from healthy elderly controls with 97% accuracy.
  • Increasing ADPN expression scores correlated with cognitive and functional decline over time.
  • ADPN expression was associated with increased metabolic connectivity between motor and language/social cognition regions.

In a study using data from the Alzheimer’s Disease Neuroimaging Initiative, researchers applied deep learning to brain imaging data to identify a metabolic network associated with psychosis in Alzheimer disease. Using a convolutional neural network trained on fluorodeoxyglucose positron emission tomography (FDG PET) scans, they developed a model that could distinguish patients with psychosis from those without, as well as from cognitively healthy elderly individuals. This model identified a distinct pattern of brain activity—termed the Alzheimer’s Disease Psychosis Network (ADPN)—that predicted the onset of psychosis with 77% accuracy, correlated with worsening cognitive and functional scores, and progressed over time, suggesting potential utility as a biomarker in clinical trials.

Psychosis in Alzheimer disease, defined in the study by the presence of delusions and hallucinations, is associated with poorer quality of life and higher rates of institutionalization. These outcomes are driven in part by disruptive behaviors, including motor and verbal agitation, which contribute to the syndrome’s more severe clinical course and underscore the need for targeted treatment strategies.

To evaluate the network’s specificity, the researchers analyzed FDG PET scans from three cohorts: 174 cognitively healthy elderly individuals, 174 individuals with Alzheimer disease without psychosis, and 88 individuals with Alzheimer’s disease who developed psychosis. The deep learning model identified brain regions where metabolic activity was consistently altered in those with psychosis. These regions included the prefrontal cortex, parietal and temporal cortices, visual processing areas, limbic structures such as the hippocampus and amygdala, and motor-related areas like the supplementary motor area.

The ADPN distinguished psychosis from normal aging with 97% accuracy and from non-psychotic Alzheimer disease with 77% accuracy. Expression scores within the network were higher in individuals with psychosis and continued to increase over 24 months, whereas scores in healthy controls remained stable. Higher expression scores were associated with greater impairment on the Clinical Dementia Rating Scale Sum of Boxes and lower scores on the Mini-Mental State Examination.

Compared with individuals without psychosis, those with AD + P also showed enhanced metabolic connectivity across the ADPN. These changes were particularly evident in links between the supplementary motor area and regions involved in auditory, language, and social cognitive processing.

“Alzheimer’s Psychosis Network holds promise as a biomarker for AD + P, aiding in treatment development and patient stratification,” the authors concluded.

Reference:
Nguyen N, Gomar JJ, Truong JN, et al. An artificial intelligence-derived metabolic network predicts psychosis in Alzheimer's disease. Brain Commun. 2025;7(3):fcaf159. doi:10.1093/braincomms/fcaf159