2026 ECNP Congress

PS01-1036

Predicting Parkinson’s psychosis: a machine learning and structural imaging approach

Linda Bryant — M. Vignando, M. Mehta, P.A. Lalousis · King's College London, Institute of Psychiatry- Psychology- & Neuroscience, London, United Kingdom

2026 ECNP Congress

Background

Psychosis is a common and debilitating neuropsychiatric complication of Parkinson’s disease (PD), typically manifesting as visual hallucinations and delusions and affecting more than half of patients over the disease course [1,2]. Parkinson’s disease psychosis (PDP) is associated with accelerated cognitive decline, increased risk of dementia, greater caregiver burden, and earlier institutionalisation [1,3]. Despite its clinical significance, there are currently no reliable tools to identify individuals at highest risk of developing psychosis. Structural magnetic resonance imaging (MRI) studies have demonstrated group-level differences between PD and PDP; however, these findings have limited translational utility at the individual level [5]. Machine learning (ML) approaches offer the potential to integrate multiregional neuroanatomical features to generate individualized risk predictions and identify biologically meaningful markers of PDP [4].

Methods

We analysed structural MRI data from 481 individuals with PD, including 135 patients with psychosis (PDP) and 346 without psychosis (PD). Atlas-based parcellations were derived from three complementary structural modalities: grey matter volume (GMV), cortical thickness, and thalamic nuclei volumes. Regional measures were adjusted for age, sex, total intracranial volume, and site to minimise confounding effects. For each modality, we trained separate support vector machine (SVM) classifiers using a repeated nested cross-validation framework. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), balanced accuracy, sensitivity, and specificity. Feature weights were examined to identify neuroanatomical regions contributing most strongly to classification.

Results

All models demonstrated above-chance classification performance. The GMV-based model achieved an AUC of 0.67, the thalamic nuclei model achieved an AUC of 0.66, and the cortical thickness model achieved an AUC of 0.62. Across modalities, distinct but converging neuroanatomical signatures of PDP emerged. In the GMV model, higher volumes in the brainstem and bilateral pallidum were the strongest predictors of PDP. In the thalamic nuclei model, larger bilateral habenula volumes were most predictive. In the cortical thickness model, increased thickness in bilateral extrastriate visual cortex (V2/V3) and somatomotor regions was associated with PDP.

Conclusions

This study demonstrates the feasibility of applying machine learning to structural MRI data for individual-level classification of PDP. Across modalities, predictive features converged on interconnected subcortical, brainstem, and cortical systems implicated in dopaminergic regulation, salience processing, and visual perception [3,5]. The prominence of brainstem, pallidal, and habenular structures suggests a key role for subcortical and midbrain circuitry in psychosis vulnerability in PD. Cortical findings further highlight the contribution of visual and sensorimotor networks, consistent with the phenomenology of visual hallucinations. To our knowledge, this is the first study to classify PD with versus without psychosis using multimodal structural MRI and machine learning. These findings provide proof of concept for imaging-based risk stratification and support the development of clinically deployable tools for early identification and monitoring of PDP in clinical settings.

References

  1. Dag Aarsland, et al. Psychosis in Parkinson’s disease: clinical features, risk factors, and management. Lancet Neurol. 2009;8(5):458–468.
  2. James G Goldman, Samantha Holden. Treatment of psychosis and dementia in Parkinson’s disease. Curr Treat Options Neurol. 2014;16:281.
  3. Dominic H Ffytche, et al. The psychosis spectrum in Parkinson disease. Nat Rev Neurol. 2017;13:81–95.
  4. Nikolaos Koutsouleris, et al. Prediction models in psychiatry: toward clinical implementation. Biol Psychiatry. 2016;79(3):187–195.
  5. Miriam Vignando, et al. Neuroimaging biomarkers of psychosis in Parkinson’s disease: a systematic review. Neurosci Biobehav Rev. 2022;132:123–140.