In this report, the application of a novel deep learning methodology to the brain FDG-PET scans of those who developed psychosis in AD has identified a predictive and quantifiable topographical metabolic covariance network, the ADPN. In those with AD who developed psychosis, baseline expression of key network regions was higher than in those without psychosis and longitudinal ADPN expression in psychosis continued to increase over time. This is similar to the expression and progression of previously identified disease-specific metabolic networks including non-psychotic Alzheimer’s disease and Parkinson’s disease, whose network dysfunctions tracks with cognitive and motor symptomatology respectively [2, 5, 43]. In previous studies, those who experience psychosis in the context of AD have been shown consistently to decline more rapidly and evidence greater functional impairment, which are hallmarks of a more aggressive neurodegenerative condition [12, 19, 22, 42, 44, 45]. Consonant with this, in the current report, individuals with higher ADPN expression scores evidenced increased impairment on CDR and lower MMSE scores, reflecting the functional and cognitive correlates of the psychotic syndrome.
While the previously reported metabolic network that predicts the cognitive syndrome of AD comprises the precuneus and temporoparietal regions [5], the ADPN identified in the current report that predicts psychosis in AD overlaps in temporoparietal regions but extends to a broader network that includes prefrontal regions and occipital regions encompassing critical structures for visual processing- the lingual and fusiform gyri. As psychosis in AD is typified by a more severe cognitive phenotype together with delusions and/or hallucinations that are primarily visual [7], the integration of prefrontal cortex with hippocampal/parahippocampal regions and visual cortex in the ADPN represents a logical extension of previous imaging and neurocognitive findings. Imaging studies consistently highlight the critical role of frontal systems in delusions in AD [22, 46, 47]. Psychosis in AD has also been associated with advanced hippocampal and parahippocampal atrophy [48, 49], a finding that is consistent with the known accelerated cognitive decline that accompanies the behavioral symptoms [6]. Hallucinations in AD have not been as extensively studied in the absence of delusions as perceptual disturbances are less common and predominantly appear together with elaborated delusional ideation, but published studies have implicated the lingual gyrus in particular and the occipital cortex in general [50, 51]. Within the visual cortex, the lingual gyrus is critical for the processing of visual images and provides inputs to the neighboring fusiform gyrus [52], a structure that plays a key role in facial affective processing [53]. Impaired facial affective processing has been reported in psychotic AD [54], and the inability to decode the valence of facial emotion from network dysfunction in the visual cortex may contribute to the elaboration of delusions when malevolent intentions are misattributed.
Key regions within the ADPN network space also include nodes not previously implicated in AD psychosis, particularly the insula and the supplemental motor area (SMA). Several lesion studies have highlighted the participation of the insula in psychosis-relevant facial affective processing [55]. The SMA is a region that is critical to both the planning and execution of voluntary motor behaviors and to the motor aspect of speech production [56]. As psychosis in AD is known to be associated with a poorer quality of life and increased rates of institutionalization due in part to disruptive behaviors that include motor and verbal agitation [57], the involvement of the SMA in the ADPN could provide a neurobiological underpinning for this well-known association. Additionally, the SMA plays a critical role in working memory and is a component of the motor-language SMA syndrome [58]. Working memory deficits have been found to be a defining feature of the cognitive syndrome of psychotic AD that have been linked to frontal hypometabolic patterns [22, 42, 59].
Within the ADPN space, we found that connectivity between individual nodes was uniformly increased- rather than decreased- in those with AD psychosis relative to those with non-psychotic AD, predominantly linking frontal with temporal and parietal regions. The SMA was found to be a particularly active node, exhibiting connectivity with structures critical for language and social cognition including Heschl’s gyrus, the superior temporal gyrus, the supramarginal gyrus and the inferior parietal lobule [60–62]. Alterations in functional connectivity have been previously reported utilizing resting-state functional MRI in primary (non-neurodegenerative) psychotic illnesses. In those with schizophrenia and bipolar disorder there is strong evidence of disruption of functional network connectivity that include frontoparietal (dorsolateral/dorsomedial prefrontal and lateral parietal), together with the default mode and ventral attention networks [63, 64]. The degradation of the frontoparietal networks in psychotic illness is thought to be reflected in the transdiagnostic dysexecutive cognitive profile observed in psychotic and severe mental illness [63]. There is also recent evidence of well-defined patterns of disruption of striatal connectivity that comprise increases and decreases in intra- and extra-striatal functional connections reflecting abnormalities of dopaminergic neurotransmission that predict treatment response and may collectively have a value as a diagnostic imaging biomarker in primary psychotic conditions [65, 66]. In non-psychotic AD, functional connectivity has generally been reported to be eroded[67] especially in posterior networks that include the hippocampus [68], while connectivity has also been reported to be augmented, especially in networks that include frontal structures [69–71]. It has been suggested that the trend towards increased frontal connectivity in AD may be compensatory for impaired temporal and posterior network coherence [67]. In AD psychosis, in the only published study of a small cohort of AD subjects with delusions, reduced connectivity of a single cluster in the inferior parietal lobe that includes the angular gyrus with the other nodes of the default mode network was reported [72]. In light of the findings in the current report that suggest that psychotic AD may be distinguished by increased metabolic connectivity linking nodes with relevance for motor behavior with regions critical for language and social cognition, it may be that delusions and the aggressive behavior that follow them in AD emerge from compensatory engagement of degraded network structures that are critical for the narrative integration of experience with those regulating behavioral planning and execution.
Although neuropathologic biomarkers that could be relevant to ADPN expression and the cognitive and functional decline associated with it that include quantifications of amyloid and tau pathology were not studied in the current report, Braak stating of tau pathology has previously been found to correlate with the expression of metabolic network patterns in non-psychotic AD [73]. Additionally, several studies have reported that increases in phosphorylated tau pathology in frontal cortex is associated with psychosis in AD [16, 17]. In a study of the association of the topography of fibrillar tau pathology with the development of psychosis in AD, increases in the retention of the tau PET tracer [18F]-AV1451 in frontal regions; precuneus; parahippocampal and supramarginal gyri; and lingual and fusiform gyri that overlap with key regions within the ADPN were observed in those destined to become psychotic over the course of the study [19]. It follows that tau pathology may be responsible for metabolic abnormalities that adumbrate the ADPN. Further research is required to determine whether individual ADPN expression predicts AD neuropathological burden in AD psychosis, and whether the regional distribution of amyloid and/or tau pathology implicates key ADPN nodes.
Our findings are subject to several limitations. One such limitation is our inability to control for medication status, including antipsychotic medications, cholinesterase inhibitors, and NMDA antagonists. Consequently, the derivation of the ADPN may have occurred under varying medication states, potentially influencing the distribution of metabolic activity. Notably, the use of antipsychotic medications could make cognitive decline in psychotic patients more severe [74], which may in turn amplify the observed longitudinal increase in ADPN expression in AD + P subjects. Furthermore, the sensitivity of the NPI-Q to drug-induced behavioral changes, as highlighted by Lai [75], suggests that it may not be the most reliable method for establishing psychosis [74], thereby raising the possibility of misdiagnosis in some subjects. Lastly, the absence of an independent validation dataset limits the generalizability of our findings. Future studies demonstrating validation in an independent dataset would enhance confidence in the ADPN and support its utilization in a clinical context.
4.1 Conclusion
The application of CNN within an explainable AI framework to FDG PET scan data together with graph theoretical analysis has identified a metabolic network pattern, the ADPN, predictive of psychosis in AD and the related accelerated cognitive and functional decline. As there are currently no FDA-approved treatments for psychotic AD and currently available antipsychotic medications are associated with an increased risk of mortality [6], the development of novel therapies and the identification of those most likely to respond could be facilitated by validated imaging biomarkers that may include the ADPN.