Utility of the pareidolia test in mild cognitive impairment with Lewy bodies and Alzheimer's disease

Previous research has identified that dementia with Lewy bodies (DLB) has abnormal pareidolic responses which are associated with severity of visual hallucinations (VH), and the pareidolia test accurately classifies DLB with VH. We aimed to assess whether these findings would also be evident at the earlier stage of mild cognitive impairment (MCI) with Lewy bodies (MCI‐LB) in comparison to MCI due to AD (MCI‐AD) and cognitively healthy comparators.


| BACKGROUND
Visual hallucinations (VH) are a feature of several psychiatric, neurological and ophthalmological disorders. 1 In dementia, VHs are particularly associated with the clinical syndrome of dementia with Lewy bodies (DLB) with an estimated prevalence of 55%-78%. 2 Complex VH, characteristically of well-formed images of people or animals, are one of the core clinical features differentiating clinicallysuspected Lewy body aetiology from the competing diagnosis of Alzheimer's disease (AD) in both dementia 3 and mild cognitive impairment (MCI). 4 VH may be context-dependent, and therefore difficult to quantify in clinical or research settings; these are often assessed through clinical interview after self-report by the patient, or report of this apparent phenomenon by an informant. This may limit the detection of VHs in the absence of insight by the patient, or unavailability of an informant. Comparable visual illusory phenomena (pareidolias, misperceptions of meaningful forms within ambiguous or visually-noisy stimuli) which may be elicited on demand in an experimental setting have therefore been proposed as proxies of VH. These pareidolic misidentifications have been shown to be more common in DLB than in AD or healthy controls, and to be positively correlated with the frequency of VH. 5 This research suggested that human and animal faces and bodies were the most common illusions in these tasks, supporting a phenomenological link to DLB-associated complex VH. A simplified pareidolia test, where participants report the presence or absence of illusory faces amongst visual noise scenes, has similarly shown to be reliable in discriminating DLB from AD (sensitivity of 81% and specificity of 92%). Pareidolia responses were more common in DLB than AD and healthy controls, were more common within cases of DLB with clinically-judged VH, and positively correlated with neuropsychiatric inventory (NPI) hallucination (of any sensory modality) scores. 6 While the pareidolia test shows apparent utility in discriminating hallucinations and DLB at the dementia stage, this utility has not yet been demonstrated in the prodromal stages of cognitive decline of MCI with Lewy bodies (MCI-LB) when cognitive impairments have begun to manifest. While neuropsychiatric symptoms, including VHs, may be present at this stage, they may be less common than in DLB while the full clinical syndrome is still emerging. 7,8 We therefore aimed to test the rate of pareidolic misidentifications in MCI-LB in contrast to MCI due to AD (MCI-AD) and age-matched healthy control subjects using the noise pareidolia test, and to consider the utility of this test in detecting clinically-judged complex VH and MCI-LB. Based on the above findings from the dementia literature, our hypotheses to test were: (1) MCI-LB patients would produce more pareidolic responses than MCI-AD or controls; (2) pareidolic responses would be more common in clinically-judged visual hallucinators than non-hallucinators, and correlate with severity of hallucinations; (3) the pareidolia test would acceptably classify clinically-judged visual hallucinators, and MCI-LB cases.

| Patients
Recruitment for this longitudinal cohort has been described in depth previously. 9 Briefly, participants over 60 years old were recruited from memory services, neurology and geriatric medical clinics in North East England. Prospective participants provided informed consent before undergoing more detailed screening by a research study medical doctor, and magnetic resonance (MRI) brain imaging. Those with possible frontotemporal or vascular aetiologies, parkinsonism preceding onset of cognitive symptoms by more than one year, dementia, or absence of objective cognitive impairment at screening were excluded. Inclusion criteria were age ≥60 years, and diagnosis of MCI at screening in accordance with NIA-AA criteria; concern about, and objective evidence of decline in cognition with maintained ability to function independently, 10 requiring a Clinical Dementia Rating (CDR) no higher than 0.5.

| Controls
Healthy participants were recruited from families of patients and local research involvement services and similarly screened as with patients undergoing medical review, neurological examination, MRI 1408brain imaging, and comprehensive neurocognitive assessment. Inclusion criteria were being age ≥60 years and cognitively healthy, with no known brain disease and a CDR of 0.
All participants, both patients and controls, were required to be medically stable on study entry. Local deprivation was calculated for each participant from the 2019 English Indices of Multiple Deprivation (IMD); IMD scores are divided at country-wide deciles so that a rank of one corresponds to living within one of the 10% most deprived neighbourhoods in England, and a rank of 10 being within the 10% least deprived neighbourhoods. 11  Images were visually rated as normal or abnormal by a five-person consensus panel of FP-CIT imaging experts, blind to clinical information. 123 I-mIBG cardiac sympathetic innervation imaging (cardiac mIBG) was also offered to all participants at baseline; delayed images (taken ∼ 4 h post-injection with medium energy collimators) were quantified with a heart:mediastinum ratio cut-off of <1.86 considered abnormal based on local data from healthy controls. 13  Diagnoses and classifications were repeated and updated after each follow-up assessment. In the case that participants were seen to have lost functional independence at follow-up assessment, all-cause criteria for dementia were considered. 14 No further follow-up was undertaken after diagnosis of dementia. Additional assessments were administered, but not considered for this work, having being detailed elsewhere. 8

| Pareidolia test
The 40-item visual noise pareidolia test 6 was administered at baseline and repeated at annual follow-up in the same manner: forty black-and-white visual noise images were presented sequentially on laminated cards. Individually differing human face images were presented within the noise in eight of these stimuli, and the remaining 32 contained only visual noise. After being shown three example stimuli to become acquainted with the task (two with faces, one without), participants were allowed up to 30 s to view each of the 40 test pages and asked to report if they did, or did not, see a face in each image. The test administrator recorded responses, out of view and without feedback or correction to the participant, as either correct (correctly identifying a face which was present, or correctly identifying a non-face stimulus), missed (missing a face which was present), or a pareidolia (where the participant identified a face as being present in a noise-only image). When providing ambiguous responses (e.g., 'maybe'), participants were prompted to provide either a 'yes' or HAMILTON ET AL.
'no' answer. As in previous studies, the count of pareidolia responses was the outcome of interest.

| Analysis
To assess group differences in the production of pareidolia responses, incorporating repeat assessments over time to maximise data availability and account for any time trends (e.g., increased pareidolia rates as MCI progressed), a generalised linear mixed model with log link function was estimated using the lme4 package for R software. Model fit was assessed by the Akaike Information Criterion

| Pareidolia analysis
Second-degree polynomials (linear and quadratic terms) were supported for the fixed effect of time only. No interactions with diagnosis were supported, and the resulting best-fitting models are presented in Table 2, with covariate effects in Table S1 In both models the marginal R 2 was relatively low compared to the conditional R 2 , suggesting that much of the variance in this measure could be attributed to individual-level differences in task performance. This is supported by the expected pareidolia values being low, even in MCI-LB groups, compared to the true observed range of pareidolia responses produced at baseline (see Table 1).
Repeated-measure correlations found no significant association between pareidolia responses and NPI-measured hallucinations score as rated by informants (r [71] = 0.03, p = 0.782), but did support a weak positive correlation between pareidolia production and total score on the NEVHI as rated by patients (r [120] = 0.22, p = 0.017).

| Classification analysis
Despite broad group differences in pareidolia response rates, the pareidolia test was found to have poor utility in classifying both hallucinating MCI cases specifically (AUC = 0.56), and MCI-LB (AUC = 0.61) in general (see Figure S1). Using cut-offs identified from the dementia stage, 6

| Summary of aims and findings
We aimed to assess if differences in performance on the noise pareidolia test observed between DLB and AD would also be present in the respective MCI stages of these.
We found only limited support for our hypotheses; probable MCI-LB were found to make more pareidolic mispercetions when completing this test than MCI-AD and controls, consistent with previous findings in DLB 6 but this association was not clearly found in the possible MCI-LB group.
There was no clear association between rates of pareidolic misperceptions and the presence of complex VHs as assessed by an expert clinical panel, contrary to the hypothesis. No association was found between pareidolia rates and hallucination severity (of any sensory modality) assessed by the NPI, but a weak association was found more specifically with simple and complex VH severity as measured by the NEVHI.
Finally, the utility of the noise pareidolia test in classifying either MCI-LB or clinical VHs was not supported; while the noise pareidolia test was able to differentiate MCI-LB from MCI-AD or healthy controls with good specificity, cut-off values from DLB had low sensitivity when applied to MCI-LB.

| Interpretation
These results partially extend previous findings to suggest that differences in the experience of pareidolias between DLB and AD 5,6 may already be apparent at the MCI stages of these aetiologies, with VH were much less common in this MCI-LB sample than is typical in DLB (22% vs. 55%-78%), 2 and pareidolias also occurred at higher rates in previous studies than in our own (mean of 3.5 in MCI-LB vs. 7.3 in DLB) 6 which may account for the limited utility of the noise pareidolia task in classifying these and MCI-LB. As our MCI patients were in the prodromal stage it remains likely that their clinical symptoms will continue to develop with more VH emerging closer to the onset of, and during, dementia. Pareidolic misidentifications may T A B L E 2 Generalised linear mixed models estimating pareidolia response production differences between diagnostic groups (Model 1) and hallucinators (Model 2). Intercept as expected count, fixed effects as incidence rate ratio  Table S1). While functional independence was highly variable at baseline in the MCI group with some particularly low IADL scores, these were assessed to include all contributions to functional dependence, including motor impairment (previously found to be correlated with baseline instrumental activities of daily living (IADL) scores in MCI-LB, while cognitive scores were not) 17 and social or cultural factors (e.g., the patient's contributions to housework were limited even prior to onset of any cognitive impairment). Despite some low IADL scores, all patients were judged to have MCI at baseline as evidenced by a CDR <1.

| Strengths and limitations
These data include a moderately-sized cohort with detailed clinical assessment and imaging for aetiological classification. We have made use of flexible modelling approaches to incorporate repeated measures to appropriately account for individual-level effects, and controlled for several anticipated confounding variables.
However, considerable variability was observed in this sample which was not explained by fixed effects. While we controlled for visual impairment reported at medical review, no objective measures of visual acuity were available, though previous research found no association between visual acuity and pareidolia rates in this test. 6 As a prospective cohort, it is not yet apparent which patients will develop VH by the time of onset of dementia, only those who have already done so (a minority of the MCI-LB group); while this clinical symptom was modelled as a timevarying predictor, it is not clear at this stage if an increased pareidolia rate in MCI may precede or predict the eventual clinical manifestation of VH.

| CONCLUSIONS
Probable MCI-LB had a higher rate of pareidolia responses in the visual noise pareidolia task than MCI-AD, who did not clearly differ from healthy controls. The relationship between hallucinations and pareidolia responses was not as clear as in dementia, with comparisons limited by low rates of hallucinations in MCI. Due to considerable inter-individual variation in task performance, the noise pareidolia test did not accurately classify MCI-LB or VH.