Development of an Ensemble Learning-based, Multi-dimensional Sensory Impairment Score to Predict Cognitive Impairment in an Elderly Cohort of Southern Italy

Background / State of the art Hearing, Vision, Olfaction, and Taste represent a complex sensory neural construct biologically linked to age (Gadkaree et al., 2016). Several studies have described their role as predictors in age-related neurodegenerative processes, particularly those related to cognitive decline (Brenowitz et al., 2020; Schubert et al., 2017). Hearing and vision are the most explored, providing robust population-based evidence (Gates et al., 2010; Panza et al., 2019; Sardone, Battista, et al., 2020) and using groundbreaking technologies, such as retinal imaging, in specific clinical settings for the latter. In its peripheral and central forms, hearing impairment has been considered one of the major modifiable risk factors for Alzheimer's dementia(Livingston et al., 2020). Among the sensory inputs related to the perception of physical events in the body, nociception and proprioception represent a set of functions that are often underestimated in their cognitive processing. Studies have shown how they are important not only in the global function of the elderly individual but also and especially in cognitive function. The interactions between aging processes and nociception are poorly investigated, particularly the relationship between chronic pain and global cognitive functions (Nadar et al., 2016). In particular, there are no precise correlates between different objective cognitive function analysis technologies, such as MRI and single or multiple interactions from sensory impairment.

Description and distribution of activities of each operating unit

Principal collaborators will have divided but integrated roles for the aims of the project. In particular, we can highlight two main branches of activities, with a list of the main activities for each research center:

1. Study Role: clinical and instrumental assessment

• ASL BA - Controls enrollment and cognitive-behavioural assessment research expertise
• Scientific Clinical Institutes Maugeri (ICS-Maugeri): Cases enrollment and functional and somato-sensorial research expertise

1. Study Role: design, modelling and feature elaboration

• IRCCS - Study Coordination, Design, and formal analysis , sensory impairment research expertise
• Polytechnic University of Bari (PoliBa)- data management and engineering, Bio-signal processing (OCT, electrophysiology, MRI), and ensemble learning architectures development Subject assessment will be divided between clinical setting (ICS Maugeri) and population setting (ASL BARI). IRCCS "S. De Bellis" will provide expertise for both design refinement, creation and monitoring of deliverables and milestones, and technological support for sense measurements. In fact, the close collaboration between Dr. Rodolfo Sardone and the PI and Co-PI, shown by numerous publications in the field of sensory impairment and functional decline, ensures a well-established and effective synergy in the proposed area of expertise.

ICS Maugeri, has also already collaborated with the group in other areas, but their role in this project will be truly innovative. They will be responsible for ensuring the recruitment of cases of demented individuals, whether with probable Alzheimer's dementia or vascular dementia. They will also contribute Dr. Pavese's expertise in introducing functional and lifestyle measures to be administered to the entire sample of subjects. Prof. Natoli, on the other hand, will contribute her expertise in the field of chronic pain assessment, both with subjective and functional scales and in the interpretation of biomedical signals from morphological and functional MRI. The technology group, essentially led by Bari Polytechnic, will be responsible for both the creation of the infrastructure for data collection, the processing of physiological signals (both electrophysiological and retinal imaging and MRI) and all variables collected in the study populations. Their expertise and synergistic collaborations with the group are showcased by a series of shared publications.

Specific aim 1 To assess the specific association between different sensory measures (central and peripheral hearing loss, retinal abnormalities measured by OCT, smell and taste objective measures, chronic pain, and proprioception subjective and electrophysiological measures.

Specific aim 2 To develop a multi-dimensional score using sensory features and clinical and lifestyle variables to predict the different types of dementia (Alzheimer's Disease, Fronto-Temporal Dementia, and Vascular Dementia) at different stages (Mild Cognitive impairment and Normal Cognition).

Specific aim 3 To create a connectomic map of the MRI morphologic and dynamic features of the Dementia cases and their relationships with sensorial features, describing the patterns differences in respect to the normal cognition controls.

Experimental design aim 1

Study Population:

This longitudinal study will be conducted using a multi-centric cross-sectional nested case-control design using data from the Casa della Salute project in Castellana Grotte and from the Clinical Neuro-rehabilitation center of the ICS Maugeri - Pavia.

The cognitive impairment cases will be selected from the ICS Maugeri - Pavia, Unit of Neurorehabilitation, and from the Neuro-psychological Outpatients Unit of the Casa della Salute - ASLBA. The cognitive normal control group will be derived from the secondary data of the SALUS in the Apulia Study baseline. SALUS is an ongoing study, started in 2012, on a representative population of older residents in Castellana Grotte (Puglia Region, Southern Italy). The study design and data collection method are described in detail elsewhere (Sardone et al. 2021). The sample included 2038 participants from the elderly (65+) residents in Castellana Grotte at the baseline (2012 to 2014). From this sample will be derived two sub- sample: one of the cognitively normal subjects and one with mild cognitive impairment according to the inclusion\exclusion criteria.

Cases: The cases included will be on two different stages of cognitive impairment: MCI and dementia.

The inclusion criteria for the MCI cases group will be: 1) to be at least 65 years old at the moment of enrollment: 2) to have an MCI (Montreal Cognitive Assessment, MOCA test, between 15.5-26) 3) to have a complete examination of the clinical, neuropsychological and sensory evaluations. The exclusion criteria for the case group will be 1) do not have the mental capacity to express consent at the follow-up; 2) have developed major malignancies or undergo major therapies could cause a loss in sensory function or cognitive decline; 3) be diagnosed with depression. The total number of MCI cases will be 45.

The inclusion criteria for the dementia cases group will be: 1) to be at least 65 years old at the moment of enrollment: 2) to have a mild to moderate cognitive impairment (Montreal Cognitive Assessment, MOCA test, below 15.5) to have a complete examination of the clinical, neuropsychological and sensory evaluations. The exclusion criteria for the case group will be 1) do not have the mental capacity to express consent at the follow-up; 2) have developed major malignancies or undergo major therapies could cause a loss in sensory function or cognitive impairment; 3) be diagnosed with depression. The total number of dementia cases will be 22 (among all the study sites).

Control Cohort: The inclusion criteria for the control cohort group will be: 1) to be at least 65 years old at the moment of enrollment: 2) to have a complete examination of the clinical, neuropsychological and sensory evaluations 3) to have no cognitive impairment (measured using MOCA greater 26). The total number of controls will be at least 151 (considering 1:5 allocation).

Methods: Ophthalmologic, audiological, physical, sensorial (olfactory and taste), pain and cognitive data derived from the examination at each assessment time will be recorded using a cloud-based data input form, associating every measure and quantitative variable to a root code for every subject examined. Further details are in section "Methods of data collection".

Experimental design aim 2 All data derived from the examination at each assessment time will be recorded using a cloud-based data input form, associating every measure and quantitative variable to a root code for every subject examined. The personal data (e.g., name, age, sex, address) will be detached from the clinical data using a subdivision matrix with a translation code (unknown to the evaluators) that links the master data to the subject code.

Outcome: The cognitive decline will be considered the primary outcome, notably the transition for every subject from MOCA 15.5-26 to a MOCA score lower than 15.5. In addition, a reduction in MOCA score of 10% points will be considered a secondary outcome of cognitive decline at every assessment.

Multisensory Score: the score will be calculated using different approaches to allow different combinations of sensitivity analyses to select the score with the most accurate prediction power.

Phenotype score will create an ordinal variable that cumulates every global impairment on each sense (e.g., hearing loss + vision loss yes\no + hypogeusia + hyposmia) with a variable from 0 to 4.

Ensemble Weighted Score: Every sense will be considered an independent weighted category depending on the number of sub-impairments that constitute the impaired sense. For example, the vision loss category will be created using the cut-offs for every OCT-a macro-variable impaired (under the 25th percentile) and the low visual acuity cut-off. Every variable will be implemented in a Random Forest machine learning model with the primary outcome as a dependent variable. The derived ranking (in terms of prediction power for cognitive decline) of variables will be used to weight each variable using the inverted rank as a multiplier for a weighted average using the total number of impairments as the denominator. Unsupervised Score: every continuous variable used to measure each sense (independently from the impaired side for ears and eyes) will be implemented in an autoencoder neural network algorithm, able to reduce the dimensionality of the variables. The encoders will create small coefficients (codes) explaining the total information derived from single variables. The codes will be used as predictors for each subject in further analysis.

Covariates: different covariates will be used as confounders of the association, particularly smoking, occupational and environmental exposure, education\socio economical status, physical activity, and BMI.

Modelling: A number of adjusted cox proportional models and other non-parametric time-event models will be run to assess the association between the different scores proposed and the cognitive decline. In addition, sensitivity analysis will be used with different methods to determine the model in terms of best fitting and prediction power. Due to the small number of observations in the models and considering the high number of covariates, particular attention to overfitting will be devoted, including regularised regression models.

Experimental design aim 3

Brain MRI Acquisition and Processing: A subsample of the participants will undergo a non-contrast-enhanced MRI scanning on a 3T scanner. Selected subjects already enrolled in the study will be selected according to the following criteria:

n.5 with a central auditory processing disorder diagnosis n.5 with a central auditory processing disorder and cognitive impairment (MOCA lower than 17.5) n. 5 only with cognitive impairment (MOCA lower than 17.5) n. 15 normal hearing and cognitive controls The aim of this sub-study will be the description of morphological differences between the four small groups using both supervised and unsupervised machine learning feature extraction and non-parametric statistical learning methods. The differences should create new knowledge on the structural brain interactions between cognition and hearing functions in the older subjects.

The controls will be enrolled all among the controls of the Bari center, and the cases between Bari and Pavia, according to a convenience sampling until the reaching of the enrollment objective.

The MRI protocol will include a T1- weighted sequence, a proton density-weighted sequence, a fluid-attenuated inversion recovery (FLAIR) sequence, and a T2¿ -weighted gradient echo sequence. For all sequences, the slice thickness will be 1.6 mm (zero-padded to 0.8 mm), except for the FLAIR sequence for which this will be 2.5 mm. An automated brain tissue classification method, based on a k-nearest-neighbour-classifier algorithm, will be used to quantify the following: total intracranial volume (ICV), total brain volume, grey matter (GM) volume, white matter (WM) volume, and cerebrospinal fluid volume (in mm3).

Voxel-based morphometry (VBM) will be performed according to an optimized VBM protocol. FMRIB software library (FSL) will be used for VBM data processing, and all GM and WM density maps will be non-linearly registered to the standard ICBM MNI152 GM and WM template (Montreal Neurological Institute) with a 1 mm × 1 mm × 1 mm voxel resolution. Subsequently, spatial modulation and smoothing procedure with a 3 mm (FWHM 8 mm) isotropic Gaussian kernel will be applied to all images (Ikram et al. 2015).

Methods of data collection The personal data (e.g., name, age, sex, address) will be detached from the clinical data using a subdivision matrix with a translation code (unknown to the evaluators) that links the master data to the subject code. The research group from POLIBA will design, develop and implement an overall interactive system, including interfaces for data extraction, artificial intelligence (AI) algorithms for data analysis, user interfaces for data visualization and interaction for the data aforementioned. POLIBA will adopt a Human-Centred Design approach [ISO9241-210], thus actively involving the other partners in the iterative activities of requirements analysis and design of system prototypes of incremental complexity. POLIBA will also study recommendation algorithms for data composition, privacy-compliant, from heterogeneous sources and AI algorithms for data analysis. POLIBA will identify visualization characteristics based on sensorial/perceptual mechanisms which allow a more effective visualization, presentation and interaction.

Laboratory assessment: Blood samples will be collected from all the participants and analyzed for inflammatory and metabolic biomarkers.

Physical \& cognitive assessment: The subjects will undergo a physical examination by a physician who will document all pertinent anamnestic information regarding sensory impairment and general health, particularly in the occupational or iatrogenic environment (e.g., exposure to ototoxic substances, optic and/or ionizing radiation), and lifestyle (e.g., nutritional status, smoking, education). Actigraphs will be adopted to collect data on physical activity and sleep quality. The physical performance will be evaluated with motion technologies. Further, data on socioeconomic status and mental and mood disorders will be collected. The MOCA test will be used to determine the subjects' cognitive level during the same session. The 30 item-Geriatric Depression Scale (GDS-30) will be used to identify individuals with depression.

Pain assessment: Physical examination by a pain specialist and a specialist in physical medicine and rehabilitation who will evaluate the cause, severity, and nature of pain and its effects on daily function and quality of life. Pain intensity will be evaluated according to the 11-point NRS (Numerical Rating Scale). The Brief Pain Inventory (BPI) questionnaire will be submitted to patients. Quality of life will be assessed by the SF-36 questionnaire. Pain thresholds will also be evaluated in all subjects by Quantitative Sensory Testing (QST).

Hearing loss assessment: otoscopic and tympanometric examination to check for middle ear conditions. Pure tone audiometry will be used to assess peripheral ARHL in a soundproof studio with specialised headphones. The Speech Discrimination Score (SDS) will be defined as the percentage of recognition of a list of 10 phonetically balanced Italian words at a 30 dB sensation level over the PTA threshold for each ear. The Synthetic Sentence Identification With Ipsilateral Competitive Message (SSI-ICM) test, which measures central auditory dichotic processing, will be utilised to identify age- related CAPD.

Eye assessment: Every participant will undergo a basic ophthalmological exam: 1) Manual refraction and assessment of each eye's best-corrected visual acuity (BCVA); 2) intraocular pressure and slit-lamp biomicroscopy; 3) The retinal Optical Coherence Tomography with Angiography (OCT-A) will be scanned after pupil dilatation.

Taste and olfaction assessment: The olfactory function will be assessed using the battery "Sniffin' Sticks". Individual pens will be provided to the participants, who will then be asked to categorise and describe the smells using four distinct descriptions for each cell. The four basic tastes (sweet, salty, bitter, and sour) will be tested for taste function with taste strips at concentrations above the threshold (Burghart GmbH, Wedel, Germany).

Statistic plan The sample size was calculated in a case-control logic: considering the prevalence of the least common sensory impairment (age-related hearing loss) of 18% and an odds ratio (OR) of 2.1 for the probability to observe cognitive impairment in subjects with age-related hearing loss with a confidence level of 95% and a desired power of the 80%, based on our previous study (Sardone et al. 2020), the sample size will be of 22 cognitive impairment cases and 45 mild cognitive impairment for the cases and at least 151 subjects for the control cohort.

Patients will be enrolled according to a non-probabilistic convenience sampling with a target accrual of 218 patients, included and subdivided according to the inclusion/exclusion criteria described above. Given the case-control research design of the study, the entire sample will be observed cross-sectionally for all the features included in the model. To comply with the proportion of the entire sample that is representative (15%) of the older population of an Italian region, we will increase controls to at least 2 times the number of cases (n = 151), matched to cases using the nearest neighbor technique for age, sex, and education.

Feature Engineering:

All the data will be recorded using cloud-based input forms which can be used by all users participating in the research project in the various participating centers. The data will be interfaced via connectors directly to sense-measurement equipment and will integrate images from OCT and MRI with DICOM technology. Following collection, measurements of the completeness of the collected variables will be made. In case of missing a variable up to 30% of the entire dataset, Multiple Imputation Chain Equation will be operated above 30% the variables will be excluded from the analysis. Each variable that will contribute to the creation of the score will be scaled in order to normalize its distribution and make its magnitudes comparable. This step is critical in handling large amounts of variables in the same model, especially if they are of different sources. The brain images will be post-processed and analyzed in the analysis phase, and they will be included in the database only after an accurate quality check.

Statistical analysis

The sample will be divided into cases (two different stages of cognitive impairment: MCI and dementia) and controls (neither condition). Normal distributions of quantitative variables will be tested using the Kolmogorov-Smirnov test. Therefore, clinical and functional differences will be described for the two groups in terms of frequency and associations between groups for all the variables and the features recorded during the study. Normal distributions of quantitative variables will be tested using the Kolmogorov¿Smirnov test. Data will be reported as mean±standard deviations (M±SD) for continuous measures and frequency and percentages (%) for all categorical variables. In order to focus on the practical differences between the groups, in terms of effect size (ES) instead of p values, a statistical approach based on the null hypothesis significance test (NHST) will be not used. Therefore ES differences between continuous variables will be calculated using Cohen's d difference between means, Hedge's g when the assumption of a similar variance will be violated, and their ES using the confidence intervals. Their 95% confidence intervals (CI) will be calculated to assess essential differences in the magnitude of association. An ensemble machine learning algorithm for feature selection will be adopted to rank the predictive power of each covariate derived from the biological sample analysis using a Shapley values ranking. Statistical and Machine Learning Methods:

To select variables a machine learning model will be used: the Random Forest (RF). The RF is an ensemble model using bagging as the ensemble method and a decision tree as the individual model. This method is able both to select the variables and to rank them in terms of prediction power and most able to classify the selected outcome. Every sense will be considered an independent weighted category depending on the number of sub-impairments that constitute the impaired sense. For example, the vision loss category will be created using the cut-offs for every OCT-a macro-variable impaired (under the 75th percentile). Every variable will be implemented in a Random Forest with the cognitive decline as a dependent variable. The derived ranking (in terms of prediction power for cognitive decline) of variables will be used to weight each variable using the inverted rank as a multiplier for a weighted average using the total number of impairments as the denominator. Ensemble model will be selected because combining individual models, tends to be more flexible (less bias) and less data-sensitive (less variance).

Furthermore, to compare the association with the outcome, different multiple linear and/or logistic models (or other non- parametric models, depending on the variables) will be run to assess the association between the different levels of the scores proposed and the cognitive impairment. In addition, sensitivity analysis will be used with other methods to determine the model in terms of best fitting and prediction power. Different covariates will be used as confounders of the association, particularly smoking, occupational and environmental exposure, education\socio economic status, physical activity, and BMI. Considering the high number of covariates, particular attention to overfitting will be devoted, including a regularized regression approach using different penalty regression meta-learners to find the best penalty term for each model (Mahani and Sharabiani, n.d.).