PRedicting the EVolution of SubjectIvE Cognitive Decline to Alzheimer's Disease With Machine Learning

The Regional Reference Center for Alzheimer's Disease and Cognitive Disorders of Careggi Hospital, Florence, started about 25 years ago to collect clinical, neuropsychological, personality and lifestyles data for patients with of Subjective Cognitive Decline (SCD). Starting from these data the aim of the present study is to increase the number of variables collected on SCD subjects and expand the clinical follow-up to delineate more accurate pathway of disease and to identify very early subject at risk of conversion in Alzheimer's Disease (AD).

The advantages of detection of cognitive impairment at the early stages are critical. Such detection has traditionally been performed manually by one or more clinicians based on reports and test results. Machine learning algorithms offer a method of detection that may provide an automated process and valuable insights into diagnosis and classification. With this tool clinicians could design optimal screenings to predict, already at the stage of SCD, possible evolutions of the patient's conditions, and, through decisional protocols, establish the degree of appropriateness of the investigations/exam to perform for each patient, based on current state of risk.

The aim of this study is to evaluate, through machine learning tools, the accuracy of clinical data, neuropsychological assessment, personality traits, cognitive reserve, genetic factors, cerebrospinal fluid (CSF) neurodegeneration biomarkers and features from EEG and event-related potentials (ERP) in predicting conversion from SCD condition, to Mild Cognitive Impairment (MCI) and AD.

All participants will undergo a comprehensive familial and clinical history, general and neurological examination, extensive neuropsychological battery (about 19 tests exploring all cognitive domains: memory, attention, executive functions, language, praxis), estimation of premorbid intelligence (TIB - Test di Intelligenza Breve, an Italian version of the National Adult Reading Test - NART), personality traits (Big Five Factors Questionnaire - BFFQ), and leisure activities evaluation (structured interviewed regarding participation in intellectual, sporting and social activities) as well as assessment of depression (Hamilton Depression Rating Scale - HDRS). Moreover, all patients will undergo a CSF analysis to assess established biomarkers (Aβ42, total tau, p-tau) and blood samples for DNA analysis will be collected to identify the Apolipoprotein E (APOE) and brain-derived neurotrophic factor (BDNF) genotypes. In particular, will be analyzed three different single nucleotide polymorphisms (SNPs):

• BDNF SNP Val66Met (valine-methionine substitution) was already associated with poorer episodic memory and abnormal hippocampal activation assayed with MRI
• APOE gene rs429358 and rs7412 that are involved in the Amyloid plaque formation Only clinical and neuropsychological assessment are repeated yearly in order to evaluate progression of decline.

The sub-sample will be selected including all new cases of SCD included in the study (about 50), all subjects converted to MCI (about 25), and all patients, already studied, who report a subjective cognitive disorder for less than 10 years (about 75).

This selected sub-sample will also perform additional investigations: CSF biomarkers, resting state EEG recordings and ERP registration.

For about 150 subject the EEG activity will be recorded. The participants will be administered an ERP test battery with concurrently recorded EEG. Analysis of EEG data, in particular, will start with a standardized early-stage EEG processing pipeline (PREP Pipeline) focusing on high-filtering at 1 Hz, the identification of bad channels and the calculation of a robust average reference. The investigators will perform signal epoching using fixed length epochs for resting state measurements, and stimulus-locked epochs for signals acquired during tasks. Epochs containing artifactual signal will be removed with a semi-automatic procedure. The investigators will then apply independent component analysis (ICA) to the data, followed by a semiautomatic detection of artifactual components based on measures of autocorrelation, correlation with EOG electrode, focal channel topography and quality of mono-source dipole fitting. Finally, the investigators will compute power features at the sensor space, i.e. frequency bands powers, and linear/non-linear functional connectivity metrics, i.e. direct transfer function and mutual information, at the source level after a source reconstruction procedure, i.e. Low Resolution Brain Electromagnetic Tomography (LORETA) method (Pascual-Marqui et, al 2002).

Machine learning analysis will proceed as follows. The first analysis will include all the patients that were diagnosed with SCD in the preliminary work. The investigators will define a set of multi-modal features including all the 19 neuropsychological tests, a selection of the personality assessments and the genetic profile. The investigators will define a common metrics based on the dispersion of each single dimension of the profile and then they will train a machine learning algorithm to associate each profile "vector" to the associated evolution of the disease after a fixed time (at first the possible categories will be only SCD, MCI, AD with no further sublevels, but this can be improved in further rounds). First, the investigators will perform this classification with standard machine learning procedures as support vector machine (SVM, using binary decisions trees) and k-nearest neighbors (kNN). Then they will exploit the fact that neurophysiological assessments are repeated yearly (see above) to use the SuStaIn algorithm previously applied to neuroimaging data (see state of the art and Young 2018). Briefly, the algorithm z-scores each dimension of the profile and models with a piece-wise linear function the different paths of accumulations of each marker, i.e. the progression from healthy to abnormal values. This approach does not only outperform standard algorithms in predicting the final condition of the subjects based on the starting profile but highlight the different paths that can lead to the same outcome. In a separate set of analysis, the investigators will follow a deep learning approach, using exactly the same subject profiles to train a multi-level feedforward artificial neural network (ANN) predicting the condition of the patient after a fixed time. The networks will be trained with standard gradient descent and backpropagation techniques and by means of dropout and batch normalization procedures the investigators will obtain robust automatic classification results. Comparing SuStaIn and ANN classification performance they will decide the most convenient approach for the task. Moreover, they will apply standard dimensionality reduction techniques to extract the most salient features and repeat the procedure described above to assess whether it is possible to achieve the same results with a subset of the screenings.

The selected approach will be then repeated to the novel patients recruited during PREVIEW and those undergoing follow up analysis during the project. The only difference will be the novel set of factors included in the analysis, including all the data used in the first version but also BDNF, CSF biomarkers, and the novel features extracted from resting state EEG recordings and ERP. This will require defining a second multi-modal metrics across all features, and a consequent reshaping and retraining of the machine learning network. As described above, even for the second version of the algorithm the investigators will aim at defining the minimal set of features achieving the optimal performance. The final outcome of the algorithm will provide a prediction of the evolution of the condition for each patient based on his/her complete profile.