Treffer: Predicting progression of Mild Cognitive Impairment patients through four distinctive groups obtained by a dimensionality reduction algorithm.

Background: The application of machine learning (ML) algorithms can help improve diagnosis and predict progression/conversion to Alzheimer's disease (AD) at prodromal stages (Mild Cognitive Impairment – MCI)1. Due to the heterogeneity of the population at early stages and the large amount of data available, the use of unsupervised learning techniques can show distinctive patterns and help move towards personalized medicine2,3. In this work, we implemented a dimension reduction algorithm with the intent of characterizing MCI patients and develop a multistate model of progression to AD. Method: We processed 57 variables including demographic, clinical, neuroimaging, fluid biomarkers, and genetic information from a mixed population of 1041 MCI patients (544 followed at the Dementia clinic of Centro Hospitalar Universitário de Coimbra (CHUC) and 497 obtained by ADNI). Data was normalized and scaled. Afterward, we applied a dimensionality reduction algorithm (UMAP) in Python. Missing data was handled accordingly. From those with longitudinal measures (n = 351), we generated transition probability estimates by Non‐Markov Multi‐State modelling, based on the change in the status of diagnosis through time by the MMSE z‐score at each visit. Result: We obtained 4 clusters associated to: 1) increased cognitive reserve; 2) AD signature; 3) behavioral/mood alterations; and 4) mixed presentation with cardiovascular risk factors. We obtained theoretical progression rates (from faster to slower) in MCI individuals categorized in groups 2, 4, 1 and 3. After the onset of symptoms, initial decline within MCI is expected to occur in the 1st to 3rd year (in order: groups 2, 1, 4 and 3) with 60‐83% probability, and from MCI to AD from those more advanced as early as 5 (group 2) to 66 months for those in an initial phase with slower decline (group 3). Conclusion: Our results show added value for the application of ML techniques in processing data with the intent to improve the characterization and follow‐up of MCI patients. We suggest that these four groups should be targeted and considered for monitoring patient care. 1. Silva‐Spínola A, et al. Biomedicines 2022; 2.Chun MY, et al. Front Aging Neurosc. 2022; 3.Tam A, et al. J Prev Alz Dis. 2022. [ABSTRACT FROM AUTHOR]