AI-based Ecological Monitoring of Handwriting to Early Detect Cognitive Decline

The early detection of Mild Cognitive Impairment (MCI) is fundamental to initiate treatments for delaying the onset of dementia. Currently, the Mini Mental State Examination (MMSE) is one of the most common clinical scales used by geriatricians to assess cognitive function. A deviation of 1 to 3 points from the maximum score (30) is considered as sign of relevant cognitive decline. However, objective and affordable tools are needed to complement the screening process. The quantitative analysis of handwriting represents a suitable solution, as the gesture is significantly impaired in MCI subjects in terms of time, speed, fluency and applied pressure. This works presents the development and testing of classification models able to separate subjects at risk of cognitive decline (MMSE <= 28) from controls (MMSE > 28), starting from free-content handwriting data acquired with a smart ink pen, used on paper, from which 36 indicators were computed. Data were collected in 2 phases. The former involved 45 subjects and served for models training. In the latter, data were acquired from 23 subjects in a domestic longitudinal framework and were partially used for model refinement, but mainly for testing. Three different algorithms were tried (support vector machine, random forest and Catboost) The best test performances on the longitudinal data were obtained by a Catboost classifier, achieving accuracy 93.33%, precision 88.89%, recall 100% and f1 score 94.12%. The results support the use of computerized handwriting analysis as screening tool for cognitive decline detection