Premium
P1‐255: LONGITUDINAL CHANGES IN IRIS MURDOCH'S LATE DIARY AND LETTER MANUSCRIPTS: A MACHINE LEARNING ANALYSIS
Author(s) -
Kassiss Maya,
Madanat Louay,
Clarke Natasha,
Miller Dayna,
Garrard Peter
Publication year - 2019
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1016/j.jalz.2019.06.810
Subject(s) - handwriting , generalizability theory , artificial intelligence , cognition , sample (material) , period (music) , set (abstract data type) , natural language processing , psychology , computer science , art , developmental psychology , chemistry , chromatography , neuroscience , programming language , aesthetics
Background: The novelist Iris Murdoch (IM) suffered late-life cognitive decline, and Alzheimer’s pathology was confirmed post mortem. Alzheimer-like changes also appeared in the language of her final novel, which was completed before cognitive symptoms emerged. IM was also an assiduous writer of letters and diaries, large numbers of which have been collected and stored in Kingston University’s Iris Murdoch Archive. Cognitive decline is evident from those written in her last years, but there may be more subtle language changes in the period coinciding with composition of the final novel, or possibly earlier. We examined this hypothesis using a machine learning model trained on features of transcripts (linguistic features) and manuscripts (handwriting features) of diaries from 1986 onward. We predicted that a regression model trained on these features would identify, with above chance accuracy, the time-window of a sample’s composition. Methods: The training set consisted of 152 samples of IM’s diary entries written between September 1986 and August 1996. Sample dates were coded as a time variable equal to the number of months since September 1986. Digital transcripts and images of the manuscripts (the latter overlaid with a measuring grid to allow physical handwriting features to be coded) were obtained. Twenty-eight features were extracted from transcripts and manuscripts using automated text analysis tools and hand measurements respectively. Results: The features were entered in to a machine learning regression model which was trained to predict the time variable. We validated the model using 4-fold cross validation on the training set. Generalizability was evaluated by determining accuracy in assigning a time variable to documents in the hold-out set. Conclusions: The results indicate that both linguistic and physical features of IM’s handwritten text underwent change over the final decade of her life, which included many years without any symptoms or changes, reinforcing the idea that language may be a sensitive marker of presymptomatic cognitive change. The generalizability of the model suggests that the approach also has potential for aiding historical research requiring document dating.