Machine learning approaches to studying the role of cognitive reserve in conversion from mild cognitive impairment to dementia

OBJECTIVES

The overall aim of the present study was to explore the role of cognitive reserve (CR) in the conversion from mild cognitive impairment (MCI) to dementia. We used traditional and machine learning (ML) techniques to compare converter and nonconverter participants. We also discuss the predictive value of CR proxies in relation to the ML model performance.

METHODS

In total, 169 participants completed the longitudinal study. Participants were divided into a control group and three MCI subgroups, according to the Petersen criteria for diagnosis. Information about the participants was compared using nine ML classification techniques. Seven relevant performance metrics were computed in order to evaluate the accuracy of prediction regarding converter and nonconverter participants.

RESULTS

ML algorithms applied to socio-demographic, basic health, and CR proxy data enabled prediction of conversion to dementia. The best performing models were the gradient boosting classifier (accuracy (ACC) = 0.93; F1 = 0.86, and Cohen κ = 0.82) and random forest classifier (ACC = 0.92; F1 = 0.79, and Cohen κ = 0.71). Use of ML techniques corroborated the protective role of CR as a mediator of conversion to dementia, whereby participants with more years of education and higher vocabulary scores survived longer without developing dementia.

CONCLUSIONS

We used ML approaches to explore the role of CR in conversion from MCI to dementia. The findings indicate the potential value of ML algorithms for detecting risk of conversion to dementia in cognitive aging and CR studies. Further research is required to develop an ML-based procedure that can be used to make robust predictions.

Diagnostic transitions in mild cognitive impairment by use of simple Markov models

BACKGROUND

Mild cognitive impairment (MCI) is a complex entity, which can involve persistence of the symptoms, conversion to dementia or improvement. The aim was to study the transitions between normal cognitive ageing and three MCI subtypes by using Markov transition models for different intervals between baseline and the follow-up assessment.

METHODS

A total of 294 participants over 50 years old attending primary care centres were assessed and diagnosed at baseline as multi-domain amnestic MCI (22 participants), single domain amnestic MCI (44), non-amnestic MCI (non-amnestic MCI) (26) or controls (202). We adopted an overlapping interval strategy by constructing six different mid-point time intervals according to the time between the baseline and the follow-up assessment. We used Markov transition models to study diagnostic changes in the groups in the different time intervals

RESULTS

The rate of change was lowest in the control group. In the single domain amnestic MCI and non-amnestic MCI groups, the same diagnosis was usually retained or changed to normal cognitive functioning. In the multi-domain amnestic MCI group, the rate of transition to normal functioning was lowest, and the conversion to dementia was the highest of all groups. The best fit to the Markov models was found for the period between 18-21 months, whereas the worst fit was for the period between 9-15 months

CONCLUSIONS

Markov models provide a comprehensive view of transitions between MCI and normal cognitive functioning. Time interval strategies seem to provide a good opportunity to monitor diagnostic transitions, although wider intervals including subsequent assessments are needed. The low rates of conversion to dementia are discussed.