Supervisors info:
Ηλίας Σ. Μανωλάκος, Καθηγητής, Τμήμα Πληροφορικής και Τηλεπικοινωνιών, Εθνικό και Καποδιστριακό Πανεπιστήμιο Αθηνών
Summary:
Early and accurate diagnosis of Mild Cognitive Impairment (MCI) is crucial as it is often a
precursor stage to Alzheimer’s disease, allowing for timely intervention and treatment that
can potentially slow the progression of cognitive decline. A non-invasive diagnostic approach
would be highly beneficial, as it would be more comfortable and accessible for patients,
increasing the likelihood of early detection and enabling prompt action to preserve cognitive
function and quality of life. Developing easy and non-invasive digital medicine methods for
diagnosing MCI could lead to earlier intervention and better management of this condition,
which is a major step towards preventing or delaying the onset of Alzheimer’s disease and
its devastating effects on individuals and their families.
Actigraphy, which involves the continuous monitoring of physical activity and rest-activity
cycles using wearable devices, can offer valuable insights into the presence of MCI. This
graduate thesis project was conducted in collaboration with the Aiginition Longitudinal
Biomarker Investigation of Neurodegeneration (ALBION) study at the University of Athens,
Greece. 7-day actigraphy time series data from individuals with normal cognitive function
and those with MCI were collected via an actigraph device. We extracted parametric
features, including Mesor (average activity), Amplitude (highest magnitude of activity), and
Acrophase (the timing of the largest peak), via a multi-cosinor analysis. We then developed
a machine-learning pipeline that uses exclusively these parametric features of recent
actigraphy to classify MCI vs. Normal samples, trying to predict the “Diagnosis” endpoint, a
binary label assigned by expert clinicians after a patient’s biannual evaluation. Our workflow
performs a wide range of signal preprocessing steps, examining a variety of machine
learning setups, and thus explores the full range of the parametric feature set’s class
separation capabilities. The selected models are analyzed via a thorough SHAP (SHapley
Additive exPlanations) visualization pipeline to distinguish the most informative features and
their effect on the classification results, providing valuable insights to clinicians.
We have established that for the ALBION dataset in its current state, lower Mesor,
Acrophase, and Amplitude feature values are associated with the MCI class. Specifically,
lower Mesor, Acrophase, and Amplitude values have a strong correlation with the MCI
samples, while higher Acrophase is associated with the Normal class. Our results are in
general agreement with the available recent literature. Moreover, we identified a collection
of machine learning models reaching expected specificity 0.88 at recall (sensitivity) level of
0.5 on unseen data. Notably, the developed computational pipeline allows us to fully explore
the very large space of predictive model solutions in terms of their sensitivity vs. specificity
tradeoff. As ALVION is an ongoing longitudinal study, we expect the generalization abilities
of interpretable machine learning models to get better as the dataset size increases and the
class balance improves over the years.
Keywords:
Alzheimer's disease, Minor Cognitive Impairment, MCI, actigraphy, time series, machine learning, classification, data science, cosinor analysis, Shapley Values
