Data Collection Workflow
Key Features:
Digital Biomarkers: high level information related to daily routines
Extraction: 1. Data purging > 2. Intermediate features (δ intervals); > 3. pattern matching with pre-defined rules
Data Validation Experiment
OVERALL STATISTICS OF DAILY DIGITAL BIOMARKERS EXTRACTED ACROSS THE MCI AND HC GROUPS. THE P VALUES ARE COMPUTED BASED ON KOLMOGOROV-SMIRNOV TEST WHICH HOLDS NO ASSUMPTION OF THE DATA DISTRIBUTIONS.
Performance of machine learning models in identifying MCI cases based on weekly averaged biomarker features
Distribution of biomarker records with missing features in the weekly averaged SINEW dataset
Top two rules for the MCI class extracted from Fusion ART
Based on the weekly biomarker data set, the predictive modeling experiments were conducted on a stratified 10-fold cross validation methodology.
Fusion ART performed exceptionally well achieving the best F1 score of 0.906 despite presence of noisy data with a higher missing rate with 80% of the records having more than 50% of their feature values missing.
Fusion ART appears to be well suited for handling input patterns with high missing rates and is compatible with rule-based representation whereby the knowledge learned can be translated into IF-THEN symbolic rules for further analysis.