Cognitive Impairment Detection through Recurrent Neural Networks and Mobile Health Technologies
Author(s)
Melissa Ouellet, BSc, MSc1, Clauirton Siebra, PhD2.
1Digital Health Cluster, Hasso Plattner Institute, Potsdam, Germany, 2Informatics Center, Federal University of Paraíba, João Pessoa, Brazil.
1Digital Health Cluster, Hasso Plattner Institute, Potsdam, Germany, 2Informatics Center, Federal University of Paraíba, João Pessoa, Brazil.
Presentation Documents
OBJECTIVES: Our study investigates the potential of leveraging mobile health technologies and recurrent neural networks (RNNs) to detect cognitive impairment (CI) in individuals.
METHODS: Using 2011-2014 National Health and Nutrition Examination Survey data, we aggregated minute-level ActiGraph GT3X+ accelerometer data to hourly levels, identifying activity periods (sleep, light, moderate, vigorous) based on Monitor Independent Movement Summary (MIMS) thresholds. Features included MIMS statistics (mean, SD, min, max, median, kurtosis, entropy) and day of the week. Cognitive tests comprised Immediate and Delayed Recall, Animal Fluency, and Digital Symbol Substitution. Scores were converted to z-scores and summed, with a cutoff of -2.20 (lowest quartile) indicating CI, while higher scores were non-impaired. We predicted CI using Logistic Regression, Random forest and Gradient Boosting classifiers, and explored temporal patterns with RNNs (standard RNN, Long Short-Term Memory, Gated Recurrent Unit) using a 24-hour sliding window. Models were adjusted for class imbalance, safeguarded against data leakage, and evaluated on a 20% test set using Monte Carlo cross-validation.
RESULTS: Our sample included 2,559 individuals (aged 60-80), 24.6% cognitively impaired, with 429,912 hourly data points over 7 days. The Gated Recurrent Unit model achieved the highest performance (area under the curve, AUC=0.966). Physical activity patterns were compared between actual and predicted CI and non-CI groups across hourly, daily, and weekly levels. CI cases showed greater sleep and reduced moderate-to-vigorous activity, with light and sedentary activity having the greatest impact on AUC (18.8% and 16.8% respectively). The model preserved hourly patterns but left room for improvement in capturing daily and weekly trends.
CONCLUSIONS: This research highlights the distinct behavioral activity patterns observed in CI sufferers and their impact on model performance. These findings emphasize the potential of wearables and sequence modeling in uncovering behavioral patterns linked to CI, offering opportunities to enhance predictive models and improve intervention strategies.
METHODS: Using 2011-2014 National Health and Nutrition Examination Survey data, we aggregated minute-level ActiGraph GT3X+ accelerometer data to hourly levels, identifying activity periods (sleep, light, moderate, vigorous) based on Monitor Independent Movement Summary (MIMS) thresholds. Features included MIMS statistics (mean, SD, min, max, median, kurtosis, entropy) and day of the week. Cognitive tests comprised Immediate and Delayed Recall, Animal Fluency, and Digital Symbol Substitution. Scores were converted to z-scores and summed, with a cutoff of -2.20 (lowest quartile) indicating CI, while higher scores were non-impaired. We predicted CI using Logistic Regression, Random forest and Gradient Boosting classifiers, and explored temporal patterns with RNNs (standard RNN, Long Short-Term Memory, Gated Recurrent Unit) using a 24-hour sliding window. Models were adjusted for class imbalance, safeguarded against data leakage, and evaluated on a 20% test set using Monte Carlo cross-validation.
RESULTS: Our sample included 2,559 individuals (aged 60-80), 24.6% cognitively impaired, with 429,912 hourly data points over 7 days. The Gated Recurrent Unit model achieved the highest performance (area under the curve, AUC=0.966). Physical activity patterns were compared between actual and predicted CI and non-CI groups across hourly, daily, and weekly levels. CI cases showed greater sleep and reduced moderate-to-vigorous activity, with light and sedentary activity having the greatest impact on AUC (18.8% and 16.8% respectively). The model preserved hourly patterns but left room for improvement in capturing daily and weekly trends.
CONCLUSIONS: This research highlights the distinct behavioral activity patterns observed in CI sufferers and their impact on model performance. These findings emphasize the potential of wearables and sequence modeling in uncovering behavioral patterns linked to CI, offering opportunities to enhance predictive models and improve intervention strategies.
Conference/Value in Health Info
2025-05, ISPOR 2025, Montréal, Quebec, CA
Value in Health, Volume 28, Issue S1
Code
MSR118
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Neurological Disorders