Biomarkers | JoVE Visualize

Area of Science:

Neuroscience
Computer Science
Transportation Science

Background:

Daily cognitive performance fluctuations are increasingly measured but their impact on real-world behaviors like driving is unclear.
Traditional linear models struggle to analyze complex driving data due to nonlinear relationships and temporal dependencies.
This study introduces a novel hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model to predict cognitive functioning from driving metrics.

Purpose of the Study:

To assess the efficacy of a hybrid CNN-LSTM model in predicting cognitive functioning based on driving behavior metrics.
To explore the relationship between daily cognitive variations and driving patterns.
To develop a novel method for analyzing complex, sequential driving data.

Main Methods:

Collected simultaneous data from smartphone-based cognitive tests (four times daily) and driving behavior metrics over months.
Engineered features such as Distance Ratio and Short Trip Ratio, applied normalization and noise augmentation.
Utilized a sliding window approach to create sequential data and trained a hybrid CNN-LSTM model with Conv1D and stacked LSTM layers.
Employed Adam optimizer, Huber loss, and early stopping for model training and evaluation using MAE and R-squared.

Main Results:

Feature analysis indicated weak but significant correlations for Distance Ratio (negative) and Short Trip Ratio (positive) with cognitive performance, with Distance Ratio being the most predictive.
The CNN-LSTM model achieved high accuracy, with an R-squared of 0.9856 and Mean Absolute Error of 0.0321.
Model predictions closely matched actual cognitive values, demonstrating strong generalization and robustness, with learning curves confirming no overfitting.

Conclusions:

Hybrid CNN-LSTM models are effective in capturing nonlinear and temporal dependencies between cognitive functioning and driving behavior.
Feature engineering and advanced modeling techniques significantly enhance predictive capabilities for sequential data.
Future research can explore attention mechanisms and external factors (weather, traffic) to further improve predictive accuracy in applications like predictive analytics.