Method
The planned analysis uses earlier behavioral and physiological measurements to predict subsequent task performance. [1]
Under refinement
Research questions defined; candidate features and evaluation protocol under refinement.
Sustained attention · Ongoing research
Multimodal Prediction of Attentional Lapses
I am investigating whether behavioral and physiological data can predict subsequent performance in a sustained-attention task. The project brings together behavioral measures, pupillometry, electrocardiography (ECG) and electroencephalography (EEG). My goal is to understand which sources of information support prediction, how models generalize across participants, and whether combining modalities improves on simpler baselines.
My work spans literature review, signal processing, feature extraction and the design of model validation.
Research questions
Questions guiding my research
Methodological overview
Open figure
01 · Define the prediction task
What is being predicted?
Which outcome is predicted, and what information is available at prediction time?
02 · Signal processing and quality control
Are the features suitable for prediction?
Are the features reliable and available before prediction?
Method
Physiological recordings are aligned with behavioral events. Preprocessing and feature extraction include checks for artifacts, missing data and agreement with reference measurements. [2]
Current focus
Signal-quality checks, reference comparisons and assessment of feature availability.
03 · Feature and model selection
Which features improve prediction?
Which features and observation windows improve prediction?
Method
The planned analysis uses nested cross-validation for feature and model selection. Data-dependent preprocessing is fitted on training data, and performance is assessed on held-out data. [3]
Planned
Feature discovery and model selection within the common validation design.
04 · Generalization and calibration
Does prediction generalize across participants?
Does prediction generalize to new participants, and does calibration help?
Method
The planned evaluation compares single-modality models with and without participant-specific calibration. Calibration uses earlier data from new participants; later observations are reserved for testing. [4]
Planned
Participant-level evaluation under explicit training, calibration and test-data boundaries.
05 · Multimodal fusion and baseline comparison
Does multimodal fusion improve performance?
Does multimodal prediction improve on behavioral and single-modality baselines?
Method
The planned analysis compares behavioral, single-modality and multimodal models on matched participants and observations. Ablation analysis evaluates each modality's predictive contribution by removing it and refitting the model. [5]
Planned
Comparative evaluation after the common feature-validation and single-modality analyses.
06 · Robustness and prediction lead time
How robust are the predictions, and how early?
Are predictions robust to confounding, and how early can they be made?
Method
The planned analysis tests for confounding and evaluates predictive performance at different lead times. Prediction lead time is evaluated separately from observation-window length. [4]
Planned
Control analyses specified during study design; systematic evaluation still to come.
Selected methodological references
- scikit-learn documentation. Data leakage and recommended practices.
- Mathôt & Vilotijević (2023). Methods in cognitive pupillometry: Design, preprocessing, and statistical analysis.
- scikit-learn documentation. Nested versus non-nested cross-validation.
- Scheinost et al. (2019). Ten simple rules for predictive modeling of individual differences in neuroimaging.
- Baltrušaitis, Ahuja & Morency (2019). Multimodal Machine Learning: A Survey and Taxonomy.