Empirical comparison of deep learning models for fNIRS pain decoding

Area of Science:

• Neuroscience
• Medical Technology
• Artificial Intelligence

Background:

• Objective pain assessment is crucial for non-verbal patients, as clinical judgment can be subjective and unreliable.
• Functional near-infrared spectroscopy (fNIRS) offers a promising non-invasive method for measuring brain activity related to pain.
• Previous research utilized machine learning with hand-crafted features for pain assessment from fNIRS data.

Purpose of the Study:

• To explore deep learning models (CNN, LSTM, CNN-LSTM) for automatic feature extraction from fNIRS data for pain assessment.
• To compare the performance of deep learning models against classical machine learning models using hand-crafted features.
• To investigate the potential of deep learning in developing objective pain assessment tools for non-verbal individuals.

Main Methods:

• Utilized functional near-infrared spectroscopy (fNIRS) to collect brain activity data during pain stimuli.
• Applied deep learning models: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a hybrid CNN-LSTM model.
• Compared deep learning models with classical machine learning approaches employing hand-crafted features.

Main Results:

• Deep learning models demonstrated effective identification of different pain types using fNIRS data.
• The hybrid CNN-LSTM model achieved the highest accuracy (91.2%) in pain assessment.
• Statistical analysis confirmed that deep learning models significantly improved accuracy compared to baseline methods.

Conclusions:

• Deep learning models can automatically learn relevant features from fNIRS data, reducing reliance on manual feature engineering.
• The CNN-LSTM model presents a viable approach for objective pain assessment in non-verbal patients.
• Further research is necessary to validate the generalizability of this method across diverse populations and real-world clinical settings.