# Adaptive weighted dual MAML: Proposing a novel method for the automated diagnosis of partial sleep deprivation

## TL;DR

This paper introduces a new automated method using EEG signals and machine learning to diagnose partial sleep deprivation with high accuracy.

## Contribution

The novel Adaptive Weighted Dual MAML method combines two models to improve classification accuracy and robustness for sleep disorder diagnosis.

## Key findings

- The proposed model achieved 99% classification accuracy and 99% F1 score for sleep deprivation diagnosis.
- The model showed more stable performance and reduced fluctuations compared to conventional MAML.
- The method demonstrates strong robustness and generalization for unseen tasks.

## Abstract

Sleep disorders significantly disrupt normal sleep patterns and pose serious health risks. Traditional diagnostic methods, such as questionnaires and polysomnography, often require extensive time and are susceptible to errors. This highlights the need for automated detection systems to enhance diagnostic efficiency. This study proposes a novel method for the automated diagnosis of partial sleep deprivation utilizing electroencephalogram (EEG) signals.

We utilized time-frequency images obtained from continuous wavelet transforms applied to two EEG channels for the automated diagnosis of sleep disorders. Although convolutional neural networks (CNNs) are commonly used for detecting these conditions, their performance is inadequate when applied to our heterogeneous and limited-scale EEG data. To overcome these limitations, we developed a Few-Shot Learning-based Model-Agnostic Meta-Learning (FSL-based MAML) approach aimed at improving classification accuracy and generalization abilities. Our method, Adaptive Weighted Dual MAML, combines two base models—a ResNet and a CNN-Transformer—within the MAML framework, which leverages multi-shot tasks to improve the EEG signal classification,

Our findings demonstrated that the FSL-based MAML method, with a combined base model, achieves an average classification accuracy of 99% and an F1 score of 99%. Additionally, the proposed model achieved a more stable range of evaluation metrics, resulting in reduced performance fluctuations across tasks compared to the conventional MAML. This indicates stronger robustness and improved generalization to unseen tasks,

The results confirm the efficacy of our proposed approach as a robust solution for diagnosing partial sleep deprivation with enhanced accuracy and efficiency in an automated manner. This model provides a groundwork for addressing various sleep disorders through advanced EEG analysis techniques.