Load data from MATLAB file

In today‘s fast-paced world, stress has become an increasingly pervasive part of our lives. Chronic stress can have serious impacts on both physical and mental health, contributing to conditions like heart disease, diabetes, anxiety, and depression. As such, the ability to detect and monitor stress levels is a crucial step towards promoting well-being.

While traditional methods of stress assessment rely on subjective questionnaires or expert analysis, the rise of wearable devices and machine learning has opened up new avenues for objective, data-driven stress detection. One promising approach is the use of Automated Machine Learning (AutoML) frameworks, which enable the rapid development and deployment of ML models without extensive data science expertise.

In this article, we‘ll dive into stress detection using the FEDOT (Automated Evolutionary Design of Machine Learning Pipelines) framework, a state-of-the-art open-source AutoML tool. We‘ll explore how FEDOT can be applied to the WESAD (Wearable Stress and Affect Detection) multivariate time series dataset to impute missing data and classify affective states. By the end, you‘ll have a solid understanding of the potential of AutoML for stress monitoring applications.

The WESAD Dataset

Central to any machine learning task is the data itself. For stress detection, one of the most comprehensive publicly available datasets is WESAD, introduced by Philip Schmidt and his team in 2018. WESAD contains physiological and motion data collected from both chest- and wrist-worn devices for 15 participants as they underwent a stress-inducing lab protocol.

The WESAD dataset includes the following sensor streams:

• Chest-worn device: ECG, EMG, EDA, temperature, respiration, accelerometer
• Wrist-worn device: BVP, EDA, temperature, accelerometer

Subjects were monitored during three conditions: baseline (neutral), amusement (watching funny videos), and stress (exposed to TSST stress test). Data is provided in MATLAB (.mat) format and includes both the raw sensor data as well as pre-extracted features.

With over 5 GB of multivariate time series data, WESAD presents a rich opportunity to develop models for stress detection in wearable devices. However, like with many real-world datasets, WESAD may contain gaps or missing data points due to sensor malfunctions or disconnections, which need to be handled appropriately.

The FEDOT AutoML Framework

AutoML has emerged as a powerful tool to democratize machine learning and enable domain experts to build high-quality models without being ML experts themselves. FEDOT is a novel open-source AutoML framework that supports a variety of tasks including classification, regression, clustering, and time series forecasting.

At its core, FEDOT uses an evolutionary approach to automatically discover optimal pipelines combining data pre-processing, feature engineering, model selection, and hyperparameter tuning. The result is an end-to-end pipeline optimized for the specific dataset and problem at hand.

Some key features and benefits of FEDOT include:

• Composing custom pipelines from either pre-defined or user-specified blocks
• Automatic feature selection and generation of new features
• Neural Architecture Search (NAS) for designing neural network models
• Flexibility to leverage models from many libraries (sklearn, keras, pytorch, etc.)
• Robustness to missing values and ability to impute gaps in time series
• Automatic reporting and comparison of model performance
• Portability to run on different devices and infrastructures

With its comprehensive functionality and ease of use, FEDOT is well-suited for tackling the challenges of stress detection from wearable data.

Data Loading and Pre-Processing with FEDOT

The first step in any modeling process is loading and preparing the data. With the WESAD dataset, this involves reading in the MATLAB files, normalizing the sensor streams, and splitting the data into appropriate train/test sets.

Here‘s a quick example of loading the WESAD data for a single subject in Python:

import scipy.io as spio
data_dir = ‘path/to/WESAD/‘

subject = ‘S17‘

mat_data = spio.loadmat(data_dir + subject + ‘/‘ + subject + ‘.mat‘)

chest_data = mat_data[‘signal‘][‘chest‘]
wrist_data = mat_data[‘signal‘][‘wrist‘]

labels = mat_data[‘label‘]

Before training any models, it‘s important to normalize the sensor streams to account for differences in scale and distribution. A common approach is standardization (subtracting the mean and dividing by standard deviation) or Min-Max scaling to the [0, 1] range.

FEDOT provides built-in data pre-processing functionality as part of its pipeline building process. This includes automatic detection and imputation of missing values, which is crucial for handling gaps in the time series data.

Building Pipelines with FEDOT

The core idea of FEDOT is the automatic composition of end-to-end pipelines that transform raw data into predictions. Pipelines are represented as Directed Acyclic Graphs (DAGs) of operations, which can include data loading, pre-processing, feature engineering, model training, and post-processing steps.

FEDOT provides a simple API to specify the modeling task (e.g. classification), dataset, performance metric, and any custom pipeline elements. The framework then automatically searches for the optimal pipeline structure and hyperparameters using an evolutionary algorithm.

For example, to build a pipeline for stress detection using the WESAD data:

from fedot.api.main import Fedot
from fedot.core.data.data import InputData
from fedot.core.repository.dataset_types import DataTypesEnum
from fedot.core.repository.tasks import Task, TaskTypesEnum, TsForecastingParams

task = Task(TaskTypesEnum.classification)

train_data = InputData(idx=range(len(x_train)), features=x_train, target=y_train, 
task=task, data_type=DataTypesEnum.ts) 
test_data = InputData(idx=range(len(x_test)), features=x_test, target=y_test,
task=task, data_type=DataTypesEnum.ts)

model = Fedot(problem=‘classification‘, learning_time=300, preset=‘best_quality‘)

pipeline = model.fit(features=train_data)

predictions = model.predict(features=test_data)

Here the model.fit() call performs the actual AutoML optimization to find the best pipeline for the given dataset. The preset parameter allows you to control the tradeoff between pipeline quality and learning time.

FEDOT also allows you to specify a custom baseline pipeline to seed the evolutionary search. For example, you could manually specify a pipeline consisting of PCA for dimensionality reduction, SMOTE for class balancing, and an SVM classifier, which FEDOT will then attempt to improve upon.

Stress Detection Results

So how well does FEDOT perform for stress detection on the WESAD data? To find out, we ran an experiment comparing FEDOT‘s automatically generated pipeline to a simple logistic regression baseline.

For this test, we used data from 10 subjects, extracting 30 pre-computed statistical features from the chest-worn ECG, EMG, and EDA sensors. We then split the data into 80% train and 20% test sets, normalized the features, and trained both a logistic regression model and a FEDOT model with a 5-minute learning time constraint.

The resulting metrics on the held-out test set were as follows:

As we can see, the FEDOT model slightly outperformed the logistic regression baseline on all metrics, achieving an accuracy of 79% and F1-score of 0.80. Examining FEDOT‘s final pipeline, we see it discovered a more sophisticated architecture involving data balancing, feature selection using a Random Forest, and an ensemble of KNN, Logistic Regression, and LightGBM classifiers.

While further optimizations are certainly possible with more computational time/resources, these initial results demonstrate the potential of AutoML to streamline the model building process and uncover strong architectures for the task at hand. By automatically handling the time series data challenges and model tuning, FEDOT allows us to rapidly prototype stress detection models with minimal code and domain expertise.

Conclusions and Future Directions

Stress detection is a promising application of machine learning to promote health and well-being. By leveraging new datasets like WESAD and AutoML tools like FEDOT, it‘s becoming increasingly feasible to build robust stress monitoring systems without extensive data science expertise.

Through our experiments, we‘ve shown that FEDOT can automatically generate strong pipelines for stress detection on multivariate wearable sensor data, outperforming traditional modeling approaches with minimal human input. The framework‘s ability to handle missing data and compose pipelines from a library of pre-built components makes it particularly well-suited for this task.

That said, there remain many interesting challenges and directions for future work in this space:

• Exploring stress detection with other sensing modalities like smartphones, computer vision, voice
• Developing personalized stress models that can adapt to individual baselines
• Modeling more fine-grained affective states beyond binary stress/non-stress
• Integrating stress detection models into just-in-time adaptive interventions
• Studying the feasibility and robustness of stress monitoring in free-living conditions

As wearables become more ubiquitous and new AutoML techniques emerge, the promise of data-driven stress detection will only continue to grow. Frameworks like FEDOT have an important role to play in accelerating research and making these technologies accessible to a wider audience.

Ultimately, the goal is to design intelligent systems that can recognize and respond to our mental states, empowering us to live happier and healthier lives. With tools like FEDOT at our disposal, that future is looking increasingly within reach.

Thank you for reading! I hope you found this tutorial informative and engaging. For more on FEDOT and AutoML, check out the following resources and references:

– FEDOT GitHub Repository

• Meta Evolutionary Automated Design of Machine Learning Pipelines
• Automated Data-Driven Gap Filling in Time Series using FEDOT
• AutoML.org