The Machine Learning Model
If you wish to work to replicate our research or look into our model, here is the link to the repository.
Our Fibrometer project is at the forefront of diagnostic innovation and we utilize advanced machine learning to streamline the detection of Fibromyalgia Syndrome (FMS). Designed to overcome the challenges of traditional diagnostic methods, our tool synthesizes data from biosensors and polysomnography (PSG) to analyze key metrics like heart rate variability (HRV), galvanic skin response (GSR), sleep motion, and EEG readings. The model integrates these multimodal inputs to allow the Fibrometer to provide faster, cost-effective, and more accurate results than conventional approaches.
The machine learning model at the heart of the Fibrometer is designed to process this data pipeline to identify patterns and biomarkers indicative of FMS. During the diagnostic process, the tool consolidates complex data into actionable insights for primary care physicians and patients, displayed through an intuitive interface within our app. The goal is to enhance both diagnostic accuracy and patient understanding, creating a bridge between advanced technology and accessible healthcare.
Overview of the Fibrometer Machine Learning Model
The Fibrometer machine learning model is the diagnostic tool we designed to identify Fibromyalgia Syndrome (FMS) accurately and efficiently. The model is vital part of the device and leverages multimodal data collected from the biosensors to analyze physiological patterns that correlate with FMS symptoms. Key components include data collected from EEG readings, heart rate variability (HRV) sensors, galvanic skin response (GSR) patches, and sleep motion detectors. These inputs are synthesized through a pipeline that integrates multiple advanced machine learning algorithms as components.
The process begins with data collection from non-invasive biosensors which ensures comfort for patients while capturing a comprehensive dataset. Preprocessing steps include signal denoising and normalization to standardize input quality across modalities. Key features, such as time-series trends and frequency-domain characteristics, are extracted via a U-Net model and converted into numerical sets to be passed down.
At the heart of the pipeline, multimodal data fusion combines inputs from different sources through a fusion layer that uses concatenation and attention mechanisms. This approach emphasizes the most relevant data points for a holistic view of the patient’s physiological condition. The fused data is analyzed using a Graph Neural Network (GNN) to process sequential, spatial, and relational data. The Convolutional Neural Network (CNN) identifies specific patterns indicative of FMS symptoms. The final output is a classification score that finds the mean squared error (MSE) between the patient's metrics and the ideal metrics of a FMS patient of the individual’s demographics.
Validation Experiment: Multimodal Data Processing for Fibromyalgia Diagnosis
This study validated the Fibrometer’s machine learning pipeline by testing its capability to process multimodal data and detect patterns indicative of fibromyalgia. The experiment utilized 3 biosensors’ data from electroencephalogram (EEG), heart rate variability (HRV), and galvanic skin response (GSR) signals. Due to the challenge of limited datasets, a Generative Adversarial Network (GAN) was employed to create synthetic but realistic samples which increased the diversity of the data and enabled the model to capture subtle variations associated with fibromyalgia.
The pipeline integrated four primary components to enhance data processing and classification. The GAN generated synthetic data which was fed into a Convolutional Neural Network (CNN) to extract hierarchical features from time-series graphs and identified both local and global signal patterns. Autoencoders were utilized to reduce noise and enhance the clarity of biomarkers so as to only preserve the most relevant characteristics of the physiological signals. A Graph Neural Network (GNN) then fused the multimodal data and unified representations that captured temporal dynamics and cross-sensor relationships critical for identifying fibromyalgia-specific patterns.
The model was trained on a dataset of 200 samples which were evenly split between fibromyalgia and control cases and used fivefold cross-validation. It achieved an impressive accuracy of 96.5%, correctly classifying 193 out of 200 samples. Performance metrics, including precision at 95.8%, recall at 97.2%, and an F1-score of 96.5%, demonstrated the model’s reliability. The model also received a ROC-AUC score of 98.0% which confirms its strong ability to distinguish between fibromyalgia and non-fibromyalgia cases.
For an in-depth explanation, please reference this attachment.
