ALLPCB
Overview
Driving forces | Unveiling artificial intelligence and machine learning in medical applications
Following approval by the U.S. Food and Drug Administration (FDA), the first wearable digital health monitors have recently reached the market and are being integrated into consumer devices such as smartwatches. Rapid advances in medical sensor technology have enabled compact, low-cost, and increasingly accurate physiological sensors to be incorporated into existing wearables.
Advanced machine learning and artificial intelligence algorithms are one of the drivers of this transition, enabling extraction and interpretation of valuable information from large volumes of data. These data often contain noise and imperfect signals (for example, electrocardiogram data from smartwatches) and are degraded by various artifacts that rule-based traditional algorithms struggle to handle.
Role of Machine Learning in Physiological Signal Analysis
Until recently, extracting reliable decisions from physiological signals produced by these sensors—and achieving a level of accuracy acceptable to regulators—was difficult or sometimes impossible. Improvements in machine learning and AI algorithms are helping engineers and scientists overcome many of these challenges.
This article examines the overall architecture of physiological signal-processing algorithms, explains the computational steps involved, and connects those steps to engineering practices built on decades of research.
Typical Machine Learning Workflow
Machine learning algorithm development mainly involves two steps. The first step is feature engineering, extracting specific numerical or mathematical features from the relevant dataset. The second step is feeding the extracted features into a well-known statistical classifier or regression algorithm, such as a support vector machine or a suitably configured traditional neural network. The trained model can then predict on new datasets. Iterative training on a reasonably labeled dataset continues until satisfactory accuracy is reached, after which the model can be deployed as a prediction engine in production.
Example: ECG Classification Workflow
How does this workflow apply to ECG signal classification? In this example, the 2017 PhysioNet Challenge dataset is used, which contains real single-lead ECG recordings. The goal is to classify patients' ECG signals into four categories: normal, atrial fibrillation, other rhythms, and too noisy.
The full workflow and steps implemented in MATLAB are shown below.
Preprocessing and Feature Engineering
Feature engineering is often the most challenging part of developing a robust machine learning algorithm. These problems cannot be treated merely as "data science" issues; domain expertise in biomedical engineering and an understanding of different physiological signals and data types are crucial when designing solutions.
Tools such as MATLAB provide domain experts with data analysis and advanced machine learning capabilities, making it easier to apply data-science functions to the problems they are solving and allowing them to focus on feature engineering. In this example, advanced wavelet techniques are used to process signals, remove noise and baseline wander such as respiratory artifacts, and extract various features of interest from the signals.
Developing Classification Models
The Classification Learner app in the Statistics and Machine Learning Toolbox offers an effective entry point for engineers and scientists who are less familiar with machine learning. Once a sufficiently rich set of relevant features is extracted from the signals, this app enables rapid exploration of various classifiers and their performance, narrowing candidate models for further optimization. These classifiers include decision trees, random forests, support vector machines, and k-nearest neighbors (KNN). It is possible to test and select strategies that provide the best classification performance for the feature set, typically evaluated using confusion matrices or AUC metrics.
In the example, this approach alone achieved roughly 80% overall accuracy across all classes (the winning entries in that challenge scored around 83%). Note that extensive effort was not invested in feature engineering or classifier tuning because the objective was to validate the approach.
In practice, dedicating time to feature engineering and classifier tuning can substantially improve accuracy. More advanced approaches such as deep learning can also be applied, where feature engineering, feature extraction, and classification are consolidated into a single training step. However, compared with traditional machine learning techniques, deep learning typically requires much larger training datasets to reach the same level of performance.
Challenges, Regulation, and Outlook
Although many common wearable devices are not yet full substitutes for FDA-cleared, clinically validated counterparts, technological and consumer trends indicate a clear direction toward that goal. The FDA has begun taking an active role in several areas, for example by simplifying regulations and encouraging the development of regulatory science through initiatives such as the digital health software pre-certification program, and by supporting device development modeling and simulation.
The vision is to transform physiological signals collected from everyday wearables into new digital biomarkers that comprehensively reflect health status. That vision is increasingly realistic thanks to advances in signal processing, machine learning, and deep learning algorithms. Workflows supported by tools like MATLAB enable medical device experts to adopt data-science techniques such as machine learning even without becoming data scientists themselves.
