💻 computer science

Machine Learning-Based Classification of Normal, Fast, and Sleep-Apnea-Related Respiratory Patterns Using Software-Defined Radio Sensing

This paper presents a statistically validated benchmark of Multilayer Perceptron, XGBoost, and Random Forest models for contactless respiratory pattern classification using Software-Defined Radio, demonstrating that MLP and Random Forest achieve superior accuracy over XGBoost while identifying fast and sleep-apnea patterns as the most challenging to distinguish and highlighting Random Forest's higher generalization gap due to overfitting.

Original authors: Taiwo Samuel Aina, Wisdom Okafor, Grace Okafor, Babatunde Iyaomolere

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Taiwo Samuel Aina, Wisdom Okafor, Grace Okafor, Babatunde Iyaomolere

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to listen to a person's breathing without ever touching them. Instead of using a stethoscope or a belt, this paper describes using a special radio device (called Software-Defined Radio, or SDR) that acts like a "super-sensitive ear." It sends out invisible radio waves that bounce off a person's chest. When the person breathes, their chest moves, slightly changing the radio waves. The computer analyzes these tiny changes to figure out if the person is breathing normally, breathing fast, or having a sleep apnea event (where breathing stops and starts).

The researchers wanted to see which of three different "brain" styles (Machine Learning models) was best at this listening game. They treated the radio data like a massive library of sound files and asked three different types of students to sort them into three piles: Normal, Fast, and Sleep-Apnea.

Here is how the three "students" performed, explained through simple analogies:

The Three Contestants

The Multilayer Perceptron (MLP): Think of this as a deep-thinking student who studies the patterns very carefully. It looks at the data through many layers of logic, trying to find complex connections. It's like a detective who connects tiny clues to solve a mystery.
The Random Forest (RF): Imagine a committee of 100 different experts. Each expert looks at the data and makes a guess, and the final answer is just the majority vote. Because there are so many experts, they usually get it right, but sometimes they get too confident in their own opinions.
The XGBoost: Picture this as a serial learner. It makes a guess, sees where it was wrong, and then tries again to fix that specific mistake. It does this over and over, getting slightly better with every attempt.

The Big Test

The researchers gave all three students the same homework: a huge dataset of breathing recordings taken from different angles (like standing in front of the sensor, to the side, or behind it). They split the data so the students studied 80% of it and were tested on the remaining 20% they had never seen before. This is crucial because it prevents the students from just memorizing the answers.

The Results:

The Winners: The MLP (the deep thinker) and the Random Forest (the committee) tied for first place. They both got about 90% of the answers correct on the new test.
The Runner-Up: XGBoost (the serial learner) came in third, getting about 87% correct. The researchers used a statistical "ruler" to prove that this difference wasn't just bad luck; XGBoost was genuinely slightly worse than the other two on this specific dataset.

The Tricky Part: Where They Got Confused

Even the winners made mistakes, and they all struggled with the same specific confusion: Fast breathing vs. Sleep Apnea.

The Analogy: Imagine trying to tell the difference between someone jogging in place (Fast) and someone gasping for air after a sprint (Sleep Apnea). To a radio sensor, these two sound very similar because the chest is moving quickly in both cases.
The Result: The "Fast" breathing category was the hardest for everyone to identify. The models often mixed up "Fast" breathing with "Sleep Apnea" breathing.
The Easy Part: "Normal" breathing was the easiest for all models to identify, likely because it is steady and predictable, like a calm metronome.

The "Overconfidence" Trap

One of the most interesting findings was about the Random Forest.

In Practice: When the students were studying (training), the Random Forest got 100% of the answers right. It was perfect.
On the Test: When faced with new, unseen data, its score dropped to 90%.
The Lesson: The paper explains that the Random Forest was "overfitting." Think of it like a student who memorized the textbook word-for-word but didn't understand the concepts. When the test questions were slightly different from the book, the student stumbled. The other models (MLP and XGBoost) didn't memorize as obsessively, so their drop in score was smaller, showing they were better at handling new situations.

What This Means (According to the Paper)

The paper concludes that for this specific job of listening to breathing via radio waves:

MLP and Random Forest are the best tools right now, both achieving about 90% accuracy.
Random Forest is a strong candidate for real-world use (like in a home) because it is generally easier to run on simple computers, but it needs to be tuned so it doesn't memorize the training data too strictly.
The "Fast vs. Apnea" confusion is a real hurdle. The paper suggests that until the models can tell these two apart more clearly, this technology is best used as a first alert system (to flag a recording for a doctor to check) rather than a final diagnosis tool on its own.

In short, the researchers built a radio "stethoscope," tested three different AI brains, and found that while they are very good at spotting normal breathing, they still need help distinguishing between someone breathing fast and someone having a breathing pause.

Technical Summary: Machine Learning-Based Classification of Respiratory Patterns Using SDR

Problem Statement
Respiratory monitoring is critical for the early detection of conditions such as sleep apnea, COPD, and cardiac disease. However, conventional diagnostic methods like polysomnography are intrusive, require specialist equipment, and are unsuitable for continuous home use. While non-contact Radio Frequency (RF) sensing using Software-Defined Radio (SDR) has emerged as a viable alternative for capturing respiratory waveforms, there is a significant gap in the literature regarding the comparative evaluation of different machine learning architectures. Specifically, few studies provide a direct, same-dataset comparison of classical ensemble methods against deep learning under a rigorous, leakage-controlled evaluation protocol with full per-class reporting and statistical validation. This lack of controlled evidence hinders practitioners from making informed decisions regarding model selection for edge-deployed healthcare systems, where trade-offs exist between interpretability, accuracy, and computational cost.

Methodology
This study benchmarks three distinct model families—Multilayer Perceptron (MLP), XGBoost, and Random Forest (RF)—on a publicly available SDR dataset collected using a USRP X310 device. The dataset comprises three respiratory classes: Normal Respiration (NR), Fast Respiration (FR), and Sleep-Apnea-Related (SAR) patterns, recorded at three sensor-to-subject angles (45°, 90°, and 180°).

Data Preprocessing: Raw recordings were processed to extract 2,000-dimensional feature vectors, combining time-domain statistics (mean, variance, peak-to-peak amplitude) and frequency-domain descriptors (dominant frequency, spectral power).
Leakage Control: To prevent information leakage common in time-series data, the dataset was split at the recording-file level (80% training, 20% held-out test). This ensures that no samples from the same continuous breathing session appear in both training and testing sets, providing a conservative estimate of generalization.
Model Configurations:
- Random Forest: 100 trees with unconstrained depth (no regularization on tree complexity).
- XGBoost: 100 estimators, learning rate 0.1, max depth 6, with L2 regularization ( $\lambda=1$ ) but no minimum split gain constraint ( $\gamma=0$ ).
- MLP: A deep network with two hidden layers (512 and 256 neurons), ReLU activations, dropout (0.2), and Adam optimization.
Evaluation: Performance was assessed using accuracy, precision, recall, F1-score, and confusion matrices. Statistical significance between model accuracies was determined using a two-proportion z-test. An empirical generalization gap ( $\Delta = \text{Err}_{test} - \text{Err}_{train}$ ) was calculated to analyze overfitting behavior.

Key Results

Overall Performance: On the held-out test set, MLP and Random Forest achieved the highest accuracy (0.90) with macro-F1 scores of 0.89–0.90. XGBoost achieved a lower accuracy of 0.87.
Statistical Significance: The difference between XGBoost and the other two models was statistically significant ( $z=5.39, p < 0.001$ ). There was no statistically significant difference between MLP and Random Forest ( $z=0.00, p=1.00$ ).
Class-Specific Performance:
- Normal Respiration (NR): Consistently the most reliably classified class across all models (test accuracy 90–93%).
- Fast Respiration (FR): The most difficult class to classify, with the lowest recall across all models.
- Sleep-Apnea-Related (SAR): Well-identified by MLP and RF, but prone to confusion with FR in XGBoost.
Confusion Boundaries: The dominant misclassification boundary for XGBoost and Random Forest was between FR and SAR. For MLP, the primary confusion occurred between FR and NR.
Generalization Gap:
- Random Forest exhibited the largest generalization gap ( $\Delta \approx 0.10$ ), achieving perfect training accuracy (1.00) but dropping to 0.90 on the test set. This indicates significant overfitting due to the unconstrained tree ensemble structure.
- XGBoost showed a moderate gap ( $\Delta \approx 0.06$ ).
- MLP showed the smallest gap ( $\Delta \approx 0.02$ ), attributed to dropout regularization.

Significance and Claims
The paper claims to provide the first statistically validated, controlled benchmark of these three model families on a common SDR respiratory dataset under a leakage-managed protocol. The study highlights that while Random Forest can achieve performance comparable to deep learning (MLP) on test data, its near-perfect training accuracy masks a high sensitivity to the training sample, a risk that must be managed via regularization for edge deployment.

The authors identify the FR/SAR classification boundary as the primary limitation for current SDR-based systems, suggesting that fast breathing and sleep apnea events share acoustic or spectral characteristics that are difficult to separate using amplitude and spectral features alone. Consequently, the authors position these models as suitable for triage or pre-screening roles (flagging recordings for clinician review) rather than as standalone diagnostic tools, given the current misclassification rates (7–12%) between critical classes.

Limitations and Future Directions
The authors acknowledge that the study relies on a single dataset and a file-level split rather than a subject-level split, which may not fully guarantee independence. They recommend future work prioritize:

Subject-aware cross-validation to obtain robust generalization gap estimates.
Hyperparameter tuning for XGBoost and Random Forest to determine if baseline configurations represent optimal performance.
Architectural enhancements (e.g., hybrid CNN-LSTM) or additional temporal features to resolve the persistent FR/SAR confusion.
Real-time inference assessment on edge hardware to validate deployment feasibility.