In the field of wearable health monitoring, the continuous estimation of oxygen saturation (SpO₂) from photoplethysmography (PPG) signals represents a major technical challenge. The quality of dual-wavelength PPG signals (red and infrared) frequently degrades due to user movement, sensor pressure, or changing physiological conditions, distorting wave morphology and reducing the accuracy of SpO₂ predictors. Until now, conventional methods of denoising and reconstruction focused on optimizing waveform fidelity or heart rate characteristics, but neglected the preservation of spectral information critical for oxygenation. This article explores how a time-frequency reconstruction approach guided by an SpO₂ predictor can overcome those limitations, offering a robust solution for wearable devices and opening up new opportunities in custom software development for digital health.
The technical proposal is based on a staged training framework that integrates a masked reconstruction model with losses in the time and frequency domain, using the Short Time Fourier Transform (STFT). The innovation is the incorporation of a pre-trained SpO₂ predictor as an additional constraint, forcing the reconstructed signals to preserve the information relevant to oxygenation instead of simply minimizing the waveform reconstruction error. This approach not only improves accuracy (achieving absolute mean errors per subject of 2,882% in public datasets and 2,359% in private datasets), but also lays the foundation for high-value clinical and continuous monitoring applications.
Behind this technology is a fundamental principle: artificial intelligence can learn to distinguish between real physiological artifacts and patterns if provided with the right constraints. The SpO₂ predictor acts as a 'guide' that forces the reconstructer to maintain coherence with the variable of interest. This idea can be directly extrapolated to other domains where signal quality is critical, such as artificial intelligence for companies that develop medical devices or remote monitoring systems. At Q2BSTUDIO, we understand that integrating AI models into signal processing pipelines requires not only advanced algorithms, but also a software architecture that ensures scalability, security, and real-time performance.
From a business perspective, the ability to reconstruct high-quality PPG signals from noisy data has a direct impact on the reliability of wearable devices. Companies that develop activity trackers, smartwatches, or implantable medical sensors need tailored software solutions that incorporate these algorithms without compromising autonomy or latency. In addition, managing the data generated – often terabytes daily in large-scale deployments – requires robust AWS and Azure cloud services that enable distributed model training, secure signal storage, and inference deployment at the edge. At Q2BSTUDIO we offer just that: optimized cloud infrastructure combined with customized applications for biomedical signal processing.
SpO₂-guided time-frequency reconstruction also raises interesting questions about the interpretability of the models. How do we know that the reconstructor is not hallucinating waveforms that match the expected SpO₂ but not the physiological reality? This is where the need for rigorous clinical validation and the use of metrics not only of error, but of physiological plausibility come into play. Business intelligence tools, such as Power BI, can help visualize large volumes of reconstructed signals and correlate them with clinical events, allowing R+D teams to identify biases or anomalous patterns. At Q2BSTUDIO we develop customized dashboards that integrate these analyses, facilitating data-driven decision-making.
Another relevant aspect is cybersecurity. PPG signals contain unique biometric information (they can even be used for identification), and their transmission to the cloud or storage on devices must be protected from unauthorized access. Our cybersecurity services include end-to-end encryption, pentesting audits, and compliance with regulations such as HIPAA or GDPR, essential for any digital health platform. Combining advanced AI algorithms with secure infrastructure is the only way to bring these innovations to market without legal or reputational risks.
In addition, the staged approach described in the literature—predictor pretraining, reconstructor training with constraints, and fine-tuning—is a clear example of how AI agents can collaborate on complex tasks. Instead of a single monolithic model, multiple specialized modules are used and optimized together. This modular architecture is precisely what we apply in Q2BSTUDIO for process automation projects and custom software, where we break down complex problems into manageable sub-problems that are solved with independently trained models and then integrated using APIs or data flows.
From a practical point of view, the implementation of this system in a wearable device requires hardware with limited computing capacity. That's why techniques such as model quantization, neural network pruning, and the use of NPU accelerators are key. At Q2BSTUDIO we offer model optimization services for edge computing, ensuring that custom applications work with low latency even on microcontrollers. Our team has worked with healthcare customers to adapt PPG algorithms to platforms such as Arduino, ESP32, and ARM Cortex-based systems, reducing power consumption without sacrificing accuracy.
SpO₂-guided PPG reconstruction research also opens the door to new applications beyond oxygenation. For example, the same architecture could be used to predict blood pressure, respiratory rate, or even glucose levels from the same signals, provided that a suitable predictor is available. This expands the potential market for companies developing multifunction wearables. The key lies in the flexibility of the framework: by separating the predictor from the reconstructor, predictors for different physiological variables can be exchanged without the need to redesign the entire system. At Q2BSTUDIO we design modular software architectures that allow for that scalability, facilitating rapid iteration from prototypes to final product.
However, the adoption of these technologies in clinical settings requires overcoming regulatory barriers. AI algorithms used in medical devices must be validated under standards such as IEC 62304 or FDA, which involves thorough documentation, traceability, and version control. Our AI services for enterprises include advice on the entire model lifecycle: from labeled data collection to clinical validation and post-deployment maintenance. We collaborate with startups and labs to bring their innovations into compliance with regulatory requirements without slowing down development.
Finally, it is important to emphasize that time-frequency reconstruction is not a magic bullet. The quality of the input signal remains critical; No algorithm can retrieve information that never existed. Sensor design and placement is therefore still an active area of research. However, when you combine good hardware with smart software, the results are amazing. At Q2BSTUDIO we believe that the synergy between data engineering, AI and custom software development is the path to a future where health monitoring is continuous, accurate and accessible to everyone. If your company is exploring solutions based on PPG or any other type of biosensory signal, we have the experience to help you build from proof of concept to deployment in production, integrating cloud services, cybersecurity and business intelligence according to your needs.
