ISO TS 24971-2:2026 Compliance for OEMs: Managing ML Model Drift and Bias in Smart Wearables via Continuous PMS

The integration of machine learning into medical wearables has transitioned from a futuristic concept to a clinical reality. Today, smartwatches, smart rings, and continuous blood pressure monitors rely on complex algorithms to filter motion artifacts, estimate physiological parameters, and detect anomalies. However, the non-deterministic nature of machine learning (ML) introduces a completely new category of risks that traditional software validation frameworks simply cannot address. When an algorithm’s performance degrades silently over time, or when it exhibits demographic bias in real-world populations, the consequences extend far beyond poor user experience—they cross into patient safety and regulatory non-compliance.

For B2B buyers, OEM brand owners, and medical device distributors, understanding the shifting regulatory landscape is no longer optional. The international community has responded to these challenges with the publication of ISO TS 24971-2:2026. This technical specification provides the definitive guidance for applying the ISO 14971 risk management process specifically to machine learning-enabled medical devices (MLMD). For Original Equipment Manufacturers building the next generation of connected health platforms, mastering this standard is the key to avoiding costly recalls, regulatory setbacks, and market exclusion. This comprehensive guide breaks down the requirements of ISO TS 24971-2:2026, explores the mechanics of model drift and bias in wearables, and outlines how continuous Post-Market Surveillance (PMS) serves as the engine for ongoing compliance.

The Regulatory Evolution: From ISO 14971:2019 to ISO TS 24971-2:2026

ISO 14971:2019 remains the internationally recognized framework for medical device risk management. It provides a systematic approach to identify, assess, control, and monitor risks across the total product lifecycle. For traditional deterministic software—where a specific input always yields the exact same output—ISO 14971:2019 works exceptionally well. Developers can map every potential failure mode, test the software against those modes, and lock the code. However, machine learning models do not operate on fixed rules; they operate on probabilistic patterns derived from training data. As the real-world environment changes, the model’s outputs can change unpredictably.

Recognizing this gap, regulatory bodies and standards organizations initiated a multi-year effort to supplement existing frameworks. The culmination of this effort is PD CEN ISO/TS 24971-2:2026, published on June 30, 2026, with the IEC releasing ISO TS 24971-2:2026 as a Technical Specification on June 17, 2026. This document explicitly provides guidance on applying the ISO 14971 risk management process to MLMD, addressing the unique hazards introduced by artificial intelligence in healthcare.

A critical distinction in ISO TS 24971-2:2026 is its scope regarding AI types. The standard explicitly excludes machine learning-enabled medical devices employing large language models (LLM) or generative AI. This exclusion is highly relevant for OEMs focusing on physiological monitoring. While an LLM might be used in a patient-facing chatbot for symptom triage, the core clinical functions of a smartwatch ECG device, a heart rate chest strap, or a blood pressure monitor rely on discriminative ML models (like convolutional neural networks for signal processing or regression models for blood pressure estimation). ISO TS 24971-2:2026 is written precisely for these discriminative, signal-processing ML models embedded in hardware.

The FDA’s Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device Action Plan, initially released in 2021 and updated with discussions on Predetermined Change Control Plans in 2023, laid the groundwork for this international standard. The FDA emphasized that traditional ‘locked’ algorithms are insufficient for devices that learn and adapt. ISO TS 24971-2:2026 translates these regulatory expectations into a globally harmonized technical specification. For OEMs, this means that risk management is no longer a one-time pre-market activity; it is a continuous, data-driven process that spans the entire operational life of the wearable device.

Identifying Machine Learning-Specific Risks in Smart Health Wearables

To manage risk, engineers and quality teams must first accurately identify it. ISO TS 24971-2:2026 categorizes ML-specific risks into several distinct domains: training data quality, algorithmic bias, explainability, transparency, autonomy, cybersecurity, model drift, and continuous learning. In the context of smart wearables, these abstract concepts translate into very specific hardware and software failure modes.

Consider the optical heart rate sensors found in smart bands and smart rings. These devices use photoplethysmography (PPG) to measure blood volume changes. An ML model is typically deployed to filter out motion artifacts and ambient light interference. If the training data used to develop this model primarily consisted of recordings from young, sedentary individuals in controlled laboratory settings, the model will likely fail when deployed on elderly patients with arterial stiffness or construction workers experiencing high-vibration environments. This is a training data risk that directly impacts clinical accuracy.

Algorithmic bias represents another severe risk category. A 2022 study published in the Journal of the American Heart Association highlighted significant accuracy disparities in optical wearable sensors across different Fitzpatrick skin types. If an OEM trains a blood oxygen saturation (SpO2) algorithm predominantly on lighter skin tones, the device may consistently overestimate or underestimate readings for users with darker skin. Under ISO TS 24971-2:2026, this demographic bias is not just a performance flaw; it is a documented safety risk that must be mitigated and controlled.

The following table contrasts traditional software risks with the new ML-specific risks outlined in the 2026 technical specification, illustrating the paradigm shift required for wearable development:

Identifying these risks requires a cross-functional approach. Hardware engineers must understand how sensor degradation affects data inputs, while software engineers must evaluate how the ML model handles edge cases. Quality and regulatory teams must then document these hazards in the risk management file, applying the specific mitigation strategies demanded by the new technical specification.

Tackling Model Drift and Algorithmic Bias in Continuous Monitoring Devices

Model drift is arguably the most insidious risk in ML-enabled wearables. Unlike a mechanical component that breaks suddenly, an ML model can drift slowly, producing increasingly inaccurate results while still outputting a plausible-looking number. In the context of ISO 14971 and ISO TS 24971-2:2026, drift is categorized into two main types: data drift and concept drift.

Data drift occurs when the statistical properties of the input data change over time. For a smart blood pressure monitor, this could mean the physical optical sensors degrade due to micro-scratches on the lens, or the LED intensity diminishes as the battery chemistry ages. The raw signal entering the ML model is now fundamentally different from the pristine data it was trained on. Concept drift happens when the underlying relationship between the input data and the target output changes. For example, a heart rate algorithm trained on a general population might experience concept drift if deployed exclusively in a high-altitude corporate wellness program, where physiological baselines shift due to lower oxygen levels.

Managing these risks requires proactive engineering and rigorous data governance. OEMs must implement techniques such as continuous model monitoring, periodic retraining, and fallback mechanisms. If the model’s confidence score drops below a certain threshold, the device should either alert the user to recalibrate or revert to a simpler, deterministic algorithm.

Mini Case: Overcoming Demographic Bias in Smart Ring Manufacturing

A European corporate health management client approached Xdun Medical Technology to develop a customized smart ring for continuous employee wellness tracking. Initial field tests revealed a 12% accuracy drop in heart rate variability (HRV) readings for employees with darker skin tones, indicating severe algorithmic bias in the baseline PPG model. Leveraging our in-house R&D capabilities, the Xdun engineering team collaborated with the client to collect a new, highly diverse training dataset across multiple Fitzpatrick skin types. We retrained the edge-computing algorithm and optimized the hardware LED wavelengths. The updated model eliminated the demographic bias, improving overall accuracy by 14% and ensuring the device met the stringent fairness requirements expected by European health regulators.

This case highlights a fundamental principle of ISO TS 24971-2:2026: risk management for ML is not just about writing good code; it is about curating representative data and aligning hardware capabilities with algorithmic requirements. OEMs must ensure their manufacturing partners possess the technical depth to iterate on both the physical sensors and the ML models simultaneously.

Implementing Continuous Post-Market Surveillance (PMS) for ML-Enabled Devices

Historically, Post-Market Surveillance (PMS) in the medical device industry relied on periodic reviews, complaint handling, and literature searches. Manufacturers would compile an annual PMS report, analyze adverse events, and update their risk management files accordingly. This reactive, batch-processing approach is fundamentally incompatible with machine learning. An ML model can drift or exhibit bias long before a formal complaint is filed, especially if the errors are subtle or if users simply assume the device is inaccurate and stop wearing it.

ISO TS 24971-2:2026 mandates a shift toward continuous, proactive PMS. For connected health platforms and smart wearables, this means establishing a live telemetry pipeline that feeds real-world performance data back to the manufacturer. The EU Medical Device Regulation (MDR), under Article 83, already demands robust PMS systems, but the application of these rules to AI requires a much higher degree of automation and data granularity.

To comply with the continuous PMS requirements for ML wearables, OEMs must establish a system that incorporates the following core components:

• Real-World Performance Telemetry: The device must periodically transmit anonymized performance metrics to a secure cloud platform. This includes signal-to-noise ratios, sensor contact quality, ambient light interference levels, and the model’s internal confidence scores.
• Automated Drift Detection Algorithms: The cloud platform must run secondary, lightweight algorithms to monitor the primary ML model’s output distribution. If the statistical distribution of the device’s heart rate or blood pressure readings shifts significantly compared to the baseline training distribution, the system triggers an alert for engineering review.
• Demographic Stratification: PMS data must be segmented by user demographics (age, sex, skin tone, comorbidities). This allows manufacturers to detect algorithmic bias in the field that was not apparent during pre-market clinical trials.
• Feedback Loops for Retraining: The continuous PMS system must feed high-quality, validated real-world data back into the data pipeline, enabling the periodic retraining of the ML model to correct for concept and data drift.
• Cybersecurity and Data Privacy Integration: Continuous data collection introduces significant privacy risks. The PMS architecture must comply with IEC 81001-5-1 for health software cybersecurity, ensuring that all telemetry is encrypted, anonymized, and transmitted securely without exposing Protected Health Information (PHI).

Implementing this level of continuous PMS requires a robust cloud infrastructure and deep integration between the wearable hardware, the mobile application, and the backend servers. OEMs cannot treat the cloud platform as an afterthought; it must be designed as a core component of the medical device’s risk control strategy.

Engineering Strategies for OEMs: Building Compliant AI Wearables from the Ground Up

Compliance with ISO TS 24971-2:2026 cannot be bolted onto a finished product at the end of the development cycle. It must be engineered into the device from the initial prototyping phase. This requires a fundamental shift in how OEMs approach hardware-software co-design for medical wearables.

The first strategy is designing for explainability and transparency. While deep learning models are inherently opaque, engineers can implement techniques to make their behavior more transparent. This includes using SHAP (SHapley Additive exPlanations) values to understand which input features (e.g., specific frequency bands in a PPG signal) are driving the model’s predictions. When a device outputs an anomalous ECG reading, the system should be able to flag the specific signal artifact that triggered the alert, providing transparency for the reviewing clinician.

The second strategy involves rigorous data governance and hardware optimization. The quality of an ML model is strictly bounded by the quality of the data it receives. OEMs must specify and source medical-grade sensors and components that provide high-fidelity, consistent raw data. If the physical sensor produces noisy, unreliable data, even the most advanced ML algorithm will fail. Furthermore, the mechanical design of the wearable—such as the fit of a smart ring or the strap tension of a heart rate chest strap—directly impacts sensor contact and, consequently, data quality. Risk management must encompass the physical ergonomics of the device.

The third strategy is leveraging edge computing to manage autonomy and cybersecurity risks. Processing ML models directly on the wearable device (edge computing) rather than in the cloud reduces latency, preserves battery life, and minimizes the attack surface for cybersecurity threats. However, edge devices have limited computational power. OEMs must employ model quantization and pruning techniques to reduce the model’s size without sacrificing clinical accuracy.

Executing these engineering strategies requires a manufacturing partner with deep technical expertise. At Xdun Medical Technology, our team of over 100 dedicated engineers specializes in the end-to-end development of complex medical wearables. From initial R&D design and prototyping to mass production, we ensure that hardware and software are perfectly aligned to meet stringent clinical requirements. Our devices are built with medical-grade sensors and components, providing the high-fidelity data foundation necessary for advanced ML algorithms. You can explore our comprehensive capabilities and view our full range of medical wearables, including our advanced smartwatch and ECG device lines, to see how we integrate hardware precision with software intelligence.

To support these engineering efforts, OEMs should implement the following development best practices:

• Establish a Predetermined Change Control Plan (PCCP): Define exactly how the ML model will be updated over-the-air. Specify the triggers for retraining, the validation protocols required before deployment, and the rollback mechanisms if a new model version underperforms.
• Implement Hardware-in-the-Loop (HITL) Testing: Do not rely solely on software simulations. Test the ML models using physical hardware prototypes subjected to real-world environmental stressors, including temperature extremes, humidity, and physical vibration.
• Design for Fallback Modes: If the ML model’s confidence drops below a safe threshold, the device must seamlessly transition to a deterministic, rule-based algorithm or prompt the user to seek alternative measurement methods.
• Maintain a Comprehensive Model Card: Document the model’s intended use, training data demographics, known limitations, and performance metrics across different subgroups. This document serves as a critical reference for both internal engineering teams and external regulatory auditors.

Navigating the Certification Journey with an Experienced Manufacturing Partner

Developing a technically sound, ML-enabled wearable is only half the battle. The other half is proving compliance to global regulatory bodies. The documentation required for CE marking under the EU MDR, FDA 510(k) or De Novo clearance in the United States, and other international markets is immense. For MLMDs, regulators expect to see a complete traceability matrix from the initial hazard analysis to the final validation report, including detailed explanations of the training data and the continuous PMS strategy.

This is where the choice of an OEM/ODM partner becomes a critical business decision. Many manufacturers claim to offer ‘turnkey’ solutions, but lack the specific regulatory engineering experience required for AI-enabled medical devices. A partner who does not understand the nuances of ISO TS 24971-2:2026 will produce documentation that fails regulatory scrutiny, leading to costly delays and redesigns.

It is important to clarify how a manufacturing partner fits into the certification ecosystem. Xdun Medical Technology brings over 15 years of specialized experience in the medical device manufacturing sector. We understand the rigorous demands of global regulatory frameworks. However, we operate strictly within the bounds of our expertise as a manufacturer. We do not claim that our standard products or facilities hold specific medical device certifications. Instead, our value proposition lies in our ability to build the foundation for your success. We support and assist customers through their own certification processes, including CE, FCC, FDA, RoHS, BSCI, and ISO 13485. We provide the comprehensive technical documentation, design history files (DHF), and manufacturing traceability records that our clients need to successfully navigate their specific certification journeys.

By partnering with an experienced OEM that understands the intersection of hardware engineering, ML software development, and regulatory documentation, brand owners can significantly de-risk their path to market. We ensure that the physical devices are manufactured to the highest quality standards, utilizing medical-grade components, and that the accompanying technical files are structured to meet the exacting requirements of international notified bodies and regulatory agencies.

Conclusion: Proactive Risk Management as a Competitive Advantage

The publication of ISO TS 24971-2:2026 marks a definitive milestone in the regulation of artificial intelligence in healthcare. It acknowledges that machine learning is not just another software feature; it is a fundamental shift in how medical devices operate, learn, and potentially fail. For OEMs, brand owners, and distributors of smart wearables, ignoring the specific risks of model drift, algorithmic bias, and the need for continuous PMS is no longer an option. The regulatory scrutiny on AI-enabled medical devices will only intensify in the coming years.

However, compliance should not be viewed merely as a regulatory burden. A robust risk management framework, aligned with ISO 14971 and ISO TS 24971-2:2026, is a powerful competitive advantage. It results in safer, more reliable products that clinicians trust and patients rely on. It minimizes the risk of costly post-market recalls and protects brand reputation. By embedding risk management into the core engineering process, leveraging continuous PMS, and partnering with experienced manufacturing experts, OEMs can confidently bring innovative, AI-powered wearables to the global market.

Partner with Xdun Medical Technology for Your Next AI Wearable Project

Building compliant, high-performance medical wearables requires a partner who understands both the cutting-edge technology and the rigorous regulatory landscape. With over 15 years of experience, a dedicated team of 100+ engineers, and end-to-end capabilities from R&D prototyping to mass production, Xdun Medical Technology is equipped to support your most complex OEM/ODM projects. We build devices with medical-grade sensors and components, and we actively support and assist customers through their own certification processes to ensure a smooth path to market.

Ready to discuss your next smart wearable, smart ring, or continuous monitoring device project? Contact our B2B engineering and sales team today to explore how we can bring your vision to life while maintaining strict adherence to global risk management standards.

Email: jine@xdunmedical.com
Phone / WhatsApp: +86-13925587509
Explore Our Capabilities: Visit the Xdun Medical Product Center

In projects where procurement teams need to validate device capabilities early, a a screenless monitoring band like the E910 that streams continuous vitals data for ML drift analysis can serve as a practical reference unit during pilot evaluations. OEMs often leverage such platforms to demonstrate sensor accuracy, connectivity reliability, and form-factor ergonomics before committing to full production runs. For specifications and technical details, explore the screenless monitoring band platform.

Related resources for deeper reading:

• PMS integration into ISO 14971 risk management plans
• quality agreement checklist covering PMS data obligations