Beyond Turing
A next-generation wearable device - a wristband measuring heart rate, beat-to-beat variability (heart rate variability or HRV), oxygen saturation, electrodermal activity, peripheral temperature and sleep quality - produces, in twenty-four hours, on the order of seven to ten million biometric samples. Those samples, having traversed a pipeline of edge acquisition, adaptive filtering, feature extraction and machine-learning inference, are reduced to a decision: anomaly or normality, alert or silence, engagement or suspension. Between the raw sensor signal and the firing of the decision threshold lies an engineering architecture that, from 2 August 2026, operates under a new legal regime: EU Regulation 2024/1689 on artificial intelligence, commonly known as the AI Act, in full applicability for high-risk systems (Regulation EU 2024/1689).
The engineering question that follows is precise. Which sensors, which algorithms, which confidence thresholds qualify a wearable for consumer use as wellness, and which - by intended use, marketing modality or declared diagnostic capability - place it under the qualification of medical device pursuant to EU Regulation 2017/745 (MDR) and, concurrently, under the high-risk classification of Annex III of the AI Act? The answer is not one of principle but of design: it is the architecture of the system, not the manufacturer's declaration, that determines its legal qualification (Pesapane et al., 2018; Muehlematter et al., 2021).
This article traces the thread that binds the sensor to the threshold - from the principles of sensor fusion to classifier calibration metrics, from human-oversight requirements to post-market audit - and shows how the decision threshold of a medical wearable is not a hyperparameter to be optimised in the lab, but an architectural choice subject to multiple, convergent, and now concurrent normative constraints.
§1. The sensor as measurement instrument and ineliminable uncertainty
Clinical metrology teaches a rule that consumer culture tends to forget: no measurement is exact. Every biomedical sensor - photoplethysmographic (PPG), electrocardiographic (ECG), inertial accelerometric (IMU), optical for saturation (SpO₂), electrodermal (EDA), thermal - possesses a transfer function characterised by thermal noise, temporal drift, systematic error dependent on skin and contact pressure, sensitivity to motion artefact, limited bandwidth. The engineering literature on wearable technologies (Fabbrizio et al., 2023; Patel et al., 2012) has consolidated a metric characterisation for each channel: the typical accuracy of a wrist PPG sensor in heart-rate estimation under quiet conditions stands at a mean absolute error of ±2-3 beats per minute, rising to ±10-15 under high-intensity activity.
Measurement uncertainty is the starting datum, not a residual imperfection. From it descends the first architectural principle of any wearable pipeline: the decisional output cannot be more precise than the input. A neural network trained to classify arrhythmias from PPG signal with nominal accuracy of 99.5% on laboratory datasets encounters, in real-world deployment, a documented performance degradation (Bayoumy et al., 2021) that can reach 15-25% in populations and contexts not represented in the training dataset. The difference between the value published in the paper and the value observed in deployment is what the clinical-validation literature calls performance gap.
The typical engineering pipeline articulates, from upstream to downstream, in canonical stages: sensing multimodal (one or more synchronised channels), pre-processing (artefact removal, band-pass filtering, baseline correction), feature extraction (time-frequency parameters, non-linearity indices, waveform descriptors), multimodal fusion (early, late or hybrid), inference (classification or regression model), threshold (calibration of the operational cut-off), action (user notification, clinical alert, integration into electronic health record). Each stage possesses traceable quality metrics and, under the AI Act, an obligation of registration (logging, art. 12) and technical documentation (Annex IV).
§2. Sensor fusion and the geometry of decision
A single channel is almost always insufficient. Mature engineering practice has shown that the decisional robustness of a medical wearable depends on the ability to integrate heterogeneous channels into a coherent representation - sensor fusion (Hall & Llinas, 1997; Khaleghi et al., 2013). The consolidated taxonomy distinguishes three levels: data-level fusion (concatenation of temporally aligned raw samples), feature-level fusion (concatenation of descriptors extracted independently), decision-level fusion (combination of outputs from specialised classifiers, typically via voting, stacking or Dempster-Shafer evidence theory). Each level implies a different trade-off between informational richness, computational cost, latency, graceful degradation under sensor failure.
The reference architectural approach for contemporary medical wearables is probabilistic late fusion (Fucarino et al., 2024): each channel feeds a specialised classifier that produces a calibrated probability distribution, and a Bayesian or evidence-theoretic combination function produces the final decision. The design logic is clear: late fusion isolates failure. If the PPG sensor loses contact, the associated classifier emits a low-confidence signal that the combination function takes in, reducing the weight of the degraded channel. The system continues to operate but communicates with honesty - to the downstream model and to the clinical operator - that the decision formed on partial evidence.
The threshold is the point at which the continuous probabilistic representation transforms into discrete action. Its choice is geometrically equivalent to drawing a hyperplane in the feature space, and operationally equivalent to defining the trade-off between false positives and false negatives. For a cardiovascular wearable, a too-permissive threshold generates clinical alert fatigue - the operator learns to ignore notifications; a too-restrictive threshold misses critical clinical events. Correct calibration is not a zero-cost optimisation exercise: it requires ROC curves stratified by demographic subgroup, active post-deployment calibration (Platt scaling, isotonic regression), and - under the AI Act - traceable registration of threshold variations with their clinical rationale (art. 9, risk management system).
§3. Regulatory qualification: medical device and risk classification
A wearable is a medical device, pursuant to art. 2(1) of EU Regulation 2017/745 (MDR), when intended by the manufacturer for a specific medical purpose - diagnosis, prevention, monitoring, prognosis, treatment, alleviation of disease or injury. Qualification does not depend on technology but on the declared intended use and the representation the manufacturer gives of it on the market (Pesapane et al., 2018). A materially identical device - same sensor, same algorithms, same firmware - may be marketed as a wellness product (outside the MDR perimeter) or as a medical device (inside the perimeter). The difference is an editorial decision by the manufacturer, but the consequences are engineering: the medical device requires a quality management system compliant with ISO 13485, documented clinical evaluation, a post-market surveillance system, a technical dossier approvable by a notified body.
From 2 August 2026, the AI Act overlays a second qualification. Art. 6(1) and Annex I automatically qualify as high-risk AI systems that are safety components of medical devices already subject to third-party conformity assessment. The dual qualification - MDR plus AI Act high-risk - is not a duplicate but a stratification: the device must satisfy medical-device requirements (safety, performance, clinical evaluation) and AI-specific requirements (data quality, robustness, accuracy, human oversight, transparency, logging). Recital 51 of the AI Act clarifies that high-risk classification under the Regulation does not automatically entail elevated-risk classification under MDR, but - as the sector literature has highlighted (Pesapane, Iuliano et al., 2025) - the two compliance architectures overlap substantially in engineering constraints.
The AI Act requirements for high-risk systems, summarised in articles 8-15, constitute a technical design specification: risk management system iterative across the entire lifecycle (art. 9), data governance with training, validation and test datasets characterised for representativeness, completeness and absence of errors (art. 10), technical documentation compliant with Annex IV (art. 11), automatic event logging (art. 12), transparency towards the user (art. 13), human oversight designed into the system (art. 14), measurable accuracy, robustness and cybersecurity (art. 15). The ISO/IEC 42001 standard on AI management systems, published in December 2023, provides the organisational model to fulfil these requirements verifiably, in parallel with ISO 14971 on medical-device risk management.
§4. Human oversight and the problem of clinical cognitive load
Art. 14 of the AI Act qualifies human oversight as a design requirement, not as a statement of intent. The system must be designed so that the natural persons to whom supervision is entrusted can understand the system's capabilities and limits, monitor its operation, correctly interpret its output, decide not to use it, intervene or interrupt its operation. This is a functional specification translatable into interfaces, admissible latencies, informational content of the notification, override pathways.
The operational problem that the clinical literature has identified with growing precision is the cognitive load of the operator. A general practitioner who receives, over the course of a week, several hundred alerts from patients' wearables - rhythm anomalies, pressure deviations, nocturnal episodes - does not possess the time necessary for substantive evaluation of each. The phenomenon of alert fatigue, documented in over twenty years of literature on clinical decision support systems (van der Sijs et al., 2006; Ancker et al., 2017), produces a counter-intuitive effect: human oversight formally existing becomes, under load, a procedural formality. The operator confirms by default, barring macroscopic evidence of error. Automation bias (Parasuraman & Manzey, 2010) closes the loop: uncritical trust in the system reinforces with volume, it does not erode.
Engineering design responds with specific tools, today part of the state of the art of clinical decision-support systems. Algorithmic alert triage clusters correlated events and presents them in groups rather than as isolated notifications. Confidence grading visually separates high-probability true-positive alerts from low-probability ones, activating differentiated workflows. The closed feedback loop between documented clinical outcome and threshold calibration, sustained by regulated continuous learning architectures (FDA Good Machine Learning Practice, 2021; FDA Predetermined Change Control Plans, 2024), allows the system to learn from its errors without violating the static-conformity framework. The experience matured on tele-exercise platforms (Iuliano, Zimatore, Fabbrizio et al., 2025) has shown that the clinical effectiveness of a wearable grows not with sensor sensitivity but with the adequacy of the feedback loop among device, patient and clinician.
Human oversight, in this reading, is not a periodic veto over the autonomous system: it is a system function that must be measurable. Average review time per alert, observed override rate, distribution of override reasons, correlation between override and downstream clinical outcomes: these are the operational descriptors of effective human oversight. Their absence in the technical documentation transforms art. 14 into an empty declaration - and the AI Act does not tolerate empty declarations.
§5. The threshold as design invariant
Return to the initial question. The wearable collects millions of samples; the pipeline distils a decision; the threshold transforms continuous probability into discrete action. Where, in this chain, does the principle that the device serves the person - and not the reverse - reside?
The mature engineering answer, today sustained by the convergence between AI Act, MDR, ISO/IEC 42001 and FDA guidelines, is that the principle resides at no specific point in the pipeline - it resides as a design invariant the system must satisfy in every state. The threshold is not an optimisation opportunity: it is the point at which the system recognises that it operates on a subject who structurally exceeds its statistical representation. The training distribution is finite, real-population heterogeneity is infinite, the individual case is - to a never-null degree - out-of-distribution. Uncertainty quantification techniques (Bayesian calibration, conformal prediction, ensemble disagreement) do not eliminate this excess: they declare it, quantify it, transform it into operational information that human oversight can actually use.
The methodological point may be stated cleanly: a medical wearable system that decides upon a subject must know when it does not know, and stop. The system's self-suspension is not a project failure - it is a functional specification of the project, encoded in fallback protocols towards mandatory human deliberation, signalled to the operator with honest language, registered in the logs of art. 12 of the AI Act. A device that does not know how to stop, in strict engineering sense, is a badly designed device - and badly designed, today, means non-compliant.
Between the sensor and the threshold lie millions of samples, dozens of processing stages, a confusion matrix, technical documentation of hundreds of pages, a notified body, an MDR file and an AI Act dossier. But at the bottom - at the bottom the pipeline does not see - there is a subject. The engineering maturity of a wearable health system is not measured in sensor accuracy: it is measured in its ability not to forget that subject at any point of its processing chain, not even when the pipeline produces technically impeccable output.
The sensor measures. The threshold decides. But between the measurement and the decision, between the sample and the action, between the physiological datum and the clinical notification, there lives a distance that no architecture can close - and that every well-made architecture acknowledges. The most important line of code, even in a wearable, is the one that suspends execution and asks a human being to look.