Sensing with Understanding: AI for Wearable Health

by

Author: Denis Avetisyan

New research explores how to build artificial intelligence systems that not only predict health metrics from wearable sensors, but also clearly explain why they made those predictions.

Time series decomposition enables the creation of an extended model [latex]M^\hat{M}[/latex] by transforming an original time series [latex]\mathbf{x}[/latex] into component representations [latex]{\mathbf{C}\_{\mathbf{x}}}[/latex] via a forward function FF, and then leveraging the inverse transformation [latex]F^{-1}(\cdot)[/latex] in combination with the original time series network MM, all without requiring model retraining.
Time series decomposition enables the creation of an extended model [latex]M^\hat{M}[/latex] by transforming an original time series [latex]\mathbf{x}[/latex] into component representations [latex]{\mathbf{C}\_{\mathbf{x}}}[/latex] via a forward function FF, and then leveraging the inverse transformation [latex]F^{-1}(\cdot)[/latex] in combination with the original time series network MM, all without requiring model retraining.

This review details a novel approach using Inherently Interpretable Components to enhance both the accuracy and explainability of AI models for time-series health data.

Despite the increasing reliance on artificial intelligence for real-time health monitoring via wearable sensors, a critical gap remains between predictive accuracy and model interpretability. This paper, ‘Explainable AI Using Inherently Interpretable Components for Wearable-based Health Monitoring’, addresses this challenge by introducing a novel Explainable AI (XAI) method leveraging Inherently Interpretable Components (IICs) to provide both high-performing predictions and understandable explanations for time-series data. By encapsulating domain-specific concepts within a custom explanation space, IICs preserve model performance while enabling concept-based explanations for applications like state assessment and epileptic seizure detection. Could this approach unlock more trustworthy and effective AI-driven health solutions by bridging the gap between prediction and understanding?

Unveiling the Body’s Subtle Language

The ability to meticulously track a person’s physiological condition, on a continuous basis, represents a paradigm shift in healthcare’s potential for preventative action and swift response. Rather than reacting to acute episodes, consistent monitoring allows for the detection of subtle, early indicators of distress or imbalance – potentially weeks or even months before symptoms manifest clinically. This proactive approach enables interventions tailored to individual needs, fostering better management of chronic conditions, personalized wellness strategies, and a reduced reliance on reactive, emergency-based care. By establishing a baseline understanding of an individual’s normal physiological range, deviations – even minor ones – can be flagged, prompting timely assessment and preventing escalation into more serious health events. Ultimately, continuous physiological monitoring empowers a move from treating illness to maintaining wellness, fostering a more sustainable and effective healthcare system.

Historically, evaluating a person’s physiological condition has depended on sporadic measurements taken during clinical visits, creating a fragmented picture of overall health. These infrequent snapshots often fail to capture the dynamic fluctuations inherent in biological systems, limiting the ability to detect subtle changes that may signal emerging health concerns. Consequently, critical insights into an individual’s baseline state, responses to daily stressors, or the progression of chronic conditions are frequently missed. This reliance on isolated assessments hinders proactive healthcare, as timely interventions are often delayed until symptoms become pronounced, rather than being guided by continuous, real-time data reflecting the body’s ongoing state.

The advent of wearable sensor technology, exemplified by devices like the Empatica E4, represents a significant leap forward in physiological monitoring capabilities. These compact systems facilitate the uninterrupted gathering of critical biometric data – moving beyond the limitations of infrequent clinical assessments. By continuously tracking metrics such as heart rate variability, electrodermal activity, movement, and skin temperature, a detailed and nuanced profile of an individual’s physiological state can be constructed. This constant stream of information allows for the detection of subtle shifts and patterns indicative of changing health conditions or emotional responses, offering the potential for early intervention and personalized healthcare strategies. The ability to move monitoring from controlled laboratory settings into everyday life unlocks a wealth of data previously inaccessible, paving the way for a more proactive and preventative approach to well-being.

The confluence of physiological signals gathered from wearable sensors offers an unprecedented window into a person’s holistic state. Heart Rate Variability (HRV), a measure of the variation in time between heartbeats, reflects the interplay between the sympathetic and parasympathetic nervous systems, indicating stress, recovery, and overall cardiovascular health. Electrodermal Activity (EDA), often measured through skin conductance, reveals changes in sweat gland activity – a sensitive marker of emotional arousal and cognitive load. Simultaneously tracked acceleration data provides insights into physical activity levels and sleep patterns, while skin temperature fluctuations can signal changes in metabolic rate or even early indications of illness. By integrating these diverse data streams, researchers and clinicians gain a nuanced understanding of an individual’s physical, emotional, and cognitive condition, far exceeding the limitations of isolated measurements.

Decoding Seizure Dynamics Through Continuous Monitoring

Seizure detection represents a complex challenge within physiological monitoring due to the variability of seizure manifestations and the potential for false positives arising from similar physiological phenomena, such as movement artifacts or muscle tremors. Despite these difficulties, accurate and timely seizure detection is critically important for patient safety, particularly for individuals with epilepsy who may experience recurrent, unpredictable events. Delays in detection can lead to physical injury from uncontrolled movements, aspiration, or falls, and can also contribute to status epilepticus, a life-threatening condition requiring immediate medical intervention. Consequently, ongoing research focuses on developing robust and reliable monitoring systems to improve the quality of life and reduce morbidity associated with seizure disorders.

Early and accurate detection of seizures, specifically generalized Tonic-Clonic Seizures, is critical for mitigating potential harm to patients. These events are characterized by loss of consciousness and involuntary muscle contractions, presenting risks of physical injury from falls or collisions. Timely intervention, facilitated by accurate detection systems, allows for the implementation of protective measures, such as positioning the patient to prevent trauma, and the administration of rescue medications to shorten seizure duration and reduce post-ictal complications. Reduced seizure duration correlates directly with decreased risk of both immediate physical harm and long-term neurological consequences, making rapid detection a primary objective in patient monitoring.

Acceleration data provides a key indicator of seizure events due to the involuntary, often violent, muscular contractions characteristic of many seizure types. Analysis focuses on detecting changes in movement, including the onset, duration, and intensity of accelerations across multiple axes. Specifically, tonic-clonic seizures present with pronounced, repetitive acceleration patterns resulting from the alternating muscle rigidity and relaxation. Algorithms utilize features extracted from this data – such as signal magnitude area, variance, and frequency-domain characteristics – to differentiate seizure-related movements from normal physiological activity or environmental noise. The sensitivity of accelerometers, coupled with their non-invasive nature, makes them a practical and effective component of seizure detection systems, particularly when integrated with other physiological measurements.

Integration of acceleration data with electrodermal activity (EDA) and heart rate variability (HRV) measurements demonstrably enhances the performance of seizure detection algorithms. This multi-modal approach leverages the complementary information provided by each signal; acceleration captures the motor manifestations of seizures, while EDA and HRV reflect autonomic nervous system changes correlated with seizure activity. Evaluations have shown this combined analysis achieves an accuracy of 87.8% in identifying seizure events, representing a substantial improvement over systems relying on single-modality inputs. The increased reliability is due to the algorithm’s ability to discriminate between seizure-related movements and other artifacts or physiological noise.

The IIC approach effectively distinguishes between seizure and non-seizure events by revealing divergent distributions in scalar summary components.
The IIC approach effectively distinguishes between seizure and non-seizure events by revealing divergent distributions in scalar summary components.

Illuminating the ‘Why’ Behind Physiological Predictions

Conventional machine learning models, despite often achieving high predictive accuracy with physiological data, typically function as ‘black boxes’ – their internal decision-making processes remain opaque. This lack of transparency hinders understanding of the physiological mechanisms that contribute to a given prediction. While a model might accurately classify a patient’s condition, it provides no information regarding why that classification was made, nor which specific physiological features were most influential. Consequently, clinicians are unable to validate the model’s reasoning, potentially limiting trust and hindering the translation of model outputs into actionable clinical insights. This is a significant limitation in medical applications where understanding the ‘how’ and ‘why’ is as important as the prediction itself.

Explainable AI (XAI) techniques, including Feature-SHAP and Saliency-Based XAI, are crucial for deciphering the factors driving predictions made by machine learning models applied to physiological data. Feature-SHAP utilizes Shapley values from game theory to quantify each feature’s contribution to the prediction, providing a consistent and locally accurate explanation. Saliency-Based XAI, conversely, highlights the input features that most strongly influence the model’s output, typically visualized as a heatmap overlaid on the input signal. Both methods allow researchers to move beyond simply knowing that a model made a prediction, and instead understand why, enabling validation of model logic against established physiological principles and identification of potentially spurious correlations.

Time Series Decomposition is a pre-processing technique used to dissect a physiological signal into its constituent components – typically trend, seasonality, and residuals. Applying methods such as Seasonal-Trend decomposition using Loess (STL) or classical decomposition allows for the extraction of features beyond those apparent in the raw signal. For example, the amplitude of the seasonal component can indicate cyclical patterns in heart rate variability, while the trend component can reveal long-term changes in blood pressure. The residual component, representing noise or irregular variations, can be analyzed for anomalies. These decomposed components and their statistical properties serve as valuable inputs for machine learning models, improving prediction accuracy and facilitating the identification of clinically relevant physiological changes.

Concept-Based Explainable AI (XAI) enhances interpretability of physiological signal analysis by translating model predictions into clinically relevant physiological concepts, rather than relying on raw data feature importance. This approach moves beyond identifying which data points influenced a decision to explaining why a prediction was made in terms understandable to a medical professional. Specifically, this methodology has demonstrated state-of-the-art classification accuracy, achieving 99.0% on a state assessment task, while simultaneously providing explanations grounded in physiological understanding. This allows for verification of the model’s reasoning and increases trust in its predictions for clinical applications.

Independent Interconnected Component Signals (IICS) effectively decompose a temperature signal into its constituent components.
Independent Interconnected Component Signals (IICS) effectively decompose a temperature signal into its constituent components.

Towards a Future of Proactive and Personalized Wellbeing

The convergence of continuous physiological monitoring and Explainable AI (XAI) represents a paradigm shift in healthcare’s ability to identify health deterioration before symptoms manifest. By constantly tracking vital signs and biometric data – far beyond the scope of periodic check-ups – these systems establish a baseline of individual physiological norms. Subtle deviations from this baseline, often imperceptible to the individual or detectable only in retrospect, can then be flagged by sophisticated algorithms. However, unlike ‘black box’ AI models, XAI provides clinicians – and potentially patients – with clear, understandable rationales for these alerts, detailing which specific data points triggered the concern and why. This transparency is crucial for building trust and enabling timely interventions, moving healthcare from a reactive model of treating illness to a proactive one focused on prevention and early management, ultimately improving patient outcomes and reducing healthcare costs.

The future of healthcare increasingly centers on interventions uniquely designed for each individual, moving beyond standardized treatments. By analyzing a person’s continuous physiological data – encompassing metrics like heart rate variability, sleep patterns, and activity levels – alongside their specific risk factors, clinicians can develop highly personalized strategies. These interventions aren’t simply about addressing existing illness, but proactively managing health trajectories; for example, adjusting medication dosages based on real-time biomarker responses, or recommending tailored exercise regimens to mitigate genetically predisposed vulnerabilities. This approach recognizes that individuals respond differently to stimuli and that a one-size-fits-all model often falls short, paving the way for more effective and preventative care that optimizes wellbeing based on a complete, individual profile.

The integration of Explainable AI (XAI) into healthcare isn’t simply about improved algorithms; it fundamentally reshapes the relationship between medical systems and those who utilize them. Traditional ‘black box’ AI models, while potentially accurate, offer little insight into how a diagnosis or treatment plan is reached, hindering clinician acceptance and patient understanding. XAI, conversely, prioritizes transparency, revealing the factors driving its conclusions – whether it’s identifying subtle patterns in physiological data or highlighting crucial risk indicators. This allows clinicians to validate AI-driven recommendations, strengthening their confidence and facilitating informed decision-making. Simultaneously, patients gain a clearer understanding of their health status and proposed interventions, fostering trust and encouraging active participation in their care. By demystifying complex processes, XAI moves beyond simply predicting outcomes to empowering both medical professionals and individuals with the knowledge needed to navigate healthcare effectively.

A paradigm shift in healthcare is occurring, moving beyond simply responding to illness to actively anticipating and mitigating health risks. Recent advancements demonstrate that a holistic strategy – integrating continuous physiological monitoring with explainable artificial intelligence – significantly outperforms traditional reactive methods. Studies utilizing multiple datasets reveal substantial accuracy improvements in identifying subtle indicators of health decline before symptoms manifest. This proactive approach enables the tailoring of personalized interventions, addressing individual vulnerabilities and optimizing preventative care. The result is not merely earlier diagnosis, but a fundamental change in how health is managed – fostering resilience and promoting sustained well-being through informed, preemptive action.

The pursuit of explainability in artificial intelligence, as demonstrated by this research into Inherently Interpretable Components, echoes a fundamental principle of understanding any complex system: discerning its underlying structure. Each time-series data point from wearable sensors, when viewed through the lens of IICs, reveals dependencies that would otherwise remain obscured. This methodology prioritizes interpreting the model – understanding why a prediction is made – over simply achieving high accuracy. As Immanuel Kant stated, “Begin not with the things themselves, but with their relations.” The work highlights that truly insightful analysis emerges not from the raw data, but from mapping the relationships within that data, offering a pathway toward reliable and transparent health monitoring systems.

Beyond the Black Box

The pursuit of explainable AI, as demonstrated by this work with Inherently Interpretable Components, inevitably reveals more questions than answers. Current methods often treat interpretability as a post-hoc adjustment – a narrative layered onto a decision already made. This approach, while useful, sidesteps the fundamental issue: can a system be truly transparent if its internal logic isn’t transparent by design? Future research must move beyond simply ‘explaining’ predictions and focus on building models where the reasoning process is as readily observable as the output itself.

A persistent challenge lies in scaling these interpretable components to increasingly complex datasets. While effective with time-series data from wearable sensors, the computational cost and potential loss of predictive power when applied to multi-modal data – combining, for instance, physiological signals with environmental factors – remain significant hurdles. One avenue for exploration involves developing methods to selectively ‘prune’ complex models, retaining only the most salient and interpretable features without sacrificing accuracy.

Ultimately, the value of explainable AI isn’t simply in satisfying intellectual curiosity. The true test will be its ability to foster trust and facilitate genuine collaboration between humans and machines. Perhaps the next iteration of this work should consider incorporating mechanisms for actively soliciting and incorporating human feedback into the model’s reasoning process – a system that not only explains its decisions but also acknowledges the possibility of being wrong.

Original article: https://arxiv.org/pdf/2603.12880.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-03-17 03:14