Chapter Four · failure evidence
What Wearable & Mobile Sensing got wrong, from 51 dissertations
The records document technical limitations and practical barriers encountered when developing wearable and mobile sensing systems for health, activity, and biomechanical monitoring. Common challenges include insufficient degrees of freedom from sparse inertial sensors, poor participant adherence due to device burdens, sensitivity of physiological signals to confounding environmental factors, and failure of fixed heuristics to generalize across individuals. These records come from PhD theses at 20 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Sparse or single-point inertial sensors fail to capture complex kinematics and activity states
Minimal or single-site accelerometer and inertial measurement unit configurations lacked the degrees of freedom required to detect balance perturbations, joint angles, and full-body movements. Adding more inertial units or relying solely on acceleration magnitude degraded cross-subject generalization and failed to separate distinct postures or calculate accurate ground reaction forces.
Considered and rejected
Considered and rejected: Rejected relying solely on wearable wrist sensors because they failed to capture isolated finger movements and complex full-body kinematics.
Where is the Person in Personalized Medicine? The Missing Expert in Adaptive Neurotechnology · ResearchWorks
Considered and rejected
Considered and rejected: Rejected using only total acceleration (accel_total) magnitude for wrist-worn sensors because it could not discriminate between sitting and standing.
Deep Neural Network Representations of Physiological Time-Series Sensor Data for Improved Recognition Performance · TXST Digital Repository
Tried and failed
adding triaxial accelerometry to multimodal sensing applied to sleep wake detection. Outcome: no signal. Reason: wrist accelerometry features did not provide additional discriminative value beyond combined video and PPG
Combining wearables and nearables for patient state analysis · Georgia Tech
Tried and failed
kinematic estimation from minimal wearable inertial sensors applied to detecting gait variability under balance perturbations. Outcome: no signal. Reason: sensor-derived variability metric lacked statistical significance across different perturbation prompt levels
Tried and failed
kinematic variability metrics from sparse inertial sensors applied to dynamic balance perturbation assessment. Outcome: no signal. Reason: No statistically significant correlation with normalized balance state when controlling for walking speed
Tried and failed
kinematic inertial sensor estimation applied to ground reaction force estimation. Outcome: worse than baseline. Reason: high stride-by-stride kinematic variability and non-sagittal compensatory movements degraded accuracy relative to direct force sensing
Tried and failed
multi-sensor fusion with additional inertial measurement units applied to locomotion speed estimation. Outcome: worse than baseline. Reason: additional sensors did not reduce error and degraded user-independent cross-subject generalization
Sensor Fusion Representation of Locomotion Biomechanics with Applications in the Control of Lower Limb Prostheses · Georgia Tech
Tried and failed
inertial motion capture for joint angle measurement applied to human joint kinematics validation. Outcome: worse than baseline. Reason: inertial sensors failed to agree with baseline goniometer measurements across multiple joints
Tried and failed
single-point inertial measurement unit tracking applied to complex exercise motion classification. Outcome: no signal. Reason: single-channel kinematics lacked sufficient spatial degrees of freedom without multi-site bodily strain sensing
Study Of Intelligent Wearables and Adaptive Systems to Measure Physiological and Physical Data for Human-Machine Interfaces · Georgia Tech
Considered and rejected
Considered and rejected: Rejected calculating stepping kinematics and trunk angular velocity from wearable IMUs in the real world due to sensitivity to motion artifacts and inter-sensor distance estimation errors.
Non-Treadmill Trip Training – Laboratory Efficacy, Validation of Inertial Measurement Units, and Tripping Kinematics in the Real World · Virginia Tech
Considered and rejected
Considered and rejected: Rejected measuring physical activity with simple accelerometers alone in population health studies, recommending wearable devices that incorporate continuous heart-rate monitoring.
Essays on Selection in Health Behaviors · Harvard
Fixed thresholds and generic models fail to generalize across diverse users and conditions
Rule-based heuristics, fixed thresholds, and global machine learning models experienced high false positive rates and poor accuracy across different movement heights, postures, and geographic regions. Subpopulation-specific training or rigid biomimetic profiles also degraded performance because small sample sizes reduced generalization and failed to match individual user dynamics.
Tried and failed
threshold-based electromyography state machine controller applied to wearable movement assistance. Outcome: did not generalise. Reason: fixed thresholding caused frequent false positive activations across varying task postures and movement heights
Assisting and Evaluating Upper Extremity Movements with Wearables Systems · Harvard
Tried and failed
training subpopulation-specific machine learning models applied to movement classification from wearable sensors. Outcome: worse than baseline. Reason: smaller subpopulation training dataset size hurt generalization compared to pooling all subject data
Fall detection and balance control during human movement · UT Austin
Tried and failed
proprietary vendor event classification algorithm applied to wearable kinematic sensor impact detection. Outcome: did not generalise. Reason: algorithm misclassified genuine impacts as spurious artifacts and missed true events compared to video verification
Lost to a baseline
Multi-country generalized machine learning models were outperformed by continent-specific and country-specific models for smartphone-sensing-based mood and activity inference.
The iLog methodology for fostering valid and reliable Big Thick Data · IRIS - UNITN - prod
Considered and rejected
Considered and rejected: Using a fixed percentage threshold for Mahalanobis distance outlier removal was rejected because it generalizes poorly across individual sensors compared to probability-based thresholds.
Security and Reliability in Pervasive Computing · IRIS - POLITO - prod
Tried and failed
wearable sensor heuristic activity classification applied to travel mode and physical activity detection. Outcome: worse than baseline. Reason: built-in proprietary algorithms showed low accuracy compared to validated self-reported travel diaries
Applications of causal inference in environmental policy and transport studies · Imperial
Tried and failed
applying mismatched actuation profiles across movement modes applied to assistive wearable robotics control. Outcome: worse than baseline. Reason: incorrect timing and force profile disrupted natural biomechanics and increased metabolic effort
Tried and failed
biomimetic control timing parameterization applied to wearable robotic gait assistance. Outcome: worse than baseline. Reason: biologically-inspired profiles failed to account for individual user dynamics compared to individualized human-in-the-loop optimization
Wearable setups suffer from high user burden, battery demands, and participant non-compliance
Multi-device configurations, frequent charging requirements, and complex body-worn hardware led to daily living inconvenience and participant non-compliance, particularly in cognitively impaired populations. In addition, users deactivated background smartphone sensing features to conserve battery life, creating severe data missingness.
Considered and rejected
Considered and rejected: Rejected bodily worn tracking sensors due to lack of suitability for natural walk-up-and-use public scenarios.
How can people’s spatial behaviour be used to dynamically lay out content on multi-user, interactive screens, and how does this dynamic layout affect people’s spatial behaviour? · University of Nottingham Repository
Considered and rejected
Considered and rejected: Rejected multi-device wearable setups (bilateral feet/wrists/sternum) for continuous free-living monitoring due to high wearability burden, reduced patient compliance, and battery charging demands.
Considered and rejected
Considered and rejected: Wearable sensors for hand gesture recognition were rejected because they are expensive, complex, and unnatural.
Human-Interactions with Robotic Cyber-Physical Systems (CPS) for Facilitating Construction Progress Monitoring · Virginia Tech
Considered and rejected
Considered and rejected: Excluded wearable sleep trackers in favor of under-the-mattress sensors to avoid data loss from short battery life, frequent charging, and adherence issues in cognitively impaired populations.
From behavioural patterns to rain barriers: representation learning for ageing and dementia · Imperial
Considered and rejected
Considered and rejected: Rejected wearable multi-device setups (e.g., S-Smart pocket smartphone + wrist IMU + foot IMU) due to daily living inconvenience
Smartwatch-based semantic learning of elements of human behaviour · Oxford
Considered and rejected
Considered and rejected: Rejected using smartphone passive sensing streams (e.g., Bluetooth device counts) due to participant deactivations for battery conservation causing excessive missingness.
Forecasting depressive symptom deterioration using wearable sensor data and LSTM models: a longitudinal analysis from the RADAR-MDD Study · University of Nottingham Repository
Tried and failed
wearable activity tracking applied to cognitively impaired older adults. Reason: poor user adoption, adherence, and tolerance leading to non-compliance
Considered and rejected
Considered and rejected: Rejected wearable devices for continuous tracking due to low patient compliance associated with cognitive impairments
Physiological and passive mobile signals fail to correlate reliably with target mental or physical states
Physiological metrics such as heart rate, peripheral temperature, and electromyography showed high variance overlap and were confounded by ambient heat or prior physical exercise. Similarly, passive mobile features and vital sign metrics lacked detectable predictive signals for depression scale scores or personal pollutant inhalation exposures.
Tried and failed
multimodal physiological stress detection with wearable sensors applied to autonomic tone and stress monitoring. Outcome: unstable. Reason: High ambient temperature caused GSR runaway and prior exercise disrupted HRV LF/HF ratio reliability
A Skin-Like Sternal Patch to Monitor Autonomic Tone During Cognitive Stress and Sympathetic Arousals in Disordered Sleep · Georgia Tech
Tried and failed
wearable physiological sensors for affective state detection applied to detecting emotional responses to transient events. Outcome: no signal. Reason: sensors failed to reliably isolate heart rate changes triggered by transient unexpected errors
Artificial social constructivism for long term human computer interaction · Imperial
Tried and failed
peripheral temperature sensing for affective state classification applied to affective state monitoring. Outcome: no signal. Reason: wide variance overlap across neutral, fatigued, and stressed distributions prevented class separation
A Wearable-Based Hidden Markov Model for Health Monitoring in Conflict Zones: A Case Study of Gaza · Harvard
Tried and failed
heuristic feature normalization without patient baseline applied to physiological regression from wearable sensor features. Outcome: no signal. Reason: external statistics and weight scaling failed to capture individual physiological baseline variability
Considered and rejected
Considered and rejected: Rejected electromyography (EMG) sensors for obtaining muscle forces due to signal fuzziness and sensitivity to environmental conditions
Efficient and Safe Robot Planning in Human-Robot Collaboration · DSpace at SUNY Buffalo
Tried and failed
gradient boosted trees with automated feature selection applied to depression prediction from smartphone sensing. Outcome: no signal. Reason: passive mobile sensor features lacked detectable predictive signal for depression scale scores
Tried and failed
predicting environmental exposure from wearable physiological signals applied to personal inhalation exposure estimation. Outcome: no signal. Reason: physiological metrics alone lack sufficient correlation with breathing-zone pollutant concentrations across varied activities
Assessment of Monitoring Strategies for Inhalation Exposure and Occupancy in Office Environments · EPFL
Sensor detachment and physical placement shifts degrade signal quality and accuracy
Physical loading changes and sensor detachment caused spatial shifts in activation patterns that proximity sensors failed to detect reliably. Furthermore, direct skin taping made placement replication difficult across sessions, and tensile strain modes suffered from pressure cancellation compared to compression.
Tried and failed
single sensor placement optimization for signal quality applied to biosignal decoding across mechanical conditions. Outcome: did not generalise. Reason: Spatial shifts in signal activation patterns occurred when physical loading or wearable interface conditions changed.
Learning steps: models of intent-driven lower limb prosthesis use · Imperial
Tried and failed
threshold-based proximity sensing for contact detection applied to wearable sensor coupling validation. Outcome: did not generalise. Reason: High false negative rate yielding only 29.7% sensitivity in detecting sensor detachment events.
Optimal Head Impact Signal Processing and the Description and Perception of Head Impact Exposure in Female Adolescent Ice Hockey Players · Virginia Tech
Tried and failed
tensile strain mode physiological transducer placement applied to respiratory monitoring wearable sensor. Outcome: worse than baseline. Reason: stretching mode caused pressure cancellation, resulting in substantially lower sensitivity than compression mode
Stretchable wearable wireless sensors for physiological monitoring of humans and dogs · Imperial
Tried and failed
analog thermistor temperature sensing applied to wearable physiological monitoring. Outcome: worse than baseline. Reason: non-linear resistance-temperature characteristics caused significantly higher noise than digital bandgap sensors
A Wearable-Based Hidden Markov Model for Health Monitoring in Conflict Zones: A Case Study of Gaza · Harvard
Lost to a baseline
Wrist-worn sensors underperformed chest-worn sensors across all 5 holding assessment scenarios (0.738 vs 0.870 accuracy in Scenario 1)
Leveraging pervasive data to study and support mother-infant dyads in the wild · UT Austin
Considered and rejected
Considered and rejected: Rejected attaching strain sensors directly to the skin with tape due to difficulty in consistently replicating sensor placement across sessions without a trained professional.
Stretch sensors for measuring knee kinematics in sports · Imperial
Hardware discrepancies and consumer device variations impair measurement consistency
Integrating heterogeneous wearable sensor streams introduced significant inter-device measurement variability that undermined data comparability across participants. Additionally, research-grade and smartwatch devices exhibited higher measurement errors and noise compared to dedicated hardware baselines.
Tried and failed
integrating heterogeneous consumer wearable sensor streams applied to remote physical activity tracking. Reason: inter-device measurement variability undermined data comparability across study participants
Lost to a baseline
Research-grade wearables (Empatica E4, Biovotion Everion) had higher resting HR measurement error (MAE 13.9 ± 7.8 bpm) than consumer devices (MAE 7.2 ± 5.4 bpm).
Discovering Digital Biomarkers of Glycemic Health from Wearable Sensors · DukeSpace
Considered and rejected
Considered and rejected: Rejected using wearable sensor spectral monitoring data and activity diaries for quantitative prior-light-history tracking due to sensor inaccuracies/malfunctioning and inconsistent participant reporting.
Alertness in work environments : on the role of indoor daylight exposure · EPFL
Lost to a baseline
Smartwatch audio achieved higher MAE (2.86 ml/s to 2.9 ml/s) compared to Mi A1 smartphone (2.5 ml/s) and Ultramic384 (2.6 ml/s), but was preferred for wearable convenience.
Urinary flow estimation through sound-based uroflowmetry and machine learning · DeustoTeka
Left open by the authors
Problems the authors named and did not get to.
Left open
Adapt the pressure sensing wearable hardware to monitor heels or shoulder blades, or integrate sensors directly into support surfaces. Blocker: Requires custom physical hardware prototyping, wearable sensor integration, and human subject testing without access to the physical apparatus.
A Wearable Device to Inform Pressure Injury Prevention Support Surfaces Selection and Design · MIT
Left open
Test the smart wearable system on a larger, diverse population beyond healthy young adults for dynamic cardiac and respiration monitoring. Blocker: Requires custom physical wearable hardware and recruitment of human clinical cohorts for dynamic physiological data collection
A Novel Smart Wearable System with Edge Computing AI for Cardiac Disease Detection and Continuous Monitoring · Texas Tech
Left open
Develop calibration-free generalized torque models or bodyweight-based calibration procedures for multimodal wearable sensors without isokinetic dynamometers. Blocker: Requires custom multimodal hardware (IMUs and soft sensors), human subject testing, and ground-truth biomechanical measurement equipment
Tracking Joint Kinematics and Muscle Kinetics with Multimodal Wearable Sensors · Harvard
Left open
Develop reinforcement learning algorithms for personalized stress detection adapting to individual physiological wearable data over time. Blocker: None
Mindfulness in the Moment: A Personalized Wearable Reflection System for Emotional Regulation · Harvard
Left open
Develop wearable multimodal sensing methods to detect postprandial hypotension and nocturnal non-dipping blood pressure. Blocker: Requires specialized wearable multimodal cardiopulmonary sensor hardware and clinical study data on CVAD patients.
Advancing Wearable Cardiopulmonary Monitoring Through Multimodal Sensing of Cardiovascular, Pulmonary, and Respiratory Muscle Function · Georgia Tech
Left open
Develop multi-sensor fusion combining IMU and FSR sensor data to reduce false positives in real-time step detection. Blocker: Requires physical wearable hardware equipped with synchronized IMU and FSR sensors or proprietary dual-sensor gait data.
Left open
Develop adaptive signal selection algorithms for redundant sensor arrays to dynamically correct for wearable placement misalignment without ground truth data. Blocker: Requires custom physical wearable sensor hardware (SCARS soft strain gauges, ultrasound arrays) and real-time physical acquisition apparatus
Multi-Modal Deformation Sensing for Evaluating Muscle Function · Harvard
Left open
Integrate digital accelerometers and input force sensors into wearable active vibration hardware while reducing the accelerometer count to a single output. Blocker: Requires designing, fabricating, and testing custom physical wearable sensor hardware
Novel Methods Using Acoustics and Bioimpedance to Assess Musculoskeletal Health and Performance · Georgia Tech
Left open
Evaluate joint feature estimation, spatio-temporal sEMG heatmaps, and alternative neural network architectures for multi-step muscle fatigue forecasting horizons. Blocker: Requires sEMG data from the custom wearable device developed in the thesis
Left open
Implement the dynamic EMG decomposition and control pipeline on embedded wearable hardware and evaluate hardware trade-offs. Blocker: Requires custom embedded hardware design and wearable physical apparatus/sensors for physical testing.
Non-invasive neural interfacing for wearable electromyographic systems · Imperial
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.