Edge-AI Enabled Wearables for Construction Safety: Real-Time Physiological Monitoring and Localised Data Processing

Abstract

This paper presents the design, implementation, and controlled evaluation of a proof-of-concept ear-level wearable system that integrates local artificial intelligence for real-time physiological monitoring in construction safety applications. The proposed architecture combines photoplethysmography (PPG), non-contact infrared thermometry, and nine-axis inertial sensing on a Raspberry Pi Pico microcontroller, enabling local inference that reduces dependence on cloud processing. A lightweight logistic regression model with three binary outputs, trained on a subset of the publicly available WESAD dataset (subjects S2–S4), classifies three physiological states relevant to worker safety—elevated PPG variability, drowsiness, and fatigue—directly from the device’s 2 MB flash memory. The principal contribution is demonstrating that ear-level multi-sensor fusion combined with on-device machine learning achieves high agreement with clustering-derived proxy labels under controlled conditions (average F1-score: 97.80% on an unseen test subject) while sustaining sub-second inference latency (<0.5 s). These results support timely supervisor alerting and motivate subsequent field validation in operational construction environments.

1. Introduction

The construction industry remains one of the most hazardous occupational sectors worldwide, characterised by persistent exposure to extreme environmental and physiological stressors. Workers frequently operate under high ambient temperatures exceeding 40 °C, elevated noise levels above 85 dB, and sustained physical exertion that imposes significant cardiovascular strain. These combined conditions contribute to a disproportionate safety burden, accounting for nearly 20% of global occupational fatalities despite representing only 6% of the total workforce [1]. The economic implications are equally severe, with preventable incidents generating approximately USD 11.5 billion annually in medical expenses and productivity losses in the United States alone [2]. In response, heat-related illness has been formally recognised as a critical public health concern, prompting institutions such as the National Institute for Occupational Safety and Health (NIOSH) to establish exposure guidelines that emphasise early physiological monitoring and timely intervention [3]. Within this context, AI-enabled wearable health technologies and preventive physiological risk detection at the network edge have emerged as key enablers of proactive occupational safety management [4,5].

Despite growing awareness and technological advancements, existing wearable safety solutions remain insufficient for the dynamic nature of construction environments. Current systems are limited by three fundamental challenges. First, a significant proportion of devices—estimated at nearly 80%—monitor only a single physiological parameter, thereby failing to capture multi-factorial risks such as concurrent heat stress and physical [6,7,8,9]. Second, reliance on cloud-based architectures introduces unavoidable network latency, which can delay critical intervention in time-sensitive scenarios [10]. Third, adoption of wearable sensing devices in construction remains a practical challenge; survey evidence shows that construction labourers may hesitate to use biometric or location-tracking wearable devices, motivating less intrusive and user-centred designs [11].

Recent advances suggest that ear-level biosensing, combined with edge computing, provides a compelling alternative. In-ear PPG sensors demonstrate 25–30% greater signal stability compared to wrist-based measurements due to reduced motion artefacts at the auricular site [12]. The earbud form factor is unobtrusive, compatible with standard hearing protection, and well accepted by workers. Edge computing further enables on-device machine learning inference, reducing dependence on network connectivity and achieving near real-time response [4]. While prior studies have demonstrated the feasibility of wearable heat-stress monitoring [13] and multi-modal fatigue detection [14,15], an integrated system combining these capabilities into a unified platform remains absent from the literature.

To address this gap, this paper presents a unified ear-level intelligent safety system designed specifically for construction workers. The key contributions of this work are as follows:

• A breadboard proof-of-concept prototype integrating PPG sensing (MAX30102), non-contact infrared thermometry (MLX90614), and a nine-axis inertial measurement unit (BNO055), designed for future ear-level miniaturisation;
• A multi-output logistic regression model deployed directly on the Raspberry Pi Pico’s 2 MB flash memory, achieving local inference with sub-second latency (<0.5 s) without requiring cloud connectivity;
• A controlled evaluation using the WESAD physiological dataset with clustering-derived proxy labels, achieving F1-scores exceeding 95% across all three classification targets and 97.80% average performance on an entirely held-out subject;
• A Streamlit-based supervisor dashboard receiving inference outputs via USB serial and presenting worker status through map-based visualisation and time-stamped alert logs.

The remainder of this paper is organised as follows. Section 2 reviews the technical background and related work. Section 3 describes the system methodology. Section 4 presents the experimental results. Section 5 discusses findings and compares prior work. Section 6 concludes the paper and outlines future directions.

2. Background and Related Work

2.1. Technical Background

2.1.1. Cardiovascular Signal Processing

Heart rate (HR) is derived by detecting peak-to-peak intervals in the photoplethysmographic pulse waveform. Heart rate variability (HRV), an established marker of autonomic nervous system activity and a sensitive indicator of physiological stress and fatigue, is quantified using standard time-domain metrics:

$$\mathrm{SDNN}=\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N}{(N{N}_i-\overline{NN})}^2},$$

(1)

$$\mathrm{RMSSD}=\sqrt{{\displaystyle \frac{1}{N-1}}{\displaystyle \sum _{i=1}^{N-1}}{(N{N}_{i+1}-N{N}_i)}^2}.$$

(2)

Frequency-domain analysis extracts power spectral density in the low-frequency (LF) and high-frequency (HF) bands to characterise sympathetic–vagal balance [16]. Time-domain HRV features reliably discriminate physiological stress from cognitive load in wearable cardiac sensors [17], and their fusion with accelerometer data substantially improves fatigue classification accuracy in free-living environments [18].

2.1.2. Blood Oxygen Estimation

Oxygen saturation (SpO₂) is derived from the ratio of optical absorbance at two wavelengths according to the Beer–Lambert law:

$${\mathrm{SpO}}_2={\displaystyle \frac{{\mathrm{HbO}}_2}{{\mathrm{HbO}}_2+\mathrm{Hb}}}\times 100\%,$$

(3)

$$R={\displaystyle \frac{({\mathrm{AC}}_{\mathrm{red}}/{\mathrm{DC}}_{\mathrm{red}})}{({\mathrm{AC}}_{\mathrm{IR}}/{\mathrm{DC}}_{\mathrm{IR}})}},$$

(4)

$${\mathrm{SpO}}_2=A-B\times R,$$

(5)

where A and B are empirically determined calibration constants. The expressions above (Equations (3)–(5)) represent the theoretical basis for dual-wavelength PPG sensing; the current implementation focuses on PPG amplitude and motion features, with SpO₂ estimation reserved as future work.

2.1.3. Thermal Monitoring

Non-contact skin temperature measurement is governed by Stefan–Boltzmann radiation:

$$P=ϵ\sigma A{T}^4\Rightarrow T_{\mathrm{body}}=f(S_{\mathrm{IR}},T_{\mathrm{ambient}}),$$

(6)

where $ϵ\approx 0.98$ is skin emissivity and $S_{\mathrm{IR}}$ is the calibrated sensor output [19].

2.1.4. Activity Analysis

Worker movement intensity is captured from the 9-axis IMU through the scalar acceleration magnitude and quaternion-based orientation tracking:

$$a=\sqrt{a_x^2+a_y^2+a_z^2},$$

(7)

$$q_t=q_{t-1}+\frac{1}{2}{q}_{t-1}\otimes {\omega }_t·\Delta t.$$

(8)

These metrics support posture classification, fall detection, and exertion quantification [20,21,22].

2.1.5. Composite Risk Indices

A Heat Stress Index (HSI) can be constructed to assess cumulative thermal strain:

$$\mathrm{HSI}=w_1·(T_{\mathrm{body}}-T_{\mathrm{threshold}})+w_2·(\mathrm{HR}-{\mathrm{HR}}_{\mathrm{baseline}})+w_3·\mathrm{ActivityIndex},$$

(9)

where the weights $w_i$ are tuned from empirical data or via regression [23,24]. Similarly, a multivariate Fatigue Index integrates movement history, autonomic imbalance, and ergonomic deviation:

$$\mathrm{FI}=v_1·\mathrm{CumulativeActivity}+v_2·{\displaystyle \frac{{\mathrm{HRV}}_{\mathrm{baseline}}}{{\mathrm{HRV}}_{\mathrm{current}}}}+v_3·\mathrm{PostureDeviation}.$$

(10)

Both composite indices are presented here as contextual references representing directions for future model extension; neither the HSI nor the FI was implemented or evaluated in the current system. The classifier described in Section 3.5 uses individual sensor features extracted directly from the WESAD dataset [25]. On-device deployment of such models is feasible through TinyML compression frameworks [26,27,28], and careful feature engineering can maintain high accuracy within the tight memory budgets of embedded hardware [29,30].

2.2. Related Work

The feasibility of continuous ear-level biometric monitoring has been well established in the prior literature. Adams et al. [31] demonstrated that in-ear PPG enables accurate and continuous heart rate measurement under realistic conditions, while Röddiger et al. [32] provided a comprehensive taxonomy of earable sensing modalities spanning physiological, environmental, and contextual domains.

Building on these foundations, existing wearable safety solutions can be broadly categorised according to their system architecture and deployment strategy [33].

2.2.1. Cloud-Connected Wearables

Commercial systems such as the viAct Smart Watch [34] provide real-time monitoring and predictive safety alerts through cloud-based [35,36,37]. However, reliance on remote processing introduces network-dependent latency and limits responsiveness in time-critical scenarios. CENTEGIX CrisisAlert [38] offers rapid emergency response via a wearable panic-button with precise location tracking, yet it lacks physiological sensing capabilities and operates in a purely reactive manner.

2.2.2. Body-Worn and PPE-Integrated Systems

PyroGuardian [39] exemplifies multi-sensor wearable systems for extreme environments, combining body temperature, heart rate, and GPS monitoring with long-range LoRa communication. While effective for remote supervision, processing is performed off-device and no predictive fatigue modelling is incorporated. SmartShoulder [40] focuses on ergonomic design and haptic feedback but does not include cardiovascular or thermal sensing, nor machine learning-based physiological state classification. Personalised sensing frameworks [41] and mechanical exoskeletons [42] represent complementary occupational safety approaches that similarly lack integrated physiological classification.

2.2.3. Helmet-Based Systems

Smart helmet platforms [43] enable detection of impacts, falls, and environmental conditions with real-time dashboard integration. Deep learning approaches have extended helmets to PPE compliance detection [44] and EEG-based cognitive stress recognition at construction sites [45]. Despite their utility in hazard detection, these systems do not capture cardiovascular signals or support fatigue modelling, and their dependence on cloud-based processing introduces latency unsuitable for rapid intervention.

2.2.4. Earable Research Platforms

OpenEarable 2.0 [46] represents a state-of-the-art ear-level sensing platform integrating a rich suite of biosensors with support for edge inference. Earlier work demonstrated that in-ear EEG reliably detects drowsiness onset [47], and wearable sensor systems have proven effective in rehabilitation monitoring [48], confirming the cross-domain applicability of ear-level physiological sensing. OpenEarable provides a strong architectural foundation; however, it remains a general-purpose platform and has not been adapted for construction safety applications, occupational fatigue labelling, or integrated supervisory alerting pipelines.

Table 1. Systematic comparison of the proposed system against the most relevant prior works. ^† F1-scores reflect agreement with K-means proxy labels, not clinically validated ground truth; see Section 3.5. HR = heart rate; IMU = inertial measurement unit; BLE = Bluetooth Low Energy; GPS = global positioning system.

System/Study	Application Focus	Sensors	Key Strength	Processing/Latency	Form Factor	Critical Limitation
viAct Smart Watch [34] (2024)	AI-powered industrial IoT smartwatch for worker safety	HR, Motion, GPS	Seamless IoT integration; real-time health alerting via cloud	Cloud/>2 s	Wrist-worn	Server-side only; wrist motion artefacts; no on-device physiological classification
PyroGuardian [39] (2024)	IoT wearable for firefighter health and GPS tracking	Body Temp, HR, GPS	Long-range LoRa telemetry; multi-module body-worn design	LoRa/>3 s	Helmet + wrist strap	No on-device inference; no fatigue prediction; firefighting context only
Elitac SmartShoulder [40] (2024)	Smart safety vest with haptic alerting for lone workers	IMU, Haptics	High ergonomic quality; haptic-feedback warning	BLE/>2 s	Body-worn vest	No PPG or thermal sensing; no ML physiological classifier
CENTEGIX CrisisAlert [38] (2024)	Wearable panic-button for emergency dispatch	Push-button, GPS	Precise room-level location; sub-2 s direct notification	Private net/<2 s	Badge/wristband	Purely reactive; no physiological sensors; no predictive capability
Knowit Smart Hat [43] (2024)	Connected helmet for construction accident prevention	IMU, Temperature	Fall and impact detection; live supervisory dashboard	Cloud/>5 s	Helmet-mounted	No cardiovascular sensing; no fatigue model; cloud latency incompatible with emergencies
OpenEarable 2.0 [46] (2025)	Open-source earphone platform for physiological research	PPG (3$\lambda$), 9-axis IMU, Temp, Pressure, Mic	Richest sensor suite in an earbud; fully open hardware	BLE/<1 s	Earbud	General-purpose research tool; no occupational fatigue pipeline; no supervisor alerting
Proposed system (This work) (2026)	Edge-AI ear-level proof of concept for construction safety	PPG, IR Temp, 9-axis IMU	Multi-sensor fusion; 97.80% F1 †; sub-0.5 s on-device inference	On-device/<0.5 s	Ear-level (breadboard)	Controlled lab validation; 3 subjects; proxy labels require field validation

presents a systematic comparison of the proposed system against all reviewed platforms across seven key dimensions. Despite significant progress, existing solutions exhibit a clear gap: none simultaneously combine ear-level multi-sensor biosensing, fully on-device edge AI inference, and real-time safety alerting within a unified system. Addressing this gap is critical for enabling proactive, low-latency intervention in high-risk construction environments and motivates the proposed approach in this work.

3. Materials and Methods

3.1. System Overview

The system operates in two distinct phases. In the offline training phase, raw physiological recordings from the WESAD dataset are segmented into non-overlapping 5 s windows and five scalar features are extracted per window. K-means clustering is applied independently to each target state to generate binary proxy labels, and a multi-output logistic regression model is trained on the resulting labelled feature matrix. The learned weight matrix and bias vector are then exported as a Python literal and compiled into the Raspberry Pi Pico’s 2 MB flash memory alongside the MicroPython 1.23 inference firmware. In the online deployment phase, the Pico reads the MAX30102 (PPG; Analog Devices, Wilmington, MA, USA), MLX90614 (infrared temperature; Melexis Technologies NV, Ieper, Belgium), and BNO055 (9-axis IMU; Bosch Sensortec GmbH, Reutlingen, Germany) sensors via I²C every 5 s, computes the same five-dimensional feature vector in real time, evaluates the logistic regression model through a fixed-point matrix multiplication, and applies a 0.5 decision threshold to each of the three binary outputs. Classification results are transmitted via USB serial to the Streamlit supervisor dashboard, which displays worker status, sensor trends, and time-stamped alerts. Figure 1 illustrates both phases of this pipeline.

3.2. Hardware Design

3.2.1. Prototype Configuration

Figure 2 shows the breadboard prototype. The central component is a Raspberry Pi Pico microcontroller [49] to which three sensors are connected via the I²C bus: a MAX30102 for PPG-based cardiovascular monitoring, an MLX90614 for non-contact infrared thermometry, and a BNO055 nine-axis IMU for motion analysis. The prototype demonstrates the feasibility of the three-sensor fusion architecture in a breadboard form factor. It is important to emphasise that mechanical enclosure design, ear-fit validation, contact pressure optimisation, battery integration, environmental protection against sweat, dust, and vibration, and compatibility with personal protective equipment such as helmets and hearing protection were not evaluated in this work and remain essential hardware development tasks for future ear-level integration.

3.2.2. Sensor Specifications

Table 2. Sensor module specifications.

Sensor	Model	Manufacturer	Function
PPG/SpO2	MAX30102	Analog Devices, Wilmington, MA, USA	Photoplethysmography; heart rate and blood oxygen capability
Thermometer	MLX90614	Melexis Technologies NV, Ieper, Belgium	Non-contact IR temperature (±0.5 °C)
9-Axis IMU	BNO055	Bosch Sensortec GmbH, Reutlingen, Germany	Fused acceleration, gyroscope, and magnetometer

summarises the three sensing modules. The MAX30102 [50] uses red and infrared LEDs to measure the photoplethysmographic waveform. The MLX90614 [51] is a factory-calibrated non-contact infrared thermometer with a measurement accuracy of ±0.5 °C. The BNO055 [52] integrates a triaxial accelerometer, gyroscope, and magnetometer with an onboard sensor fusion processor, providing orientation-corrected acceleration and angular velocity outputs without host-side computation.

3.2.3. Microcontroller

The Raspberry Pi Pico [49] is built around the RP2040 dual-core ARM Cortex-M0+ processor running at up to 133 MHz. It provides 264 KB of SRAM for runtime computation and 2 MB of on-chip flash memory, which is sufficient to store the full logistic regression weight matrix alongside the MicroPython firmware. In the current prototype, communication with the supervisor dashboard is achieved over USB serial; wireless BLE/WiFi integration was not implemented and is planned for subsequent hardware revisions.

3.3. Software Environment

Table 3. Principal software tools and their roles in the development pipeline.

Tool/Library	Role
Python 3.10	Primary development language for training and firmware development
Google Colab (Python 3.10 runtime)	Notebook environment for model training and feature analysis
Thonny IDE 1.4	Deployment environment for MicroPython firmware on Raspberry Pi Pico
MicroPython 1.23	Firmware runtime used on the Raspberry Pi Pico
Scikit-learn 1.3 [53]	Logistic regression and K-means clustering
Pandas 2.1	Data processing, windowing, and feature handling
NumPy 1.26	Numerical computation and feature extraction
Matplotlib 3.8	Visualisation of features, clusters, and results
Streamlit 1.32	Real-time supervisor monitoring dashboard

lists the principal software tools used in this project. The machine learning training pipeline was implemented in Python 3.10 using Scikit-learn 1.3 [53] for model training and K-means clustering, executed in Google Colab (Python 3.10 runtime) to leverage cloud-based GPU acceleration during development. Sensor firmware for the Pico was written in MicroPython 1.23 and deployed via Thonny IDE 1.4. The real-time supervisor interface was constructed using Streamlit 1.32, with Pandas 2.1, NumPy 1.26, and Matplotlib 3.8 supporting data handling and visualisation.

3.4. Dataset and Feature Engineering

3.4.1. WESAD Dataset

The Wearable Stress and Affect Detection (WESAD) dataset [54] was selected as the training and evaluation corpus. WESAD is a widely used benchmark for physiological state detection, collected under controlled laboratory conditions from 15 subjects performing baseline, stress-inducing, and amusement tasks. It provides synchronised PPG, skin temperature, and triaxial acceleration recordings that directly correspond to the three sensor modalities of the proposed hardware. Subjects S2 and S3 were used for model training, and Subject S4 was withheld entirely as an external validation set to assess generalisation to unseen individuals.

3.4.2. Feature Extraction

Five scalar features are computed from each non-overlapping 5 s window of raw sensor data: bvp_mean (mean PPG waveform amplitude over the window), bvp_std (standard deviation of the PPG waveform, capturing short-window amplitude variability), acc_mean (mean acceleration magnitude, indicating overall activity level), acc_std (standard deviation of acceleration magnitude, capturing movement irregularity), and temp_mean (mean skin temperature over the window). It is important to note that bvp_mean and bvp_std are raw waveform amplitude descriptors and are not equivalent to clinically derived heart rate or heart rate variability metrics, which require peak detection, inter-beat interval computation, and artefact rejection algorithms that were not implemented in the current pipeline. This minimal five-dimensional feature vector was deliberately chosen to satisfy the 2 MB flash constraint of the deployment hardware while preserving physiological interpretability for safety supervisors.

Although WESAD provides synchronised PPG, skin temperature, and triaxial acceleration recordings that match the sensing modalities of the proposed hardware, it was collected under controlled laboratory conditions and does not reproduce the environmental stressors characteristic of construction sites, including sustained heat exposure exceeding 40 °C, heavy personal protective equipment, airborne dust and vibration, manual material handling, and construction-specific fatigue patterns. The proxy labels derived from K-means clustering capture statistical structure in the physiological feature space but have not been validated against expert-annotated occupational ground truth. Consequently, all performance results reported in this study should be interpreted as a controlled proof-of-concept evaluation, and field validation with domain-specific labelled data is an essential prerequisite for operational deployment.

3.5. Machine Learning Approach

3.5.1. Proxy Label Generation via Clustering

Because WESAD does not include annotations for the specific occupational states targeted in this work, K-means clustering was used as a physiologically motivated proxy labelling strategy. The rationale is that the three target conditions—elevated PPG variability, drowsiness, and fatigue—produce distinguishable signatures in the five-dimensional feature space. Elevated PPG variability is identified from windows with high bvp_std; drowsiness from windows combining low acc_std with suppressed bvp_std; and fatigue from windows that show elevated temp_mean accompanied by high acc_std. Three independent binary K-means assignments (one per state) are thus used as surrogate ground-truth labels for classifier training.

It is important to acknowledge that all performance metrics reported in Section 4 reflect classifier agreement with these proxy labels, not clinically validated annotations. The results therefore demonstrate technical feasibility under controlled conditions and should not be interpreted as a guarantee of operational accuracy on real construction sites. Field validation with expert-annotated ground truth data is identified as an essential priority for future work.

3.5.2. Classifier Training

A multi-output logistic regression model was trained on subjects S2 and S3 with an 80/20 train–test split, using Scikit-learn’s built-in class-weight balancing to address label imbalance. The model produces three simultaneous binary outputs corresponding to the three physiological states. Logistic regression was selected over more expressive alternatives (e.g., gradient-boosted trees, neural networks) because its linear weight structure can be represented as a compact matrix multiplication executable within the memory and processing constraints of the RP2040. This choice also ensures that alert triggers are fully interpretable: a supervisor can understand, for example, that a fatigue alert was raised because body temperature and movement variability both crossed their learned thresholds.

3.5.3. On-Device Deployment

After training, the weight matrix and bias vector are serialised to a Python literal and written directly to the Pico’s flash memory alongside the inference firmware. At runtime, the device: (i) reads all three sensors over I²C; (ii) computes the five features from a 5 s buffer; (iii) evaluates the logistic regression via a fixed-point matrix multiplication; and (iv) applies a 0.5 decision threshold to each output. Classification results are transmitted to the dashboard over USB serial at each inference cycle.

3.6. Supervisor Dashboard

A real-time Streamlit dashboard was developed to present inference outputs to safety supervisors. In the current prototype, classification results are received via USB serial from the Raspberry Pi Pico; wireless communication via BLE or WiFi was not implemented and is identified as future work. The interface displays the current physiological status of each monitored worker (Normal or Critical), live sensor readings for PPG amplitude, temperature, and motion intensity, and time-stamped alert logs for triggered classifications. Worker positions are shown as manually assigned markers on a map of the Tabuk City construction area for visualisation purposes; no GPS, BLE-based, WiFi-based, or indoor localisation system was implemented or evaluated in this work, and formal localisation integration is planned as a future development priority. Figure 3 and Figure 4 illustrate the normal and alert display states, respectively.

3.7. Evaluation Protocol

Model performance is evaluated using four standard metrics: accuracy (proportion of correct binary classifications), precision (positive predictive value), recall (sensitivity), and F1-score (harmonic mean of precision and recall), which is used as the primary comparative metric throughout. Confusion matrices are analysed for each physiological state separately to identify systematic misclassification patterns. Inference latency is measured at batch size 1—the operationally relevant setting for single-worker real-time monitoring—and compared with an XGBoost baseline to contextualise the deployment trade-off.

4. Results

4.1. Dataset Characteristics

The combined feature dataset extracted from subjects S2, S3, and S4 comprises 4217 windows, each representing 5 s of synchronised physiological recordings. Principal Component Analysis (PCA) was used to project the five-dimensional feature space into two dimensions for visualisation; Figure 5 shows the resulting scatter plot alongside the three K-means cluster assignments, confirming that the proxy labels occupy well-separated regions of the feature space and are therefore suitable as surrogate training targets.

4.2. Classification Performance

F1-Score Results

Figure 6 presents F1-scores for both the internal test split (S2 + S3) and the held-out external validation subject (S4). The classifier achieved strong performance across all three physiological states under both evaluation conditions, with every F1-score exceeding 95%—well above the 80% design target. Drowsiness detection yielded the highest scores (internal: 98.98%, external: 98.55%), while the average F1-score across all states reached 97.53% on the internal split and 97.80% on the external test, indicating that classification performance generalises to unseen individuals without degradation.

It bears repeating that these scores reflect agreement with K-means proxy labels generated from the same dataset rather than independently validated clinical annotations. They demonstrate that the classifier has successfully learned the statistical structure of the clustering-derived label space, which serves as a proof of technical feasibility. Achieving comparable performance against expert-annotated field data is a necessary subsequent step before clinical or operational deployment.

4.3. Confusion Matrix Analysis

Figure 7 presents confusion matrices for internal and external testing across all three classification targets. In all cases, the dominant mass of predictions falls on the diagonal, confirming correct classification. Off-diagonal misclassification rates are consistently low, ranging from approximately 1% to 3%, and the pattern of errors is largely consistent between the internal and external test sets—indicating that the learned decision boundaries generalise reliably rather than overfitting to the training subjects.

4.4. Inference Latency

Figure 8 compares per-sample inference latency for logistic regression and XGBoost at batch size 1. Logistic regression completes inference in well under 0.5 s on the Raspberry Pi Pico, consistent with real-time monitoring requirements. XGBoost incurs substantially higher latency at small batch sizes owing to its sequential tree-traversal structure, rendering it impractical for deployment on resource-constrained embedded hardware. This result supports the selection of logistic regression as a well-suited model for this prototype, balancing deployment efficiency, interpretability, and memory constraints. However, this comparison is limited to inference latency at batch size 1; a comprehensive model-selection study comparing accuracy, flash memory footprint, SRAM usage, quantisation error, power consumption, and decision-threshold sensitivity across logistic regression, decision trees, quantised neural networks, and gradient-boosted ensembles remains an important direction for future work.

5. Discussion

5.1. Performance and Generalisation

The empirical results validate the core hypothesis of this work: that a compact, ear-level wearable prototype integrating three complementary sensor modalities and an on-device logistic regression classifier can achieve high-accuracy physiological state detection under controlled conditions. An average F1-score of 97.80% on a completely unseen test subject is particularly encouraging given the small training corpus of only two subjects and suggests that the feature-space structure of the proxy labels is physiologically consistent across individuals. The sub-second inference latency achieved on the Raspberry Pi Pico (<0.5 s) supports real-time alerting and substantially reduces dependence on cloud-based processing, a key limitation of existing alternatives [10].

The choice of logistic regression merits further discussion. While ensemble methods or shallow neural networks could, in principle, capture non-linear interactions among the five features, they would require substantially more flash storage and multiply-accumulate operations per inference cycle—both severely constrained on the RP2040 (264 KB SRAM, no floating-point unit). Beyond hardware constraints, the linear weight structure provides inherent interpretability: a supervisor receiving an alert can understand, at the feature level, which combination of elevated temperature, PPG amplitude variability, and motion anomaly triggered the notification. In safety-critical operational contexts, this transparency is likely to support trust and adoption more effectively than a black-box classifier with marginally superior accuracy.

5.2. Prototype Validation and Design Rationale

The breadboard prototype successfully demonstrates the technical feasibility of the three-sensor fusion architecture. Ear-level PPG acquisition with the MAX30102 benefits from the substantially lower motion noise present at the auricular site compared with the wrist, a well-documented advantage confirmed by comparative PPG accuracy studies [12,55]. Non-contact infrared thermometry with the MLX90614 provides a minimally intrusive proxy for core body temperature, which has been validated as a reliable indicator of occupational thermal strain at peripheral measurement sites [56]. The BNO055’s onboard sensor-fusion processor offloads quaternion computation from the RP2040, ensuring that the microcontroller’s limited CPU budget remains available for feature extraction and inference.

The ear-level form factor is informed by prior design guidelines for in-ear biosensors [57] and the architectural choices demonstrated in the OpenEarable platform [46]. By targeting a standard earbud enclosure in future miniaturisation, the system would be compatible with existing personal protective equipment requirements on construction sites, address the high abandonment rates associated with bulky body-worn devices [33], and support the fundamental shift from reactive to predictive occupational safety monitoring that the literature identifies as the most urgent unmet need in the field [5].

6. Conclusions

This paper presented the design, implementation, and controlled evaluation of an ear-level edge-AI wearable proof of concept for real-time physiological monitoring in construction safety. By fusing PPG, non-contact infrared thermometry, and nine-axis inertial sensing on a Raspberry Pi Pico microcontroller, and deploying a lightweight multi-output logistic regression classifier entirely on-device, the system achieves local inference at sub-second latency that substantially reduces dependence on cloud-based processing. Evaluated on the WESAD physiological dataset, the classifier reached an average F1-score of 97.53% on internal validation and 97.80% on a held-out test subject, demonstrating that strong agreement with clustering-derived proxy labels generalises across individuals even with a small training corpus. Confusion matrix analysis confirmed consistently low misclassification rates of 1–3%, and latency benchmarking established that logistic regression is well suited to the resource constraints of the RP2040 compared with alternative models such as XGBoost.

Several limitations are acknowledged. The evaluation involved only three WESAD subjects, which constrains the representativeness of the reported inter-subject generalisation results. All experiments were conducted under controlled laboratory conditions; the motion artefacts, temperature extremes, and environmental variability characteristic of real construction sites were not captured in the training data. Performance metrics reflect agreement with K-means proxy labels rather than independently validated clinical annotations, and field validation against expert-labelled ground truth is essential before any operational deployment. Continuous multi-sensor operation imposes battery constraints that have not yet been characterised over full working shifts, and the current USB-based communication channel limits deployment to tethered monitoring scenarios.

Addressing these limitations defines the roadmap for future development. Priority items include field validation studies with 30 or more subjects drawn from operational construction environments with expert-annotated physiological ground truth; PCB miniaturisation to produce a self-contained production-grade earbud enclosure; integration of BLE or WiFi for wireless dashboard communication and GPS-based geolocation; adaptive personalisation techniques to calibrate physiological thresholds to individual baselines; model compression and quantisation to extend battery life while maintaining classification accuracy; a comprehensive model-comparison study evaluating accuracy, memory footprint, and power consumption across alternative classifiers; and exploration of predictive time-series architectures capable of forecasting physiological deterioration before threshold crossing rather than classifying established states.

This work provides initial evidence that the convergence of ear-level multi-sensor biosensing and locally deployed machine learning represents a technically viable direction for next-generation construction safety monitoring. By substantially reducing dependence on cloud processing, targeting a minimally intrusive form factor, and supporting timely supervisor alerting, the proposed architecture addresses three of the most persistent barriers to wearable adoption in high-risk occupational settings. Realising the full potential of this approach will require field validation with expert-annotated data from operational construction environments, hardware miniaturisation into a self-contained earbud enclosure, and wireless communication integration.