Soft wearable flexible bioelectronics integrated with an ankle-foot exoskeleton for estimation of metabolic costs and physical effort

Musculoskeletal disorders remain prevalent among workers performing repetitive, high-effort tasks such as squatting and lifting. Wearable robots (exosuits/exoskeletons) can reduce physical effort and injury risk but require personalized, rapid tuning based on user biofeedback. Human-in-the-loop optimization typically relies on indirect calorimetry (mask-based respirometry) to estimate metabolic cost, but such systems are bulky, slow to respond (>3 min), and ill-suited for field deployment. Alternative proxies such as heart rate (HR) and heart rate variability (HRV), particularly RMSSD derived from ECG, can respond faster and be more practical if measured with comfortable, motion-robust sensors. However, conventional HR straps and rigid devices are uncomfortable and artifact-prone, and epidermal bioelectronic electrode fabrication can be complex and costly. This study introduces a soft, flexible biopatch integrated with an ankle–foot orthosis (AFO) that measures high-quality ECG/HR/IMU signals to estimate metabolic cost and physical effort via HRV-RMSSD, aiming to replace mask-based calorimetry for rapid, personalized exoskeleton assistance.

Prior work demonstrates wearable robots can reduce metabolic cost during walking, running, and squatting and can be personalized using human-in-the-loop optimization that incorporates physiological signals, often via indirect calorimetry. Commercial calorimetry systems are accurate but heavy, rigid, time-consuming to set up, and restrictive to laboratory environments. HR-based measures have been explored (e.g., HR straps) for intensity/effort estimation, but rigid form factors and poor skin conformity produce motion artifacts and discomfort. Additionally, complexity in fabricating epidermal electrodes has hindered scalability. Existing exosuit studies largely rely on respirometry for metabolic estimation and do not relate HRV-RMSSD to metabolic cost. This work addresses these gaps by providing a soft, conformal ECG biopatch with robust signal quality and by quantitatively relating HRV-RMSSD to metabolic cost and perceived exertion during exoskeleton-assisted activities.

Device design and fabrication: A soft flexible biopatch (SFB) with stretchable, laser-patterned Au electrodes on PDMS (10 nm Cr/200 nm Au) was fabricated using a femtosecond laser micromachining system for rapid, maskless patterning. Electrodes (8.5 cm long, 7.5 cm center-to-center spacing; 150 µm width; ~8 µm total thickness) and serpentine interconnects were transferred to a stretchable fabric substrate (3M 9907T) using water-soluble tape. The electronics include a flexible printed circuit (fPCB) with ADS1292 (ECG A/D), IMU20948 (IMU), and nRF52832 (microcontroller/BLE), powered by a 3.7 V, 40 mAh battery with 1.8 V and 3.3 V regulation. The mainboard (25 x 14 mm, double copper layer, 12.7 µm polyimide) is stacked over a power board (18 x 10.2 mm) to reduce rigidity. The assembly is encapsulated in soft elastomers (Ecoflex Gel and Ecoflex 30). Connections to electrodes use asymmetric conductive films and silver paint cured at 60 °C for 1 h. Battery life exceeds 9 h; magnetic charging takes ~30 min. Mechanical reliability: Finite element analysis under 30% uniaxial strain showed <4% maximum von Mises stress in electrodes and interconnects. Cyclic uniaxial stretch tests (30% strain) showed minimal resistance change per cycle (<0.05 Ω; <0.1%) and <1% change for the whole device, with optical inspection indicating no delamination. The device sustained up to 50% elastic elongation, yielded thereafter, and fractured at ~123% extension; Young’s modulus ~500 kPa. fPCB sustained 100 bending cycles (±15°) without resistance drift. Breathability testing (MVTR over 72 h) showed 3M 9907T superior to PDMS, Ecoflex, and micropore; peel tests indicated sufficient adhesion (up to ~1 N pull; peel energy ~70 J/m^2 dry, ~40 J/m^2 with water). Robotic AFO: A tethered ankle–foot orthosis emulator provided active plantarflexion via Bowden cables with ankle ROM −80° to 50° (plantarflexion) and −20° to 20° (inversion–eversion). Assistance during squat ascent/descent used an impedance-based torque profile proportional to ankle angle, with separate parameters for phases. Load cells and encoders measured torque and angle; a low-level torque controller commanded actuator velocity. Human study: Six adults (4 male, 2 female; ages 20–30) performed squatting and then walking/running while wearing the SFB; squatting was also performed with the AFO under randomized assistance conditions, including no-power and no-assist. Squat protocol: 4-min bouts with 2 s squats (1 s down/1 s up) interleaved with 6 s standing; 12 min rest between squat conditions. Walking and running bouts were 3 min each, with randomized order and rest. A mask-based indirect calorimeter (Cosmed K5) recorded VO2 and VCO2 for comparison. Data acquisition and signal processing: SFB streamed ECG and IMU data via BLE to an Android device. ECG was bandpass filtered (Butterworth, 0.5–60 Hz), smoothed (moving average), and an RMS envelope rejected noisy segments. QRS detection used an adaptive thresholding approach (Pan–Tompkins-style): peaks ≥200 ms apart were candidate fiducials; dynamic NoiseLevel and SignalLevel estimates updated after each classification; refractory logic added missed beats and rejected likely T-waves based on slope and timing. HR was computed from R–R intervals and smoothed. HRV-RMSSD was calculated over short 30 s intervals using NeuroKit2; the comparison to metabolic cost used the average RMSSD from the last 2 min of each squat condition. Indirect calorimetry: Metabolic cost (rate) was derived from VO2 and VCO2 (per Brockway), then normalized by body weight and across conditions; steady-state values from the last 2 min were used. Machine learning motion classification: 3-axis accelerometer data were segmented into 0.5 s windows and hand-labeled into six conditions (running, elevated running, walking, elevated walking, standing, squatting). An 80:20 train:test split was used. A spatial CNN with five convolutional layers, max/avg pooling, dropout, and a final fully connected softmax layer (6 outputs) was trained in PyTorch using Adam (lr=0.001), batch size 50, cross-entropy loss, and ReLU activations. The best test-accuracy model was reported. Validation: ECG signal quality was assessed by SNR across activities (standing, squatting, walking, running) and by HR agreement with a commercial chest strap (Polar H10).

- The soft flexible biopatch (SFB) delivered high-quality ECG across activities with SNR >25 dB; representative SNRs: standing 34.32 dB, squatting 33.35 dB, walking 33.17 dB, running 27.92 dB. - HR derived from SFB closely matched a commercial chest strap (Polar H10): R^2 = 0.961 (n = 447), p = 0.002, with 90% of differences within −2.27 to 1.72 bpm. - Normalized HRV-RMSSD (from SFB) strongly and negatively correlated with normalized steady-state metabolic cost (from indirect calorimetry): Pearson R = −0.758, p = 1.2e−7. - Normalized perceived effort (Borg 6–20 scale) also showed a negative correlation with normalized HRV-RMSSD: R = −0.689, p = 6.6e−6. - Machine learning classification of activities using IMU acceleration achieved high performance: training accuracy 97%, best test accuracy 91%, and overall confusion-matrix accuracy 88%. Standing and squatting were classified at 99% accuracy; running (74%) and elevated walking (82%) were lower, indicating difficulty distinguishing gradient conditions. - Mechanical robustness: FEA under 30% strain showed <4% maximum von Mises stress; cyclic stretching yielded minimal resistance drift (<0.1% per cycle; whole-device <1%). Device maintained electrical function up to ~50% elongation; fractured at ~123% extension; Young’s modulus ~500 kPa. - Breathability/adhesion: 3M 9907T substrate exhibited superior MVTR versus PDMS/Ecoflex/micropore and sustained peel forces up to ~1 N with adequate peel energy (dry ~70 J/m^2; wet ~40 J/m^2). - Practicality: Battery supported >9 h continuous operation with ~30 min charge time. - The integrated SFB+AFO system enabled real-time wireless physiological monitoring during squatting, walking, and running, supporting field-deployable exoskeleton personalization.

The study demonstrates that a soft, conformal ECG biopatch can provide robust HR and HRV measurements during dynamic lower-limb activities and that short-term HRV-RMSSD is a strong negative proxy for steady-state metabolic cost during exoskeleton-assisted squatting. This correlation enables rapid, less obtrusive estimation of effort compared to bulky, slow-responding indirect calorimetry, supporting human-in-the-loop personalization of wearable robot parameters in real-world settings. High ECG SNR and strong agreement with a commercial HR strap validate sensing fidelity. The ML-based motion classifier distinguishes activity modes, enabling context-aware estimation and control. Mechanical tests and breathability/adhesion results confirm durability and comfort necessary for prolonged, artifact-resistant use. Together, these findings address the need for fast, portable biofeedback to tune exoskeleton assistance, potentially improving user comfort, safety, and performance outside laboratory environments.

This work introduces and validates an integrated wearable system combining a soft flexible biopatch, an ankle–foot exoskeleton, and machine learning to estimate metabolic cost and physical effort from ECG-derived HRV-RMSSD. The SFB provides high-quality, wireless, real-time physiological data with strong correlation to calorimetry-based metabolic cost, indicating feasibility as a practical alternative to mask-based systems for rapid, personalized exoskeleton tuning. The device is mechanically robust, breathable, and comfortable, and its motion classification enables activity-aware operation. Future work will include larger-cohort rehabilitation studies, real-time optimization of assistance parameters, additional sensing modalities (e.g., muscle activity, temperature, stress), and enhanced motion-state estimation (e.g., combining accelerometer and gyroscope with sensor fusion) to further improve accuracy and generalizability.

- Small pilot sample size (n = 6; ages 20–30) limits generalizability; larger, more diverse cohorts are needed. - HRV-RMSSD was computed from short 30 s windows and compared to steady-state metabolic cost; while practical, ultra-short HRV measures can be sensitive to noise and may not capture all autonomic dynamics. - Activity classification accuracy decreased for gradient conditions (elevated running/walking), suggesting limitations of accelerometer-only input; additional sensors or fusion may be needed. - Validation relied on a tethered AFO and laboratory setup with indirect calorimetry; field performance and untethered exoskeleton integration require further study. - Early transient differences in HR relative to a commercial strap were observed due to smoothing; algorithms may need tuning for rapid transients. - Perceived exertion correlations were moderate and subjective, varying across individuals.