Edge-Aided Radar-Based Exercise Form Classification Using Lightweight Ensemble Learning for Personalized Healthcare.

The RehabRadar system achieved 91.5% accuracy and 92.0% macro-F1 on average across all personalized models using ten-fold cross-validation.

• Evaluation was conducted in a proof-of-concept study with 10 healthy adults performing eight clinically curated exercises.
• Ten-fold cross-validation was performed per-participant.
• The system classified exercise execution as correct or one of two common error types.
• These results outperformed single-branch baselines.

The hybrid ensemble combining MobileNetV3 for micro-Doppler images and CNN-LSTM for range-Doppler sequences with late fusion outperformed single-branch baselines.

• Parallel inference was performed using MobileNetV3 processing micro-Doppler spectrograms and CNN-LSTM processing range-Doppler sequences.
• The two branches were combined via late fusion.
• The complementary nature of micro-Doppler spectrograms and range-Doppler sequences was exploited for improved classification.
• Single-branch baselines were used as the comparison benchmark.

End-to-end latency averaged 1.5 seconds, enabling rep-level feedback on the edge device.

• All processing occurred on a Raspberry Pi 4B edge device.
• The 1.5 s latency supports real-time, rep-level exercise feedback.
• Models were optimized for the Raspberry Pi 4B via magnitude-based pruning and post-training INT8 quantization.
• The final model footprint was 21 MB.

The system uses a TI IWR1642 mmWave FMCW radar with a signal pipeline that isolates the subject via range-gating and FIR high-pass filtering.

• The radar sensor is the TI IWR1642 mmWave FMCW radar.
• Signal processing includes range-gating for subject isolation and FIR high-pass filtering.
• The pipeline extracts complementary micro-Doppler spectrograms and range-Doppler sequences as features.
• The system is designed as a non-contact monitoring solution addressing privacy and robustness concerns of cameras and wearables.

The system preserves privacy by processing all data on-device, discarding raw radar data after feedback, and transmitting only minimal encrypted metrics externally.

• All processing occurs on-device on the Raspberry Pi 4B.
• Raw radar data is discarded after feedback is generated.
• Only minimal encrypted metrics are transmitted externally.
• The design supports decentralized learning in IoT healthcare settings.

The authors acknowledge that clinical efficacy, generalizability beyond the studied cohort, broader exercise taxonomies, and patient populations are outside the present scope.

• The study cohort consisted of 10 healthy adults, limiting generalizability to patient populations.
• Only eight clinically curated exercises were evaluated.
• The work is described as a 'proof-of-concept study.'
• Multi-site clinical validation is identified as planned future work to establish clinical efficacy and generalizability.