Sensor-Stack Limits on Contactless In-Bed Body Position: A 20-Subject Multimodal Radar + Thermal LOSO Characterization

Technical Summary: Sensor-Stack Limits on Contactless In-Bed Body Position

Problem Statement
Contactless in-bed body-position inference is often limited by the sensor representation available to the classifier rather than the choice of classification algorithm itself. While prone position monitoring is clinically motivated by its association with Sudden Unexpected Death in Epilepsy (SUDEP), existing contactless approaches face significant hurdles. Prior FMCW radar work often relies on raw RF data rather than on-device Constant False Alarm Rate (CFAR) detections, and low-resolution thermal arrays struggle to resolve face-versus-back-of-head signatures. This paper investigates the performance limits of a specific, commercially available sensor stack: a bedside 60 GHz FMCW radar with host-visible CFAR point-cloud output, fused with a low-resolution (24 × 32 nominal) thermal array. The core question is whether this specific hardware configuration, under a leave-one-subject-out (LOSO) protocol with zero labeled data from the test subject, can reliably distinguish in-bed postures, particularly the prone position.

Methodology
The study utilized a cohort of 20 subjects (ages 5–68, including 13 minors) across 8 residences, representing a convenience sample rather than a population-level clinical trial. Data collection involved 273 supervised in-bed posture holds totaling approximately 149 minutes, alongside a separate bed-presence audit timeline of over 35,000 seconds.

• Hardware:
  • Radar: Texas Instruments IWR6843AOP (60 GHz FMCW) mounted at the head of the bed. It operates at ~10 Hz with on-device CFAR detection (Range: CFAR-CASO; Doppler: CFAR-CA), outputting sparse point clouds (range, velocity, SNR, azimuth, elevation) rather than raw RF or range-FFT data.
  • Thermal: Melexis MLX90640-D110 thermal array (nominally 24 × 32). Due to unmerged subpages in the cached data, the effective resolution was 24 × 16 pixels.
  • Reference: Ground truth labels were derived from operator tags (for supine/prone) and a chest-mounted IMU (for left/right lateral distinction).
• Protocol & Validation:
  • Tasks: Binary bed-presence (in-bed vs. out-of-bed) and four-class posture classification (supine, left-lateral, right-lateral, prone).
  • Validation: Leave-One-Subject-Out (LOSO) cross-validation. For each fold, the model was trained on 19 subjects and tested on the held-out subject.
  • Features: Radar features included point counts, range/velocity/SNR statistics, and spatial centroids. Thermal features included hot-region geometry and lateral centroids. Cross-modal features normalized thermal centroids against radar centroids.
  • Classifiers: Primary analysis used L2-regularized logistic regression. Additional experiments included Gradient Boosted Trees, Random Forests, and a stacked ensemble of these models.

Key Contributions

1. Sensor-Stack Characterization: A rigorous LOSO evaluation of a specific radar + thermal stack on a 20-subject cohort, distinguishing between "bed presence" (a strong result) and "posture classification" (a limited result).
2. Modality-Level Error Analysis: Identification that thermal sensing is critical for resolving left-vs-right lateral discrimination (reducing swap rates from ~35–42% in radar-only to ~8%), while the radar CFAR representation fails to reliably expose the expected supine-vs-prone breathing cue on a per-hold basis.
3. Hardware Limitation Identification: The paper demonstrates that the "prone gap" is a sensor-system data-representation issue, not a classifier tuning issue. The CFAR point-cloud output obscures the breathing amplitude signal necessary to distinguish prone from supine, which is only visible in aggregate class distributions (Cohen's d = 0.61) but not in 8.4% of individual holds.

Results

• Bed Presence: The fused radar + thermal model achieved a median LOSO balanced accuracy of 0.871 (IQR 0.760–0.922). True-in recall was 95.2%, and false-exit rates were 4.8%.
• Four-Class Posture: The best-performing pipeline (a full-feature stacked ensemble) achieved an aggregate balanced accuracy of 0.674.
  • Per-class Recall: Supine (0.80), Left-lateral (0.69), Right-lateral (0.70), Prone (0.50).
  • Modality Ablation: Thermal-only achieved 0.591 accuracy, while radar-only achieved 0.440. Direct fusion offered a marginal lift (+0.011), while the stacked ensemble provided a larger lift (+0.083) attributed to the broader feature pool rather than synergistic interaction.
• The Prone Gap: Prone recall remained at 0.50 across all configurations, including the stacked ensemble. Prone precision was 0.41, meaning 59% of prone alarms were false positives. The radar breathing cue, expected to differentiate supine from prone, was unreliable per hold (only 8.4% of holds showed a clean breathing peak in the CFAR stream).
• Body-Size Bias: A positive correlation (r = +0.41) was observed between subject height and balanced accuracy, with the model performing worst on the smallest (pediatric) subjects, though this did not reach statistical significance (p = 0.075) at n=20.

Significance and Claims
The paper explicitly frames itself as a sensor-stack characterization rather than a "best system" or deployment-ready solution.

• Not Deployable for Prone Detection: The authors state clearly that the current results are "not deployable prone detection." The prone recall of 0.50 is at the chance rate for a binary decision, and the high false-positive rate makes it unsuitable for clinical alarm systems.
• Limit of Current Hardware: The study concludes that the bottleneck is the CFAR representation. The on-device CFAR processing discards the raw range-FFT data required to reliably extract the chest-motion breathing amplitude needed to distinguish supine from prone on a per-hold basis.
• Next Steps: The paper does not claim that classifier tuning will solve the problem. Instead, it identifies raw range-FFT access (via a hardware/firmware change to expose per-bin breathing cues) as the necessary next experiment.
• Scope Boundaries: The authors emphasize that these are characterization figures on a convenience cohort under operator-directed protocols, not overnight natural sleep data. No claims are made regarding seizure detection, overnight performance, or generalizability beyond the tested hardware stack.

In summary, the paper demonstrates that while contactless radar and thermal fusion can robustly detect if someone is in bed, the current generation of host-visible CFAR radar data combined with low-resolution thermal arrays cannot reliably distinguish prone from supine positions without access to raw RF range-FFT data.