Technical Summary: A Real-Time Hot Flash Detection System for Noisy Free-Living Conditions

Problem Statement

Hot flashes (HFs) are the most prevalent and disruptive symptom of the menopausal transition, affecting up to 80% of individuals and significantly impacting daily life and sleep. Currently, objective tracking of HFs is largely restricted to research settings due to the limitations of existing measurement technologies. The "gold standard" for HF measurement involves ambulatory monitoring of sternal skin conductance (SC) using the UFI Biolog system. However, this device is bulky, requires wired electrodes, necessitates expert operation, and lacks consumer-friendly interfaces, making it unsuitable for large-scale, long-term, or free-living deployment.

While recent advancements have introduced machine learning (ML) algorithms for HF classification, these have primarily relied on research-grade equipment in controlled environments and often utilized shuffled data in cross-validation, limiting their ability to generalize to unseen populations. Furthermore, existing systems often depend solely on SC signals and offline processing, failing to account for the variability in HF physiology across different bio-behavioral states (e.g., sleep vs. wake) and environmental contexts. There is a critical need for a cost-effective, real-time system capable of detecting HFs using commercially available, multi-sensor wrist wearables in noisy, free-living conditions.

Methodology

Study Design and Data Collection

The study recruited eleven menopausal women (aged 40–65) reporting ≥5 HFs per 24-hour period. Participants underwent a 48-hour at-home data collection protocol. The experimental setup involved four distinct devices to capture multimodal data:

1. Gold Standard: UFI BiLog HF monitor (3991x/1-SCL) with Ag/AgCl electrodes on the sternum to record SC at 1 Hz.
2. Primary Sensor: Empatica E4 wrist-worn device (placed on both wrists) collecting raw data from photoplethysmography (PPG), accelerometry (ACC), SC, and skin temperature at high resolutions (up to 64 Hz).
3. Behavioral Context: Fitbit Charge 4 (FC4) on the left wrist for sleep/wake classification and step counting.
4. Environmental Context: iButton DS1923 logger for ambient temperature and humidity.

Participants pressed event-marker buttons on the E4 devices upon perceiving an HF. Expert annotations, based on manual evaluation of sternal SC fluctuations (>1.5-µmho rises within 30 seconds) from the BioLog device, served as the ground truth for HF onset.

System Architecture and Feature Engineering

The authors developed a real-time ML HF-detection system using data from the Empatica E4. Unlike previous iterations that required processing entire waveforms or offline analysis, this system was optimized for real-time operation with minimized computational load.

Key System Modifications:

• Real-Time Processing: The system computes features every 15 seconds using a windowing approach.
• Feature Sets:
  • SC Features: Differential area under the curve (AUC) and derivative (slope) calculated between past and future windows (specifically focusing on the "rising edge" to predict onset before subjective perception).
  • Temperature Features: Average temperature differentials between past/future contexts and the current window.
  • PPG/Heart Rate Features: Differential heart rate between the current window and past/future contexts.
  • Motion Features: Absolute maximum displacement (AMD) in x, y, and z dimensions, including differential calculations between axes.
• Classifier: A decision tree classifier generates binary (HF/non-HF) predictions for each 15-second interval.
• Post-Processing: Predictions are aggregated to identify discrete HF events. A 15-minute refractory period (60 intervals) is enforced to prevent multiple detections of a single physiological episode.

Experimental Tasks

The study evaluated the system through five specific tasks:

1. Real-Time Performance: Determined the optimal past/future context window sizes (tested 1s to 240s) using a leave-one-subject-out (LOSO) cross-validation.
2. Computational Efficiency: Evaluated the impact of signal down-sampling (factors of 1 to 32) and algorithm computation intervals (15s to 90s) on accuracy.
3. Noise Robustness: Tested system performance against synthetic noise (signal drops and spikes) ranging from 0.5% to 50%.
4. Noise Augmentation Training: Retrained the model on synthetically augmented data to determine if performance could be recovered under noisy conditions.
5. Free-Living Validation: Validated the final model across different bio-behavioral (sleep vs. wake, active vs. inactive) and environmental (temperature, humidity, time of day) conditions.

Key Contributions

1. Real-Time, Multi-Sensor Detection: The paper presents a novel ML system that integrates SC, temperature, pulse rate, and motion data from a commercial wrist wearable (Empatica E4) to detect HFs in real-time, moving beyond the single-signal, offline approaches of prior research.
2. Optimization for Free-Living Conditions: The system is explicitly designed and validated for "noisy" free-living environments, addressing signal loss, motion artifacts, and varying physiological states (sleep vs. wake).
3. Computational Efficiency: The authors provide specific algorithm parameters (e.g., 15-second computation intervals, 4x down-sampling) that minimize computational load while maintaining high accuracy, making the system feasible for deployment on consumer-grade hardware.
4. Robustness to Noise: The study demonstrates that the system maintains >90% accuracy even with up to 15% signal drop when trained with synthetically augmented noise, a critical finding for real-world applicability where sensor contact is often imperfect.
5. Generalizability Validation: Using a held-out testing dataset from seven women (distinct from the four women in the training set), the study verifies that the model does not merely memorize training data but possesses genuine predictive power.

Results

• Accuracy: The system achieved >90% accuracy in real-time detection of HFs under free-living conditions.
• Optimal Parameters:
  • Context: 60 seconds of past context (left context) was identified as optimal for real-time operation, balancing resolution and performance.
  • Efficiency: A computation interval of 15 seconds and a down-sampling factor of 4 (resulting in 1 Hz sampling for most sensors) provided the best trade-off between delay and accuracy.
• Noise Handling:
  • Without retraining, the system maintained >90% accuracy with signal drops up to 15%.
  • Retraining with synthetic noise (0.25% spike and 0.25% drop) improved accuracy to 99.16% for 15% signal drops, representing a 60% relative reduction in error compared to the non-augmented model.
• Condition Robustness: The system demonstrated consistent performance (>90% accuracy) across various conditions, including sleep vs. wake states, periods of activity vs. inactivity, different times of day, and varying environmental temperatures and humidity levels.

Significance and Claims

The authors claim that this work represents a significant step toward the feasibility of measuring HFs in daily life using commercial-grade wrist sensors. By demonstrating that a multi-sensor ML system can operate in real-time with high accuracy despite the "noise" of free-living conditions, the system offers a pathway to:

• Objective Tracking: Providing individuals and clinicians with reliable, passive, and longitudinal data on HF patterns and characteristics, moving away from subjective self-reporting.
• Management and Intervention: Enabling better management of this disruptive symptom and potentially supporting "just-in-time" interventions by predicting HF onset before subjective perception.
• Scalability: Offering a cost-effective alternative to the bulky, expensive, and expert-dependent gold-standard devices, thereby facilitating large-scale studies and broader clinical adoption.

The paper modestly acknowledges limitations, noting the modest sample size (11 participants), the exclusion of individuals with BMI ≥35 kg/m², and the lack of analysis regarding skin tone distribution. The authors emphasize that further large-scale validation on diverse consumer wearables is necessary to ensure the system's reliability across all demographic groups and to address potential biases related to PPG accuracy in different skin tones or body compositions.