TL;DR
This paper introduces a Kalman filter (KF) model that estimates human core body temperature (Tcore) using only heart rate (HR) as input. This approach achieves accuracy comparable to more complex models but with significantly less data, offering a promising avenue for real-time thermal injury risk assessment in various demanding environments.
What Problem Does This Paper Address?
Accurately monitoring human core body temperature (Tcore) is crucial for preventing heat-related illnesses, especially in environments where individuals wear protective gear or exert themselves heavily. Traditional methods for measuring Tcore, like rectal thermistors or esophageal probes, are impractical for continuous, non-invasive monitoring in real-world scenarios. Existing wearable sensors can capture physiological data like heart rate and skin temperature, but these are influenced by numerous factors (environment, clothing, hydration, etc.), making their direct correlation to Tcore tenuous. More sophisticated models exist, but they often require a multitude of inputs (metabolic rate, clothing, individual characteristics), limiting their practical application in dynamic, real-time situations. This paper aims to bridge this gap by developing a simpler, yet effective, method for estimating Tcore from readily available wearable sensor data.
Methods
The core of this paper is the development and evaluation of a discrete Kalman filter (KF) model. The KF is an iterative state estimation algorithm that uses a system's dynamics and noisy measurements to estimate its current state. In this context, the "state" is the human's core body temperature (Tcore), and the "noisy measurement" is heart rate (HR).
The KF model development involved:
- System Dynamics (G): The authors assumed that Tcore at the current time step is largely predicted by Tcore at the previous time step, with minimal change due to thermal inertia. This was modeled using an identity matrix for G.
- Measurement Model (H): This matrix maps the Tcore state into the HR measurement space. It was derived from training data by establishing a linear relationship between HR and Tcore, specifically defining a line between points representing low HR/low Tcore and high HR/high Tcore.
- Noise Components (u, v): The model accounts for noise in the prediction (u) and measurement (v). The prediction noise variance (Q) was derived from the discrete probability distribution of Tcore changes, and the measurement noise variance (R) was determined from the standard deviation of HR measurements.
The KF algorithm iteratively updates its Tcore estimate using a time update (predicting the next state based on current state and dynamics) and a measurement update (refining the prediction using the actual HR measurement).
Datasets
The paper utilized data collected from several sources:
- Training Data: Data from three field studies involving soldiers and athletes, comprising over 20,000 HR-Tcore pairs. This data was used to estimate the parameters of the KF model (G, H, Q, R).
- Evaluation Data: Data from a separate set of 25 physically-active subjects (N=25) who participated in various laboratory and field studies. This included over 25,000 HR-Tcore pairs from diverse environmental conditions (20°C to 40°C), work rates, and durations (2-8 hours). This dataset was used to evaluate the KF model's performance and compare it to the more complex ICDA model.
Key Insights and Ideas
- Leveraging Physiology for Simplicity: The authors successfully framed the problem as a state estimation task, using physiological knowledge (e.g., the relationship between HR, work, and heat production/dissipation) to simplify the model. This insight allowed them to use a single input (HR) where other models require many.
- Kalman Filter's Efficacy: The KF proved to be a powerful tool for inferring a latent variable (Tcore) from noisy, indirect measurements (HR). Its iterative nature allows for continuous estimation in dynamic environments.
- HR as a Proxy for Tcore: The study demonstrates that while HR is influenced by many factors, its relationship with Tcore becomes more direct and predictive when thermoregulation is stressed (i.e., during significant heat gain or loss).
- Potential for Real-Time Risk Assessment: The model's ability to estimate Tcore in near real-time from a single wearable sensor opens doors for proactive interventions to prevent heat illness, particularly in high-risk occupations like military personnel, firefighters, and athletes.
Key Results
- Overall Accuracy: The KF model achieved a root mean square error (RMSE) of 0.30 ± 0.13 °C from observed Tcore.
- Agreement: Over 85% of the KF model's Tcore estimates were within ± 0.5 °C of the observed values.
- Comparison to ICDA: In 7 out of 8 experimental conditions, the KF model provided a closer estimation of Tcore with lower variance compared to the more complex Initial Capability Decision Aid (ICDA) model.
- Individual Case Study: For a specific military cadet who experienced heat exhaustion, the KF model showed an average RMSE of 0.17°C with 94.5% of estimates within ± 0.5°C.
Limitations and Future Work
- Model Simplifications: The authors acknowledge that the KF model has simplifications:
- The prediction function (G) assumes Tcore doesn't change significantly between time steps, which might not hold during rapid physiological shifts.
- The mapping function (H) is linear, which may not accurately represent the relationship at extreme Tcore values (above 39.5°C or below 36.5°C).
- Variances for prediction and mapping functions are assumed constant, which might not be true in all scenarios.
- Extreme Temperatures: The current KF model may not correctly model extreme core temperatures above 39.5°C.
- Future Work: The authors suggest that with a few refinements, this model could offer a promising means of real-time thermal injury risk assessment. They also highlight that other human health state monitoring parameters (hydration, metabolic state, fatigue, cognitive state) are ripe for similar AI-driven estimation techniques. The paper also emphasizes the need for better data sharing practices between physiological monitoring and AI communities to advance this research.
Summary updated