Project: Noninvasive core body temperature sensor for wearable sleep monitoring devices
The objective of the project was to create a proof of concept that it is feasible to estimate the core body temperature (CBT) from the wearable sensors using the skin temperature and heat flux sensors. In the course of the project, a prototype of a standalone wearable sensor device was developed. The sensor is lightweight and allows for non-invasive continuous estimation of CBT with an average error smaller than 0.4 centigrade, using the measurements of the skin temperature and heat flux.
The core body temperature and its cyclical fluctuations are important indicators of proper functioning of the healthy human organism, in particular the metabolic rate. Many cyclical states of a healthy human, such as sleep, ovulation etc., manifest themselves via characteristic body core temperature trajectories. Departures from the regular circadian cycle can be used as important diagnostic tool in conditions such as fever, insomnia, elevated stress, jet lag etc.
Despite its high diagnostic value and potential, the core body temperature measurement cannot be currently done in a non-invasive fashion. Popular temperature measurement techniques merely measure the skin temperature at various locations - and the skin temperature is known to vary drastically from the core body temperature. It also depends on transpiration, insulation, etc. Existing techniques that attempt to measure the core body temperature (in the ear canal, rectal, radio pill), are excessively invasive for everyday, non-clinical Applications.
During the project, we compared several approaches to estimating CBT using heat flux and skin temperature using ingestible pill sensors as a reference. The following approaches were investigated:first-principle based models, machine learning-based models (ML), and hybrids thereof. In the first-principle-based models, the closed-form explicit mathematical model linked the sensor measurement with the core body temperature with one free parameter (body conductance). Our attempt to find one parameter value to yield satisfactory CBT prediction for all individuals was unsuccessful. Consequently, the free parameter was estimated for individuals from the first hour of their recording - we refer to this procedure as calibration. The necessity to calibrate the model for each individual is a drawback of the first-principle-based approach. In the machine learning approach, the transfer function from sensor measurements to CBT was learned entirely from training data. In the hybrid approach, the free parameter of the closed-form model was learned from training data. In our experiments with the ML approach, the models and their parameters were trained and tested on disjoint user sets (i.e. no data of the test user was used to train the models). The benefit of the ML approach was the fact that, as opposed to the first-principle-based model, it did not require a calibration procedure for each individual. A sample result of the CBT prediction using the first-principle-based, physical models and the ML approach is shown in Figure 1.
In conclusion, the combination of a miniaturized temperature and heat flux sensors that can be integrated into a wearable device, and an advanced algorithm allowed us to create the first non-invasive and continuous CBT measurements at different body positions under temporarily changing external conditions using a wearable device. This development opens new paths towards personalized diagnostic systems. We are currently working on the application of the developped technology to perform early detection of Alzheimer’s disease.
«Human body is nothing but a highly complex machinery – and we expect to gain new insights and predictive capacities by applying the same machine learning approach to biosignals that proved to be successful in treating signals from other physical systems.»
Dr. Krzysztof Kryszzuk