(I-335) Privacy disrupting individual identification from temperature records

Introduction::
In recent years, the development of a wide array of physiological sensors has enabled a broad range of applications, from continuous glucose monitoring to wearable activity monitors. One constant in the development of new applications for physiological data is the increasing sample rates and timescales of the data collected. Although increasing data complexity engenders better outcomes for users, attendant privacy concerns are also realized. Collection, transmission, and storage of physiological measures exposes individual privacy to potential misuse by bad actors [1]. This work will demonstrate individual classification from skin temperature monitoring data, indicating privacy risks.




Continuous monitoring of physiological parameters, such as blood pressure, blood glucose, and temperature have all shown significant potential in enabling more efficient diagnosis and treatment of many conditions. One commonality between many of the novel classifiers and symptom monitoring systems across different physiological measures and diseases is the longitudinal nature of the collected data. Longitudinal data presents a twofold privacy concern. First, it is often possible to differentiate individuals from longitudinal measurements [2,3]. Second, by its very nature, longitudinal data provides information about the activities of daily living of a user, and when these activities occur. In combination, these two concerns present a significant challenge to widespread adoption of wellness monitoring technologies, for both individuals and institutions. As a motivating example, this work demonstrates the classification of individuals from a moderately sized population (n=40) based solely upon wearable skin temperature monitoring data.




Materials and Methods::
Skin temperature is of increased interest for monitoring due to rise of COVID-19 as well as the increasing cost of healthcare acquired infections in general. While individual temperature measurements may not present an immediate privacy concern, longer recordings do. Skin temperature is correlated with varying daily activities and sensitive physiological states, such as acute stress [4], and ovulation [5]. Exposure of this information is not desirable to patients. To motivate future theoretical and practical privacy developments, classification of individual subjects from continuously monitored temperature data was undertaken.




The dataset used in this work was drawn from an initial publication on the TemPredict trial [6]. 40 subjects from the group were used, with data segmented arbitrarily into 24 hour increments. Samples were bootstraped, with 1000 training segments drawn from the first 25 days for each subject and 100 validation segments drawn from the latter 40 days. A random sampling of subject data is shown in Figure 1.




An LSTM neural network was trained to classify the 40 individuals into 40 classes. The privacy implications of the temperature dataset were evaluated via the trained model accuracy in correctly classifying an individual in the pool from a single day temperature segment and from a group (the validation set) of temperature segments.




Results, Conclusions, and Discussions::
The trained model produced an accuracy of 37.5% in individual test segments, but accuracy improved to 87.5% when the mode of each subjects validation set (all 100 samples) was used. These results were compared to encrypted sets of the same data, with markedly different results.






The temperature data set was encrypted utilizing the method described in prior work [7]. When the model was trained with this encrypted data, subject identification was not possible, with an individual trial accuracy of 2.3% and a by subject (mode) accuracy of 2.5%, in line with a random guessing of the subject’s identity out of the 40 subject pool. This work demonstrates the significant privacy concerns attendant in the collection, storage and analysis of temperature data. The results here show the necessity of stringent privacy preserving methods for the evolution of health data collection and use.




A confounding factor in the identification problem described here is the presence of a COVID-19 infection self reported by study participants beginning on day 45 of the 65 days recorded. This puts the infection window entirely within the validation segments. Model performance would be expected to improve if conditions (infection status) were to remain constant across both training and test periods.






This work relies on the assumption that a bad actor has previously gained access to labelled temperature data for individual subjects. Heath record breaches are a significant and ongoing concern, over 240 million healthcare records were improperly disclosed between 2005 and 2019 [8], lending bad actors the potential ability to train identification classifiers. In recent years, the proliferation of consumer wearable devices has created another avenue for the generation of large libraries of individual temperature records. Even unlabeled or partially labelled temperature data may present a privacy concern in some scenarios, with unsupervised learning of classifiers an ongoing research area for the authors. The 40 subject pool is another limitation of this work, as any data breach would likely contain far more subject records. Future work will seek to investigate the privacy concerns surrounding larger datasets, and methods to secure physiological data as it is collected and analyzed.








Acknowledgements (Optional): :

References (Optional): :
1] Chikwetu, Lucy, et al. "Does deidentification of data from wearable devices give us a false sense of security? A systematic review." The Lancet Digital Health (2023).




[2] Li, Ning, et al. "The Identification of ECG Signals Using Wavelet Transform and WOA-PNN." Sensors 22.12 (2022): 4343.




[3] Bidgoly, Amir Jalaly, Hamed Jalaly Bidgoly, and Zeynab Arezoumand. "A survey on methods and challenges in EEG based authentication." Computers & Security 93 (2020): 101788.




[4] Herborn, Katherine A., et al. "Skin temperature reveals the intensity of acute stress." Physiology & behavior 152 (2015): 225-230.




[5] Zhu, Tracy Y., et al. "The accuracy of wrist skin temperature in detecting ovulation compared to basal body temperature: prospective comparative diagnostic accuracy study." Journal of Medical Internet Research 23.6 (2021): e20710.




[6] Smarr, Benjamin L., et al. "Feasibility of continuous fever monitoring using wearable devices." Scientific reports 10.1 (2020): 21640.




[7] Kawase, Hiroaki, et al. “Encrypted Classification for Prevention of Adversarial Perturbation and Individual Identification in Health-Monitoring.” IEEE/ASME International Conference on Advanced Intelligent Mechatronics. (2023) Accepted




[8] Seh, Adil Hussain, et al. "Healthcare data breaches: insights and implications." Healthcare. Vol. 8. No. 2. MDPI, 2020.