Session Item

The COVID pandemic accelerated the integration of virtual care and patient monitoring in cancer. In the age of big data, alternative and continuous patient monitoring holds tremendous promise. Within the context of a Phase I/II trial evaluating the combination of Durvalumab, Tremelimumab and SBRT for oligometastatic head and neck cancers (NCT03283605), we explore the role of patient biometry to detect alterations in circadian rhythm and their association with patient disease state and prognosis.

The first 26 (22 males and 4 females) subjects who accepted to wear a Fitbit Alta HR during the clinical trial were analysed. Patients were imaged and measured using RECIST 1.1 criteria every 3 months. For each hour, we used minute-to-minute calories, metabolic equivalent of task (MET), activity intensity, number of steps, and heart rate to compute the mean, minimum, 10th to 90th percentile, maximum, as well as the 1st, 2nd, and 3rd time-derivative of these variables. Time in hour was encoded as vectors (T_h): 

where h is the time in hour and (e_i ) are the standard unit vectors in 24th dimension. This formulation was derived to ensure that for all n∈{1,2,3,…,11}, the distance ‖(T_(h+n))-(T_h)‖ in Euclidean 24-space increases as n increases but remains the same for all h∈{0,1,2,…,23}. We subsequently performed dimension reduction analyses of biometric data using uniform manifold approximation and projection (UMAP) and principal component analysis (PCA).

PCA and k-means clustering revealed 5 clusters of activity profiles, one of which was associated with low physical activity, high metabolic rates, and higher chance of death and cancer progression. Multivariate logistic regression showed that 7-day maximum total steps, minimum sedentary minutes, maximum lightly active distance, and maximum very active distance were independently associated with lower likelihood of progression in 2 months (p<0.0001). Similarly, maximum total steps and total daily distance predicted lower likelihood of death in 1 year (p<0.0001). The multivariate model had specificity and sensitivity of 100% and 94%, respectively, for predicting progression and specificity and sensitivity of 98% and 85% for predicting death in 1 year. Unsupervised dimension reduction revealed a circular cluster consisting of 24 temporally ordered subclusters containing hour-to-hour biometric data (Figure 1). Activity level was highest in the center but lower in the periphery and lowest in subclusters outside the circular cluster. At the same time, clusters associated with low physical activity level were associated with patient death and with cancer progression (Figure 2). 

In this study, we demonstrated that multivariate continuous biometric recording can be used to predict survival and progression in metastatic head and neck cancer. We are also the first to employ unsupervised dimension reduction of biometric data to assess alterations in the circadian rhythm as a tool detect disease progression.