Real-time, personalized medicine through wearable sensors and dynamic predictive modeling: a new paradigm for clinical medicine

Abstract

Accurately predicting the onset and course of a disease in an individual is a major unmet challenge in medicine due to the complex and dynamic nature of disease progression. Continuous data from wearable technologies and biomarker data with a fine time resolution provide a unique opportunity to learn more about disease evolution and to usher in a new era of personalized and real-time medicine. Herein, we propose the potential of real-time, continuously measured physiological data as a noninvasive biomarker approach for detecting disease transitions, using allogeneic hematopoietic stem cell transplant (HCT) patient care as an example. Additionally, we review a recent computational technique, the landscape dynamic network biomarker method, that uses biomarker data to identify transition states in disease progression and explore how to use it with both biomarker and physiological data for earlier detection of graft-versus-host disease specifically. Throughout, we argue that increased collaboration across multiple fields is essential to realizing the full potential of wearable and biomarker data in a new paradigm of personalized and real-time medicine.

Figure 1:. Schematic of the nonlinear evolution of disease.

The black line represents the linear model of disease whereby disease progression happens uniformly and in a constant manner to the disease state. The red curve represents the nonlinear evolution of disease whereby a stable pre-disease state undergoes a rapid transition (transition state) to a disease state. Intervention at the transition state may result in better outcomes.

Figure 2:. Examples of current wearable technologies, based on the data they collect.

Various physiological data streams available from some currently available wearable technologies. The list presented is not meant to be exhaustive, and there may be other important wearable technologies on the market.

Figure 3:. Outline of the steps to the landscape DNB method.

The l-DNB ethod takes as input a gene-expression data set. Step 1 calculates Pearson (PCC_m) and differential Pearson correlation coefficients (sPCC_m) to construct a single sample network (SSN) around gene x. The SSN is separated into first-order neighbors of gene x, N_xd (with n_xd genes), and second-order neighbors, M_xd (with m_xd genes). Next, based on the SSN found in Step 1, Step 2 computes local DNB values for each gene. sPCC_in is the average absolute sPCC_m value between gene x and its first-order neighbors. sPCC_out is proportional to the average absolute value of sPCC_m between gene x and its first-order neighbors and the second-order neighbors. sED_in is the average deviation in expression of gene x and all its first-order neighbors. I_s(x) is the local DNB score of gene x. Finally, Step 3 ranks all local DNB scores and takes the average of the top k values to compute a global DNB score, I_DNB.

Figure 4:. An example application of the l-DNB method to HCT patients using biomarker and physiological data.

Wearable technologies continuously record physiological data, and blood samples are processed for biomarker and genetic data. Next, daily signatures are extracted from the continuous physiological data to align the time points of the biomarker and physiological data sets. Finally, the l-DNB method takes as input both data streams to compute global DNB scores to find tipping points for GVHD in the first 100 days after treatment, which represents the high-risk period for aGVHD onset.