Towards Discovering Digital Twins of Dementia Patients: Matching the Phases of Cognitive Decline.

BACKGROUND: Digital twins in healthcare is an emerging area. Twins are created from data of past patients. The objective is to facilitate data-driven decision support for clinicians for more precise and personalized dementia care. The project has three phases: (1) Twin discovering algorithm development from synthetic data; (2) Algorithm validation with real patient data; (3) Clinical implementation. The current work is related to Phase 1.

METHOD: Past subjects (n=30) showing cognitive decline (based on MMSE scores) mimicking a sample monitored over three years (i.e., 365 x 3 = 1,095 days) were synthetically generated. All past subjects show monotonically decreasing MMSE scores, and have had varying numbers of cognitive tests (between Three and Nine) across three years, unevenly spaced in time. Fig. 1 shows Three examples. Five present subjects (n=5) were also synthetically generated. Present subjects have had only their first two cognitive tests, unevenly spaced in time (i.e., 90 or 180 days-see Fig. 2). Since cognitive decline is traditionally modeled linearly, linear interpolation was used over present and past subjects to fill out missing data between cognitive tests. Post interpolation, an MMSE score is available for every day between a subject's first and last cognitive assessment. Then, for each present subject, the best matching past subject and their phase of cognitive decline-i.e., a digital twin from history that is closest to the present subject-were determined via Mean Absolute Percentage Error (MAPE) calculation between MMSE scores of a present subject and past subjects.

RESULTS: Digital twin segments showing MAPE < 2%, i.e., 100 - MAPE > 98%, were identifiable for all Five present subjects. Table 1 shows the Five best matching digital twin segments for each present subject. Fig 3 graphically shows a matching.

CONCLUSION: Results to discover digital twins via matching phases of cognitive decline are positive. This study looked at only the MMSE score as a one-dimensional metric, for proof of concept. The validity for such one-dimensional arrays, indicate generalizability for multi-dimensional data (i.e., a collection of multiple streams of one-dimensional data) collected over time, for discovering more comprehensive digital twins.