Human Digital Twin for Personalized Elderly Type 2 Diabetes Management

Abstract

Managing Elderly type 2 diabetes (E-T2D) is challenging due to geriatric conditions (e.g., co-morbidity, multiple drug intake, etc.), and personalization becomes paramount for precision medicine. This paper presents a human digital twin (HDT) framework to manage E-T2D that exploits various patient-specific data and builds a suite of models exploiting the data for prediction and management to personalize diabetes treatment in E-T2D patients. These models include mathematical and deep-learning ones that capture different patient aspects. Consequently, the HDT virtualizes the patient from different viewpoints using an HDT that mimics the patient and has interfaces to update the virtual models simultaneously from measurements. Using these models the HDT obtains deeper insights about the patient. Further, an adaptive patient model fusing this information and a learning-based model predictive control (LB-MPC) algorithm are proposed. The geriatric conditions are captured as model parameters and constraints while solving the LB-MPC to personalize the insulin infusion for E-T2D management. The HDT is deployed on and illustrated with 15 patients using clinical trials and simulations. Our results show that HDT helps improve the time-in-range from 3-75% to 86-97% and reduces insulin infusion by 14-29%.

Keywords: Elderly type 2 diabetes (E-T2D); digital twin (DT); human digital twin (HDT); internet of medical things (IoMT); learning-based model predictive control (LB-MPC); personalization; precision medicine.

Figure 1

HDT architecture.

Figure 2

Precision Treatment Methods Through HDT—The HDT framework leads to new data, novel models that create additional knowledge, and improved decision support through new degrees of freedom.

Figure 3

IoMT Architecture—Data transfer using IoT sensor devices from patients’ (left), mounted devices to edge nodes (center) to the HDT architecture and finally to medical personnel and caretakers using a mobile app.

Figure 4

Simulated web page view of the patient’s data.

Figure 5

Real-time patient data collection system using temperature sensor, pulse oximeter, CGM, sensor tag, and Edge device to display the data.

Figure 6

Structure of the LSTM Network for Blood Glucose Level Predictions.

Figure 7

BGL prediction using time segmentation.

Figure 8

LSTM prediction result for P1.

Figure 9

LSTM prediction result for P2.

Figure 10

Structured Time-Series Analysis—The y-axis in the upper panel is BGL (mg/dL), and the time interval is 15 min.

Figure 11

Time-Series Analysis: Motif Discovery between days of a P1. (Top) Dependent variable (BGL) versus time whereas the bottom show independent (controllable) behaviors (insulin) and CHO in food.

Figure 12

Time-Series Analysis: Motif Discovery between days of P1—The upper panel shows a dependent variable (BGL) versus time whereas the lower panel shows independent (controllable) behaviors (insulin) and the dependent variable CHO.

Figure 13

Model predictions for P1–P5—The time step is 15 min, and the BGL is predicted for one step ahead.

Figure 14

Testing Protocol.

Figure 15

Patient data.

Figure 16

Snap-shot of one day data collected from a patient.

Figure 17

Vital Sign Monitoring Using Mobile App—Top figure shows the body temperature data, the middle shows the heart rate data, and the bottom figure shows SpO${}_2$ data.

Figure 18

ACF and PACF plot for $P_1$.

Figure 19

Time-Series Analysis: Motif Discovery between days 9 and 10 of P1.

Figure 20

Motif Discovery between patient $P_1\&P_4$: Motif shows days with similar patterns. Patient $P_4$ has lower BGL than $P_1$ due to active lifestyle and relatively lower co-morbid conditions.

Figure 21

Time-Series Analysis: Motif Discovery between $P_4\&P_{13}$—$P_4$ has a lower magnitude of BGL profile even with high average CHO and less insulin infusion.

Figure 22

Explanations for three randomly selected samples of patient P1. (a): An actual sample in the hyperglycemic range detected with high probability through XAI as basal value of 0.02, (b): sample 2 that is in Euglycemic range being detected with high probability with Insulin actual values, (c): Sample in Euglycemic range detected with probability of 0.64 through basal and bolus.

Figure 23

Explanations for three randomly selected samples of patient P1. (a): Sample showing hyper condition and explained through basal and bolus values even though CHO is low, (b): A hyper glycemic sample explained through basal values even when the food CHO is low, and (c): random sample detected to be hyper glycemic based on basal and bolus values.

Figure 24

Insulin recommendation and BGL for $P_1$—Top panel shows the comparison between uncontrolled BGL variations during the clinical trial and controlled BGL using HDT framework. The middle plot shows insulin infusion during the clinical trial, whereas insulin recommendation from HDT is shown in the bottom plot.

Figure 25

Insulin recommendation and BGL for $P_2$—Comparison between uncontrolled clinical trial and controlled BGL excursions with HDT framework is shown in the top plot. The middle plot shows the insulin recommendation during the clinical trial, and the bottom plot shows the insulin recommendation from HDT.

Figure 26

Insulin recommendation and BGL for $P_{13}$—Comparison between uncontrolled clinical trial and controlled BGL excursions with HDT framework is shown in the top plot. The middle plot shows the insulin recommendation during the clinical trial. The bottom plot shows insulin recommendations from HDT.

Figure 27

Percent improvement in Time in Range for 15 patients.

Figure 28

Percent improvement in hypo events for 15 patients with HDT.

Figure 29

Percent improvement in hyper events for 15 patients with HDT.